AI doi: 10.3390/ai7100391

Authors: Sewoong Lee Jungyub Woo Sangsu Choi

Large language model (LLM)-based AI agents combine reasoning and planning, memory, external knowledge use, tool calling, and multi-agent collaboration to perform complex tasks. However, they face planning failures, context-maintenance limitations, collaboration instability, vulnerabilities arising from external interactions, bias, unclear attribution of responsibility, and difficulties in oversight and incident tracing. This narrative literature review examined the technological evolution and technical limitations of AI agents and developed a conceptual mapping relating these limitations to selected issues of bias and governance. Selected early theoretical studies were included; the main focus was post-2022 scholarly literature and official institutional and corporate materials through July 2026, with publication-status and evidence updates during revision in September 2026. Sources were categorized into concepts and trends, single- and multi-agent architectures, technical limitations, and socio-ethical issues; limitations were compared using a problem addressed–mitigation approach–residual limitation framework. The analysis suggests that autonomous actions performed on behalf of users, ambiguity in the scope of authority, multi-agent interactions, external knowledge use, observation of external environments, and tool calling may be linked to unclear attribution of responsibility, bias propagation and amplification, and difficulties in oversight and incident tracing. Potential mediators include multiple actors, miscoordination and conformity, inherited external-source biases, and new attack surfaces. This conceptual mapping provides a basis for future research on trustworthy AI-agent design, operation, and governance.

AI doi: 10.3390/ai7100390

Authors: Sonja Djukić Popović Dejan Viduka Dragiša Žunić Aleksandar Stokić Stefan Popović

The rapid development of artificial intelligence over the past few years has significantly impacted all levels of education. Despite the widespread adoption of generative AI in education, empirical studies simultaneously analyzing the attitudes of teachers and students particularly in primary and secondary education remain relatively scarce. The emergence of generative AI systems, such as ChatGPT, Gemini, Copilot, and other tools based on large language models, has created new opportunities for personalizing instruction, enhancing the learning process, and supporting teachers in the preparation and implementation of teaching activities. The aim of this research is to analyze the attitudes of primary and secondary school teachers and students regarding the use of generative artificial intelligence in teaching, identify the key benefits and risks of its implementation, and assess the educational system’s readiness for its integration. The study was conducted in schools falling under the jurisdiction of the two largest school administrations in the Republic of Serbia and involved 1042 teachers and 2587 students. The study was guided by a multidimensional conceptual framework integrating AI literacy, perceived benefits, perceived risks, ethical considerations, adoption readiness, and perceived learning outcomes. Data were collected using structured questionnaires and analyzed using descriptive statistics, Mann–Whitney U tests, and Spearman’s rank correlation. The internal consistency of the research dimensions was assessed using Cronbach’s alpha, with coefficients ranging from 0.82 to 0.91 and an aggregate value of α = 0.92. The analysis involved comparing the attitudes of teachers and students, as well as identifying statistically significant differences between the groups studied. The results indicate that 93.4% of students and 72.6% of teachers reported having used AI-based tools. Students reported higher AI literacy (M = 4.21 vs. 3.89) and perceived benefits (M = 4.31 vs. 4.12), whereas teachers reported higher perceived risks (M = 4.32 vs. 3.83) and stronger ethical concerns (M = 4.46 vs. 4.03); these group differences were statistically significant (p < 0.001). In addition, frequency of AI use was positively associated with perceived benefits (Spearman’s ρ = 0.56, p < 0.001). At the same time, both groups identified academic integrity, inaccurate information, over-reliance on AI systems, and the need for clear guidelines as important concerns. The obtained results provide a foundation for formulating policies to integrate artificial intelligence into the education system, developing teacher professional development programs, and defining national guidelines for the safe, ethical, and pedagogically sound application of AI technologies in education. This study provides large-scale empirical evidence on the adoption of generative AI in primary and secondary education in the Republic of Serbia and offers practically relevant insights for educational institutions, policymakers, and curriculum developers. The novelty of this study lies in the simultaneous, multidimensional examination and direct comparison of teachers’ and students’ perspectives on generative AI within the same K–12 educational context, integrating AI literacy, perceived benefits, perceived risks, ethical considerations, adoption readiness, and perceived learning outcomes within a common conceptual framework.

AI doi: 10.3390/ai7100389

Authors: Nicholas Catasso

Research on large language models as linguistic judges has focused primarily on numerical outcomes, while their accompanying explanations remain understudied. This article qualitatively analyzes 750 explanations generated by five publicly accessible systems rating 150 contextually embedded German stimuli. The findings show that similar scores frequently conceal substantial differences in the accompanying metalinguistic accounts. Overt morphosyntactic violations generally elicit accurate diagnoses, whereas gray-zone constructions are often normalized or reduced to categorical standard/non-standard oppositions. Diatopic forms are usually accepted, but their regional classification is frequently broad or inconsistent. Register-sensitive items produce the clearest contextual reasoning, although often in formulaic terms. The explanations also reveal recurrent practices of norm invocation, repair, target substitution and sociolinguistic labeling. Although they do not transparently represent model-internal computation, they constitute behavioral data revealing the normative categories and sociolinguistic assumptions that models make available in interaction. Evaluations of LLMs as linguistic judges should therefore assess scores and explanations jointly.

AI doi: 10.3390/ai7100388

Authors: Yaxin Xu Yinxiang He Tianyi Liu Ningzhong Liu Han Sun Mingjing Pei Jiaquan Shen

Efficient exploration remains a key challenge in cooperative multi-agent reinforcement learning (MARL), where the novelty of a joint situation may arise not only from unfamiliar individual observations but also from previously underrepresented interaction patterns among agents. Existing intrinsic-reward approaches commonly characterize novelty through state, observation, value, or prediction representations, which may not explicitly capture such relational changes. To address this issue, we propose Relationship Graph Reconstruction Exploration (RGRE), an intrinsic exploration method that characterizes novelty from the perspective of inter-agent relationships. RGRE first constructs an attention-derived relationship representation from agents’ local observations and then employs a graph autoencoder to learn its latent relational structure. The reconstruction discrepancy is used as a proxy for relational novelty and incorporated into the environmental reward as an intrinsic exploration bonus. Experiments on the Multi-Agent Particle Environment (MPE) Spread task with 3, 6, and 10 agents and on six StarCraft Multi-Agent Challenge (SMAC) scenarios show that RGRE achieves higher mean final performance than QMIX on all nine evaluated tasks. Based on the final-performance results, the average improvements are 5.82 percentage points in landmark occupation rate across the three MPE settings and 7.84 percentage points in test win rate across the six SMAC scenarios.

AI doi: 10.3390/ai7100387

Authors: Jehad M. Hamamreh Aman M. Araf Ahmad M. Jaradat Adnan Daraghmeh Moamin Abughazala Henry Muccini

The rapid expansion of the Internet of Medical Things (IoMT) has enhanced healthcare services, but it has also exposed these systems to a growing range of cyberattacks. Addressing this challenge requires effective intrusion detection solutions. Machine learning-based intrusion detection systems (IDSs) offer a promising approach. In this research, we leverage IoMT-TrafficData, a recently published dataset specific to IoMT environments, to develop intrusion detection models based on federated learning (FL), which preserves data privacy while enabling the use of lightweight learning algorithms suited to resource-constrained devices. In addition, SHAP-based interpretability methods are employed to explain model decisions, thereby improving reliability and transparency. Compared with existing IDS solutions for IoMT, this research proposes an approach that balances performance, privacy, and interpretability while accounting for resource limitations, resulting in three unified, lightweight, and interpretable detection models tailored to different data types in IoMT environments. This work thus provides a solid foundation for future research on secure medical IoT systems.

AI doi: 10.3390/ai7090386

Authors: Adel A. Ahmed

The rapid evolution of sophisticated cyber threats has drastically increased the cybersecurity risks in Industrial Internet of Things environments due to the massive interconnection of industrial devices, sensors, programmable logic controllers, gateways, and edge computing infrastructures. Traditional intrusion detection systems are insufficient for modern IIoT networks due to challenges with dynamic attack behaviors, class imbalance, concept drift, and computational limitations of resource-constrained edge devices. Although machine learning and deep learning have significantly improved intrusion detection performance, current studies usually deal with these challenges separately and do not provide a holistic view of adaptive and real-time industrial internet of things (IIoT) cybersecurity. In this survey, we provide a structured narrative review of adaptive intrusion detection techniques, focusing on three emerging research directions, including GAN-based data augmentation, drift-aware learning, and Edge Intelligence. It provides a structured narrative review of machine learning, deep learning, and hybrid IDS models, benchmark IIoT datasets, and representative techniques addressing data imbalance, concept drift, and low-latency edge deployment. The survey further provides a comparative study of existing approaches in terms of detection capability, adaptability, computational efficiency, scalability, and deployment suitability. To fill the gap between those complementary research directions, the survey combines the literature into a unified reference architecture that merges GAN-based data augmentation, drift-aware learning, and Edge Intelligence to enable adaptive, real-time IIoT intrusion detection. Finally, we discuss key research challenges and future opportunities in autonomous, collaborative, and trustworthy IIoT cybersecurity, thus providing a practical roadmap for the development of next-generation intelligent intrusion detection systems.

AI doi: 10.3390/ai7090385

Authors: Salar Shakibhamedan Nima Amirafshar Axel Jantsch Nima TaheriNejad

Approximate multipliers have potential to improve energy efficiency in Deep Neural Networks but introduce computational errors that degrade accuracy. This paper introduces a novel method, leveraging approximate multipliers to enhance accuracy, while improving computational and energy efficiency. We propose a layer-wise heterogeneous approach using quantized approximate multipliers (INT8) in DNNs, applying varying levels of approximation across layers. This approach achieves estimated energy savings of up to 44.75% across all evaluated configurations, including up to 43.72% for the VGG models, while improving Top-1 accuracy by up to 2.07 percentage points relative to the corresponding exact INT8 baseline. Using Information Bottleneck (IB) theory, we analyze the enhanced information flow and feature extraction capabilities enabled by approximate multipliers. Through Information Plane (IP) analysis, we gain insights into DNN behavior and demonstrate how this method can overcome accuracy limitations. An analytical MAC-count comparison indicates that, under the experimental settings considered, the forward-pass-only GA search requires 146× to 22,500× fewer MAC operations than the corresponding gradient-based training schedules; this comparison reflects computational workload rather than an equivalent-objective algorithmic speedup.

AI doi: 10.3390/ai7090384

Authors: Hai Wang Junhao Wen Chunlai Yang Kamara Kekele Adnan Fayçal Jiale Gu

Transformer-based detectors model long-range context effectively, yet their representations remain dominated by RGB appearance and can become unreliable in cluttered, occluded, or crowded scenes. We present BMF-DETR, a pseudo-depth-guided detector that introduces RGB-derived geometric structure without requiring a depth sensor. DA3Mono-Large from Depth Anything 3 generates spatially aligned pseudo-depth maps offline, while two ResNet-50 streams encode appearance and relative geometry. Bidirectional cross-modal attention (BCMA) establishes two-way correspondence, and multi-strategy fusion (MSF) combines the streams through global calibration, channel allocation, and spatial gating before squeeze-and-excitation (SE) recalibration. On the fixed validation/evaluation split of the 2024 Roboflow-curated PASCAL VOC derivative, the complete model reaches 60.80 AP, compared with 52.80 AP for a capacity-matched dual-RGB control. BMF-DETR obtains 49.30 AP on COCO 2017. Its detector contains 58 M parameters and requires 103 GFLOPs; these figures exclude offline pseudo-depth generation. A shared-low-level variant retains 60.10 AP with 50 M parameters and 87 GFLOPs. The results show that pseudo-depth can serve as a useful auxiliary representation when its contribution is separated from capacity effects and evaluated under controlled fusion settings.

AI doi: 10.3390/ai7090383

Authors: Wei Zhang Hui Fang Tongle Wu Chaoqun Wang Libo Xu Jiajun Bu Yueyao Yu Qiming Zhong

The rapid advancement of large language models (LLMs) has created new opportunities for intelligent fault diagnosis, particularly in complex industrial systems, such as heating, ventilation, and air conditioning (HVAC) in urban rail transit. Although LLMs have shown strong general reasoning capabilities, adapting them to domain-specific fault diagnosis tasks remains challenging. This is particularly true for textual maintenance records, where sparse and brief entries can only provide limited context for effective knowledge adaptation. To address this challenge, we propose a multi-stage post-training framework based on LLMs. The framework consists of three components: (1) data augmentation via retrieval-augmented generation (RAG) to enrich brief maintenance records with domain knowledge and reasoning traces; (2) supervised fine-tuning (SFT) for domain-specific adaptation; and (3) reinforcement learning with group relative policy optimization (GRPO), using a task-specific reward that separately evaluates root-cause identification and maintenance action recommendation. The framework is applied to a real-world textual HVAC fault dataset derived from Ningbo Rail Transit, covering 33 equipment categories and over 100 fault types. With Qwen3-0.6B as the base model, the proposed method significantly improves diagnostic accuracy and reasoning quality, achieving a 97% increase in model-based diagnostic accuracy (from 0.323 to 0.635) and a 48% improvement in human expert evaluation scores (from 0.509 to 0.754). Moreover, conventional machine learning baselines, such as Support Vector Machine (SVM), Random Forest (RF), and Multi-Layer Perceptron (MLP), achieve accuracies below 0.4 in this task, further highlighting the superiority of the proposed LLM-based framework. These results indicate that the proposed multi-stage post-training framework effectively improves LLM performance on real-world text-based fault diagnosis. It provides a practical and extensible solution for intelligent maintenance decision support in complex electromechanical systems.

AI doi: 10.3390/ai7090382

Authors: Arif Bramantoro Mohd. Amiruddin Saddam Eko Sakti Pramukantoro M. Ali Fauzi

Retailers managing large stock-keeping unit (SKU) catalogues need a compact, auditable view of how individual products behave in order to plan replenishment, assortment and promotions. We present an interpretable analytics pipeline that derives SKU-level recency, frequency and monetary (RFM) features from one year of fashion-retail transactions (40,760 SKUs; 101,144 orders), segments the catalogue with prototype-based clustering under explicit internal validation, reads the segments alongside a co-purchase layer obtained by closed-itemset mining, and delivers the result to category managers through a deployed web application. Model selection is made explicit rather than assumed: K-Means and K-Medoids are compared at matched cluster counts on the Silhouette coefficient and the Davies–Bouldin Index (DBI). The two methods perform comparably at k=2, and K-Means is clearly superior at every larger cluster count; the selected configuration (K-Means, k=4) attains a Silhouette of 0.577 and a DBI of 0.624, against 0.522 and 0.762 for the best K-Medoids configuration. The resulting segments separate a small group of fast-moving items from a long tail of low-frequency products, and profiling them on the monetary axis shows that the separation tracks value contribution: 7.1 per cent of the clustered SKUs account for 28.8 per cent of monetary contribution, and the 19 fast-movers contribute roughly eleven times the per-SKU average. We set out how each segment maps to replenishment, assortment and promotion decisions, together with the validation each mapping would require before adoption. Two scope conditions should be read with the results: clustering uses the recency and frequency axes, with monetary value reported as a descriptive attribute of the segments rather than as a clustering input, and the co-purchase layer rests on very low support because baskets in this catalogue average 1.16 SKUs, so the itemsets are exploratory anchors rather than association rules.

AI doi: 10.3390/ai7090381

Authors: Andreas Spilz Heiko Oppel Michael Munz

Automated evaluation of movement quality can enhance physiotherapeutic treatment and sports training by providing objective, real-time feedback. However, deep learning models that assess movements captured by inertial measurement units (IMUs) are often limited by data scarcity, class imbalance, and label ambiguity. We present a data augmentation method for IMU-derived orientation data that generates additional examples by systematically modifying movement trajectories and passing them through a musculoskeletal simulation. The approach enforces the joint-range limits of a musculoskeletal model and enables automatic labeling by combining inverse kinematic parameters with a knowledge-based evaluation strategy. Across four datasets of varying complexity, augmented variants closely resemble real-world data and contribute to gains in classification accuracy, generalization to unseen subjects, and patient-specific fine-tuning from few examples. The magnitude of these gains varies with dataset properties, in particular class balance and label ambiguity. These findings indicate that musculoskeletal simulation-based augmentation can address common challenges faced by deep learning applications in physiotherapeutic exercise evaluation.

AI doi: 10.3390/ai7090380

Authors: Ivan Bulychev Andrey Savchenko

Recent advances in large language model (LLM) agents have shown promise for autonomous decision-making in recommender systems. However, existing approaches suffer from two fundamental limitations: flat agent memories that conflate different information modalities and prohibitive computational costs that prevent scaling beyond a few hundred users. We propose Hybrid-GraphRAG, a recommender system that integrates hierarchical agent memory structures, graph-based retrieval-augmented generation (Graph RAG), and knowledge distillation for scalable deployment. Our approach extends agent-based collaborative filtering by structuring agent memories into intrinsic, collaborative, and interaction tiers that disentangle different information types; performing multi-hop retrieval over a dynamically constructed heterogeneous interaction graph to enable relational reasoning; and distilling LLM-generated memory dynamics into efficient graph neural encoders with adaptive gating between full and efficient inference paths. Experiments on Amazon review datasets (CDs and Vinyl, Office Products) demonstrate that Hybrid-GraphRAG achieves recommendation quality comparable to full LLM-based agents while reducing computational cost by 85% and improving NDCG@10 by 12.7% over flat-memory agent baselines. Our results establish a principled bridge between semantic agent reasoning and scalable graph-based recommendation.

AI doi: 10.3390/ai7090379

Authors: Yuxuan Zhou Shudong Zhuang Ao Sheng Yizheng Ge Jiarui Zhu Zhizhou Wang Yuxian Lei Xinyan Cao

The reliable detection of spot-weld defects in automotive structural components is challenged by large variations in defect scale, severe background interference and limited detection accuracy. Here, we propose YOLOv8-RFA-iEMA-RH, an improved YOLOv8n-based detector for spot-weld defects. A receptive field attention convolution module (RFACM) is introduced into the backbone to strengthen local texture representation through multi-receptive-field feature modelling. An improved Efficient Multi-scale Attention module (iEMA) is incorporated into the neck to enhance global context modelling and suppress background interference. In addition, a structurally re-parameterized RepHead is integrated into the detection head to enhance feature learning during training while maintaining a simplified single-branch structure for inference. On the self-built spot-weld defect dataset, the proposed model achieves 92.7% Recall, 91.2% F1, 97.9% mAP@0.5 and 71.9% mAP@0.5:0.95, improving on the YOLOv8n baseline by 2.3, 1.3, 2.5 and 3.1 percentage points, respectively. Cross-dataset evaluation on NEU-DET further yields 78.8% mAP@0.5 and 48.7% mAP@0.5:0.95. These results demonstrate improved detection accuracy and cross-dataset adaptability; actual inference speed and memory consumption require further validation on specific deployment hardware.

AI doi: 10.3390/ai7090378

Authors: Sittinon Thanonklang Talit Jumphoo Wongsathon Pathonsuwan Kasidit Kokkhunthod Atcharawan Rattanasak Rattikan Nualsri Porntip Nimkuntod Pattama Tongdee Monthippa Uthansakul Peerapong Uthansakul

Congenital heart disease (CHD) is a major cause of infant morbidity and mortality, and timely diagnosis remains difficult where advanced imaging is not consistently available. Automated phonocardiogram (PCG) screening can support early triage, but pediatric multiclass classification is limited by subtle acoustic differences and class imbalance. This study proposes a five-class heart-sound framework (Normal, ASD, PDA, PFO, and VSD) using multi-domain feature fusion and a convolutional recurrent neural network (CRNN). A four-channel representation combining Log-Mel, PCEN-Mel, Gramian Angular Summation Field (GASF), and Markov Transition Field (MTF) is modeled with a CNN encoder, bidirectional GRU, and dual temporal pooling. Recordings are segmented with uniform 50% overlap and evaluated under a strict subject-wise protocol with macro-F1-guided model selection. Across five runs on the pediatric ZCHSound dataset, the five-class model achieves a mean macro F1-score of 78.61 ± 1.00% and a mean per-subject accuracy of 87.94 ± 0.87%. A dedicated Normal-versus-Abnormal model reaches 95.74% accuracy. These results suggest that multi-domain fusion with sequence-aware modeling is a promising approach for pediatric CHD auscultation support.

AI doi: 10.3390/ai7090377

Authors: Andrii Pukach Oleksandr Morushko Vasyl Teslyuk Yurii Kynash

This study develops a specialized method for the synthesis of intelligent platforms for cross-referral transition between cognitive basis systems (CBSs) in the field of HCI object perception subjectivization, within the context of global scientific and applied efforts to increase the intelligence level of interaction between humans and software and/or hardware products (SHPs). The proposed method comprises a conceptual model, a mathematical model, and a specialized algorithm. Practical implementation was conducted using R (within an appropriate IDE) and Python. The method was approbated by solving a relevant applied problem: synthesizing a cross-referral transition platform between a specialized CBS for HCI object perception subjectivization and an existing personality classification system based on 16 Jungian sociotypes. The results indicated a 29.5% baseline correspondence rate for pairs of dominant perception impact factors regarding their exclusive alignment with specific Jungian sociotypes. Furthermore, the accuracy rate of the built-in multilayer perceptron (MLP) artificial neural network (ANN) for dominant factor pairs reached approximately 83.3% ± 2.95%, while introducing a third dominant factor raised the accuracy threshold to a potential 96.7% within the evaluated scenario. In addition, the paper outlines prospects for further research into the potential of the proposed method for identifying and documenting cross-referral relationships between fundamental constituents of various conventional and alternative CBSs.

AI doi: 10.3390/ai7090376

Authors: Pavlo Radiuk Oleksander Barmak Liliana Klymenko Iurii Krak

Deep electrocardiogram (ECG) classifiers can achieve strong predictive performance, yet their latent evidence remains difficult to audit, and explanation pipelines can become misleading when semantic contracts or label provenance fail. In this work, we propose CARDIA-X, an electrocardiographic instantiation of the global semantic transition and rough-set rule sequence that couples a versioned 52-target semantic contract with evidence-gated class eligibility, separated primary and external branches, and end-to-end provenance controls. After correcting the compensatory-pause ratio to be nonnegative and unbounded above, patient-grouped development reconstruction achieved ratio-specific mean absolute errors of 0.2394 out of fold and 0.231 on validation. The frozen internal evaluation contained 2692 records from 1599 patients but no atrial-fibrillation-positive or atrial-flutter-positive exported labels; audit traced this to an upstream label-export discrepancy, so atrial fibrillation discrimination could not be estimated and no production rules or inference-route claims became eligible. External Lobachevsky University Database (LUDB) R-peak validation achieved an F1 score of 0.916, while single-clinician agreement on archived explanation displays reached Cohen’s kappa 0.683. CARDIA-X therefore currently supports reproducible research auditing while providing a foundation for future clinical validation, potential deployment, and evaluation of patient benefit.

AI doi: 10.3390/ai7090375

Authors: Nikola Jovišić Milica Škipina Vanja Švenda Dubravko Ćulibrk Boris Antić Branko Brkljač

Annotated medical imaging data remain scarce, and labels are often weak and noisy: in mammography, an examination comprises several high-resolution views, yet the diagnostic outcome is recorded only at the breast level. Such problems are naturally cast as Multiple Instance Learning (MIL), where the model must infer instance-level structure from bag-level labels alone. Although contemporary foundation encoders supply strong general-purpose embeddings, augmenting MIL data in this representation space remains an open problem as established techniques act on one instance at a time and ignore the statistical dependencies binding a bag together. We address this with SetFlow, a generative model that learns the distribution of complete MIL bags directly in a frozen encoder’s embedding space. SetFlow couples flow-matching training with a Set Transformer-inspired backbone, making it invariant to instance ordering while modeling intra-bag relationships. Generation is conditioned jointly on class label and per-instance scale, yielding coherent, semantically faithful bags rather than isolated vectors. Evaluating on two large public mammography datasets and two encoders, we assess distributional fidelity, nearest-neighbor behavior, and downstream augmentation utility. We show that generated bags reproduce real-data statistics and improve classification in certain configuration, with performance gains varying on the amount of synthetic data. An architecture ablation confirms each design choice contributes to performance.

AI doi: 10.3390/ai7090374

Authors: Yongqiang Yang Chi Fai Cheung

Accurate 3D topography measurement of micro-structured surfaces remains challenging due to the limitations of conventional autostereoscopic systems, particularly the intrinsic constraints of light-field imaging and dependence on single-source data. Building on a multichannel autostereoscopic measurement system that simultaneously captures a high-resolution (HR) 2D center view containing rich textural and edge information and a light-field image providing dense multiview geometric cues, a deep learning-enhanced feature fusion network is introduced. This model is a hybrid deep learning architecture featuring a convolutional local feature extractor for the HR image, a Transformer-based global feature extractor for angular relations in the light field, and a cross-channel attention fusion module for effective feature integration. The end-to-end trainable network is optimized using a composite loss function. Experiments on synthetic and real micro-structured surfaces demonstrate stable performance of the proposed approach, achieving improved accuracy and stability over current depth-estimation methods in challenging micro-scale scenarios.

AI doi: 10.3390/ai7090373

Authors: Xiangyang Deng Hongji Zhu Limin Zhang Yupeng Fu Shandong Wang

Traditional imitation learning is prone to distribution shift and trajectory divergence in highly dynamic, strongly time-varying flight tasks. To address this, we propose a temporally adaptive action-chunking framework built on temporal-derivative encoding, termed Temporal-Derivative-Encoded Adaptive Action Chunking with Transformer (TDA-ACT). First, using state temporal-derivative features as the core representation, we construct a confidence-estimation mechanism that also incorporates the latent-variable variance of the maneuver-mode representation and the action-prediction variance produced by the decoder. Second, we develop a confidence-guided adaptive temporal-ensembling strategy that uses this confidence metric to jointly adjust the fusion scope and the fusion weights of historical predictions, enabling a dynamic trade-off between long-horizon smoothing under stable conditions and high-frequency responsiveness during aggressive maneuvers. On both the Loop and AileronRoll maneuvers in JSBSim, TDA-ACT reduces action jerk and suppresses trajectory divergence relative to ACT and other imitation-learning baselines.

AI doi: 10.3390/ai7090372

Authors: Alireza Yarmohammad Tooski Ehsan Kargar Mehrnegar Foratinejad Mohammad Sadegh Javadi Amin Mirgheisari Mohammad Hossein Alizadeh Roknabadi Alireza Soleimani Anna Pinnarelli Goran Strbac

The convergence of additive manufacturing (AM) and artificial intelligence (AI) is poised to redefine the landscape of modern production; however, the literature remains fragmented across isolated applications, lacking a unified perspective on the engineering impact and practical deployment of these technologies. This review provides a comprehensive and critical synthesis of the state of the art in AI-enhanced AM, systematically covering supervised, unsupervised, and reinforcement learning paradigms, alongside deep-learning-based computer vision, natural language processing, and robotics. In contrast to prior works that focus on singular aspects, this paper consolidates progress across four core engineering domains: (i) lightweight and manufacturable design, (ii) real-time in situ defect detection and process analysis, (iii) energy-efficient process optimization, and (iv) cost-effective build-time estimation with intelligent support minimization. Beyond cataloging these advances, this review identifies key quantitative benchmarks and recurring technical challenges, including data scarcity, poor model generalizability, and the critical gap between offline prediction and real-time closed-loop control. To transcend these isolated successes and enable industrial adoption, we propose a novel, unified closed-loop AI-AM framework that tightly integrates generative design, process planning, in situ production monitoring, and continuous model updating into a cohesive digital thread. Furthermore, a domain-stratified SWOT analysis is compiled, offering a strategic evaluation of strengths, weaknesses, opportunities, and threats across the four application pillars. By bridging the gap between laboratory prototypes and production-ready autonomous systems, this review serves as a definitive reference for researchers and practitioners aiming to navigate, deploy, and advance the rapidly evolving field of AI in additive manufacturing.

AI doi: 10.3390/ai7090371

Authors: Andreas Demetriou George Spanoudis Elena Kazali Andreas Savva Nikolaos Makris Smaragda Kazi

We compared four large language models (LLMs; ChatGPT, Grok, Gemini, and DeepSeek) with humans on tests of cognitive development, assessing relational integration, linguistic awareness, general and domain-specific reasoning, and cognitive self-awareness to specify how LLMs compare with humans along cognitive development hierarchies. LLMs also discussed how Descartes’s Cogito applies to them and rated themselves on aspects of Artificial General Intelligence (AGI). Accordingly, we propose a novel interdisciplinary comparison of human and LLM capabilities that integrates developmental, cognitive, and psychometric psychology. Overall, the processes in humans and LLMs were highly similar. All LLMs attained perfect linguistic and metalinguistic performance. ChatGPT and Gemini outperformed university students in mathematics and causal reasoning. Grok performed slightly better and DeepSeek considerably worse. All LLMs underperformed in visual–spatial tasks. Self-evaluation profiles broadly mirrored performance profiles: ChatGPT and Grok rated themselves highly in reasoning and low in visualization, Gemini inflated visualization by reframing it as linguistic creativity, and DeepSeek consistently underrated itself. Each LLM restated Descartes’s Cogito differently, reflecting its own priorities, and denied having high AGI; these self-characterizations were generally stable about a year later. Therefore, LLMs displayed “subjective” task scaling, implying algorithmic or functional self-monitoring, capturing their architectural profile of performance, but they were modest in claiming above-human intelligence. We discuss implications for an integrated natural–artificial intelligence theory. We also sketch a developmental engineering model that might remove the limitations of each LLM.

AI doi: 10.3390/ai7090370

Authors: Petar Zhivkov Anton Totomanov

Information architecture may shape harmful conduct in dynamic multi-agent systems, yet its relationship to objectively measured misrepresentation remains unclear. We examine seller-only communication and market-wide public history in a repeated hidden-quality market populated by 12 GPT-4o-mini seller agents and 12 buyer agents over 20 rounds. The submitted five-replicate pilot and a fresh primary-wording replication batch with twenty replicates per condition compared five conditions. Deceptive selling is operationalized solely as observable quality overstatement, defined as listing an item above its source-quality tier; strategic intent is not inferred. The primary outcome is the fraction of unseeded sellers making at least one false quality claim. With intervention sellers held fixed, the fresh primary-wording replication batch showed mean forum-associated prevalence differences of 0.850 under private history and 0.483 under market-wide public history. These contrasts apply only to intervention-present markets because the design does not include a forum condition without intervention sellers. Two seller-listing paraphrases produced materially different private-history contrasts, indicating prompt sensitivity rather than wording invariance. Across the submitted pilot, seller conduct, realized buyer harm, and market activity diverged. These descriptive results motivate multi-agent safety evaluations that jointly examine information architecture, harmful conduct, and system utility. They do not identify message versus prompt mechanisms, buyer- versus seller-side public-history mechanisms, intent, or seed-to-peer transmission.

AI doi: 10.3390/ai7090369

Authors: Javier Martín Arista Matías Núñez

This paper raises a methodological question: How can we assess machine-made Old English when there is no parallel reference text and the standard metrics do not fit the task? We propose a pipeline with five measuring layers plus two compliance components, including lexical attestation with form linking, frequency-profile diagnostics, character-level comparison, word-embedding geometry under a verified mapping and dependency parsing, with bootstrap confidence intervals around the main contrasts. We apply the pipeline to a machine-made version of Gregory’s Dialogues: 4217 sentences, one for each sentence of the Old English original, generated under hard constraints. The unattested residue is two word types and 0.003% of tokens. The frequency profile diverges from the original by 0.008, less than the original diverges from the background corpus. At character level, in embedding space and in parsed syntax, the generated text stands at the same distance from the Dictionary of Old English Corpus as the original itself does. We propose an overall metric G, the geometric mean of seven bounded components, which scores the text at 0.991 with a confidence interval of [0.991, 0.992]. Two blind detection experiments with expert judges place the index externally: roughly two-thirds of generated sentences pass as authentic to specialists, so the divergence the pipeline measures is real at corpus scale but not available to sentence-by-sentence reading. The main contribution is a reusable evaluation method for historical language generation, together with a single interpretable score that subsumes the partial metrics without hiding them.

AI doi: 10.3390/ai7090368

Authors: Andrada Puisor Stefan Caramizoiu Stefan-Marian Iordache Bogdan Bita

Environmental-monitoring systems often bundle sensing, communication, storage, visualization, and control into one application, making later changes difficult. We designed a service-oriented platform that separates these functions through defined interfaces. It combines a Raspberry Pi gateway, a dedicated motor-control microcontroller, five environmental sensor modules, Node-RED middleware, a database, and a web interface. Deterministic code alone evaluates threshold and composite rules and controls safety-relevant alerts; an optional large language model (LLM) turns pre-computed statistics and rule outcomes into narrative reports. We examined data acquisition and rule processing during two short indoor campaigns. In the residential campaign, the SCD41 yielded 78 valid three-minute bins (234 min of recorded data) across four sessions between 09:18 and 17:12 local time; binned CO2 concentrations ranged from 679 to 1471 parts per million (ppm). Using the initial campaign for development and the residential campaign as a temporal holdout, the persistence model produced a 15 min forecast mean absolute error of 58.3 ppm and a root mean square error of 78.8 ppm. A separate controlled experiment generated 270 reports from nine deterministic synthetic scenarios. Every reporter preserved all deterministic alert identifiers, while the fixed template and seven of the nine locally hosted LLMs achieved 100% numerical fidelity. Qwen 3.5 9B was the only LLM that returned all required measured content without automated claim-review flags and produced identical outputs across repetitions for every scenario. These results confirm integration and functional separation under the tested conditions, but they do not demonstrate week-scale reliability, longer-horizon forecasting accuracy, robotic mobility performance, or load scalability.

AI doi: 10.3390/ai7090367

Authors: Weihan Shi Yinlong Liu

Visual Question Answering (VQA) in embodied settings draws on computer vision, natural language processing, and robotics. Unlike traditional VQA, Embodied Question Answering (EQA) may require an agent to acquire and retain evidence from a 3D environment before answering a natural language query. This semi-systematic survey reviews a core corpus of 72 papers selected from 210 candidates and supplements it with a separate targeted qualitative update comprising 28 additional records identified through July 2026. We use PMRA (Perception–Memory–Reasoning–Action) as an author-developed analytical framework rather than a new robot-control architecture, together with a three-level taxonomy of task formulations, method architectures, and capability dimensions. An audit of the 20 tabulated datasets, with counts recomputed from the accompanying extraction sheet, finds that five permit or require active exploration and only one requires physical object interaction. We also define four architecture-based stages without using publication year as an assignment rule. The qualitative synthesis identifies examples of explicit memory interfaces and question-conditioned information-acquisition mechanisms, but it does not infer their prevalence, temporal growth, or relative importance from publication frequencies. Finally, we discuss four recurring limitations of foundation-model approaches and nine research directions toward 2030.

AI doi: 10.3390/ai7090366

Authors: George Melville Dena Ghiassi Scott Inthathirath Julian Yeomans

AI and machine learning deployments in regulated decision contexts face an intolerance for inadmissible outputs (“hallucinations” when the model is generative) that current explainability methods address only after the fact. Bounded AI denotes prevention by architectural design. This study establishes five conditions (C1–C5) under which a bounded artificial intelligence (AI) architecture transfers from one regulated decision domain to another. Conditions C1 through C4 adapt or combine previously established principles. The most significant contribution is the discrete joint-state topology condition, C5, for which no precedent was found in this role. The claim is that these five conditions are jointly necessary for the closure property to survive an architectural transfer—while sufficiency is not claimed. Two of the five conditions are structural prerequisites governing whether the architecture’s operators can be constructed in a destination at all. The remaining three provide warrant conditions governing whether it is the appropriate instrument or not. In existing runtime-assurance architectures, the constraint acts after inference, on the output of the learned component. In contrast, the pattern developed in this paper reverses the assurance steps via a deterministic-first/learned-second approach. Namely, the assurance architecture acts before inference on the input domain: a deterministic filter admits only rule-compliant objects, and the trigger fires non-discretionarily on joint-state cell occupancy rather than on the learned score. The architecture’s domain-neutral type signatures are formalized, and three structural transfers are developed in depth: predictive maintenance, energy-grid management, and credit underwriting, each concluding with a closure proof. All three transfers remain conceptual and report no deployment outcomes. The proofs are conditional on three stated premises that establish soundness with respect to a rule set rather than a safety case. The strongest evidence of transferability is a market-surveillance destination classified as admissible in advance and later realized on a live venue. C5 is what discriminates the transferability. It is shown that the autonomous-vehicle perception case satisfies C1 through C4, but fails C5 because the required distinctions are absent from the representation at every granularity—which is a failure that no additional compute can resolve. The architecture becomes domain-neutral through the act of transfer, not before it.

AI doi: 10.3390/ai7090365

Authors: Yazmin Mariela Hernández-Rodríguez Oscar E. Cigarroa-Mayorga

This study evaluated unconditional synthetic mammography generation with Deep Convolutional Generative Adversarial Networks (DCGANs). The dataset comprised 690 anonymized mammograms, equally distributed across BI-RADS 1–6 (115 images/category). All images were standardized to 512 × 512 pixels, reoriented to a common right-breast view, and prepared in two input domains: grayscale and color-mapped intensity encoding. Two models were compared: a standard DCGAN trained directly at 512 × 512, and a progressive DCGAN trained through discrete stages from 8 × 8 to 512 × 512. Training used PyTorch (Python 3.10), latent dimension = 100, Adam, learning rate = 2 × 10−4, β1 = 0.5, binary cross-entropy loss, and batch size = 1. Both models learned the low-frequency mammographic manifold, generating breast-like silhouettes and heterogeneous internal intensity distributions. However, the progressive DCGAN produced smoother contours, more coherent internal organization, and fewer grid/line artifacts than the standard model. Grayscale-only training showed weak learning, whereas the color-mapped domain improved structural recovery, although this advantage should be interpreted as computational rather than clinical. Late-epoch checkpoint analysis showed a structurally invariant generator with 43 tensors and 19,531,127 parameters; from epochs 89–100, relative checkpoint drift remained within 0.294–0.317%, with epochs 93–96 showing the most stable regime. Despite these advances, generated images still exhibited background speckle, coarse mottled texture, extra-anatomical bright structures, and limited diversity. Thus, the results support feasibility of synthetic mammography generation, but not yet clinically reliable synthetic data for direct AI training. The generated images should presently be regarded as exploratory complementary data pending expert, metric-based, and downstream validation.

AI doi: 10.3390/ai7090364

Authors: Rekha R Nair Tina Babu Sumendra Yogarayan Siti Fatimah Binti Abdul Razak

Sustainable agricultural planning requires predictive frameworks that can capture spatiotemporal variability, sustainability dynamics, and the potential outcomes of alternative management scenarios. The research proposes TSAFI-DT, a retrospectively validated, data-driven Digital Twin prototype integrating spatiotemporal data reconstruction, sustainability-state representation, hierarchical yield forecasting, counterfactual analysis, and scenario simulation. The framework operates on historical district-level APY observations and therefore represents a retrospective approximation of Digital Twin operation rather than a continuously synchronized cyber-physical agricultural Digital Twin. The Extended Regenerative Agriculture Index (eRAI) combines crop diversity, productivity–stability, land-use efficiency, and yield-trend information to characterize district-level sustainability states. A Bayesian-optimized XGBoost model is employed for one-step-ahead yield forecasting under temporal validation, while fixed-effects and synthetic-control analyses provide complementary associational and intervention-associated evidence. Evaluation using district-level groundnut data from India during 1997–2023 demonstrates that the proposed predictor achieves an RMSE of 0.171 t/ha and R2=0.92, outperforming the evaluated baselines with statistically significant differences (p<0.05). The fully adjusted fixed-effects model identifies a positive association between higher sustainability states and yield, while retrospective Digital Twin replay demonstrates close temporal agreement between predicted and observed outcomes. Model-based scenario simulations indicate predicted yield increases of up to 12.4% under the evaluated sustainability-state perturbations; these estimates represent counterfactual sensitivity rather than guaranteed causal effects. TSAFI-DT provides a reproducible framework for sustainability-aware agricultural forecasting, comparative scenario exploration, and data-driven decision support.

AI doi: 10.3390/ai7090363

Authors: Armita Dabiri Amir H. Behzadan

Artificial intelligence (AI) is transforming civil and construction engineering (CCE) education. CCE students must develop both AI technical proficiency and ethical awareness, ensuring that AI tools reflect the varied experiences of project clients. Integrating ethical AI instruction supports the formation of students’ professional identity by encouraging them to internalize values of responsibility, integrity, and equity in future practice. Adopting a narrative literature review, this paper examines the ethical concerns surrounding AI in CCE education, implications of existing regulations, and the need for institution-specific risk mitigation policies. The synthesis of the literature indicates that while integrating AI into CCE education may enhance learning experiences and personalized instruction, it also raises key ethical risks, such as algorithmic bias, privacy concerns, lack of transparency, threats to academic integrity, and digital inequity. As such, we also offer guidelines for responsible AI use in educational settings and propose an output governance framework and practical recommendations for educators and institutions, including performing regular ethical and algorithmic audits of AI tools, involving students in technology deployment decisions, and providing faculty development on digital ethics. By detailing actionable strategies, formalizing human-in-the-loop validation workflows, and preserving chains of provenance for educational metrics, these recommendations aim to align AI innovation with the core values of engineering education, i.e., integrity, equity, and public welfare.

AI doi: 10.3390/ai7090362

Authors: Shaoting Liu Haiyan Xu Ziqi Shuai Jifeng Xuan

Crash reports are critical for software debugging and maintenance. In software maintenance, developers reproduce a crash to locate the root cause via analyzing the crash report. A crash report is a document that records details of the crash as it occurred. The crash reports cannot be reproduced by developers because numerous crash reports in real-world issue-tracking systems contain incomplete reproduction steps and ambiguous descriptions. Existing works have explored the reproduction of crash reports for Android applications. These works focus on the process of reproducing crashes based on steps to reproduce and stack traces. However, it is a challenging task to manually identify reproducible crash reports from a large number of crash reports that consist of missing reproduction steps. To address this challenge, we propose PreCrash, a multi-feature fusion approach for predicting the reproducibility of a crash report. This approach helps identify reproducible crash reports before developers expend manual effort trying to reproduce the crashes. First, PreCrash automatically extracts 49 features from each crash report, which cover basic metadata, contextual content, comments, and supplementary evidence in crash reports. Second, PreCrash integrates four text representation models, including Word2Vec, GloVe, FastText, and BERT representations. Third, PreCrash employs a TextCNN-based prediction model to capture semantic patterns based on the integrated text representation. We evaluate PreCrash on 4133 crash reports collected from 102 real-world Java and Android projects on GitHub. Experimental results show that PreCrash achieves an accuracy of 0.91 and an F1-score of 0.91. PreCrash outperforms traditional machine learning models and single text representation baselines.

AI doi: 10.3390/ai7090361

Authors: Moses Guddah Adham Atyabi

Early stroke risk prediction offers an opportunity for timely interventions and may help reduce the clinical burden associated with stroke. Artificial intelligence (AI) provides medical practitioners with tools to analyse clinical biomarkers and predict a patient’s stroke risk. However, existing models lack interpretability and explainable decision support, limiting their adoption in clinical settings. This paper proposes a hybrid Gated Recurrent Unit (GRU)-TabTransformer architecture using cross-attention for stroke-status prediction. The proposed architecture comprises two stages. In the first stage (Model A), ordered feature-sequence representations from a GRU encoder are combined with concatenated categorical and numerical tabular features from a TabTransformer encoder. The model passes these distinct learned representations through cross-attention and linear projection layers before the final prediction. In the second stage (Model B), we augment our model with Large Language Models (LLMs) and Local Interpretable Model-agnostic Explanations (LIME) to provide per-sample, post hoc, human-interpretable explanations based on predicted probabilities. Experimental results show that both models achieve competitive sensitivity and accuracy values. In particular, the Synthetic Minority Over-sampling Technique (SMOTE) yielded a more balanced sensitivity–specificity trade-off, with a sensitivity of 74.00% and a specificity of 75.10%. Moreover, the model achieved an accuracy of 75.05% with SMOTE. An ablation study further shows that Cross-Attention offers better sensitivity and Receiver Operating Characteristic-Area Under the Curve (ROC-AUC), while Gated Fusion performs better on several other metrics. Additionally, age and average glucose level were the most influential stroke risk indicators, while Body Mass Index (BMI) and ever-married status were secondary model-attributed features. A two-factor repeated-measures Analysis of Variance (ANOVA) confirmed an interaction between model choice and the class-balancing technique used in stroke risk prediction systems. The Mistral + Hybrid GRU-TabTransformer architecture also recorded a mean inference time of 57.59s using few-shot prompting. Overall, the results provide a proof of concept for integrating hybrid GRU-TabTransformer with cross-attention and LLM-based explainability to support interpretable stroke risk prediction systems, pending robust external validation before deployment.

AI doi: 10.3390/ai7090360

Authors: Xueshuang Fan Mary Jane C. Samonte Xiaofeng Wang

Multi-modal MRI provides essential anatomical and pathological information for accurate brain tumor segmentation. However, deep learning-based segmentation methods are hampered by limited annotated data and incomplete modality acquisition in clinical MRI datasets. To address this issue, we propose a synthetic enhancement framework based on SCGAN-MultiJNet for multi-modal brain-tumor MRI. Specifically, SCGAN establishes a dual-branch disentangled latent space to independently encode images’ structural contour and textural features. Combined with the multi-scale fusion capability of MultiJNet, the proposed network realizes effective modality translation. Subsequently, synthetic samples are mixed with BraTS2020 training data to optimize the U-Net segmentation model. With the optimal real–synthetic data-mixing strategy, the segmentation metrics are improved: accuracy, Dice, precision, and IoU increase from 0.9857, 0.8919, 0.8980, and 0.8054 to 0.9864, 0.8962, 0.9205, and 0.8122. The proposed synthetic augmentation experimentally demonstrates enhancements in model robustness against MRI disturbances (including Gaussian blur, brightness shift, etc.). Finally, cross-domain generalization experiments conducted on the BraTS2025-SSA-Data demonstrate that introducing synthetic data augmentation can effectively boost the model’s cross-domain generalization capability. Overall, the proposed SCGAN-MultiJNet-based data synthesis algorithm provides a feasible technical solution to break the data bottleneck in multi-modal MRI brain tumor segmentation.

AI doi: 10.3390/ai7090359

Authors: Patrick Deininger Wolfgang Slany

Behaviour-Driven Development (BDD) encodes acceptance criteria in Gherkin, but hand-authoring is laborious, and it is unclear which large language model (LLM) drafts it best. We benchmark eight LLMs generating Gherkin from three requirement corpora (requirement lists, user stories, RFP excerpts) over 2960 generations, scoring validity, runner acceptance, judged coverage and quality, similarity to gold standard, stability, and cost. Validity is near the ceiling, yet only 78% of outputs load in the Cucumber runner: a fifth emits several Feature blocks per file. Two student annotators (a small, non-expert panel) calibrate the judge on 72 blinded generations. Human score levels are matched (error 0.33 versus 0.35 between the humans) but outputs are ordered far less reliably (ICC 0.47 versus 0.64; coverage 0.21 versus 0.76): magnitudes hold, but fine rankings do not. Against that gold standard, models span 66–107% of the human–human ceiling, ordering differently again. Pareto analysis leaves three of eight models non-dominated: cost varies 157×, judged quality 0.36 points. Per-model prompt tuning yields no cross-validated gain; a restrictive token budget truncates verbose models. Two newer models displace the low-cost front: tier-level findings transfer, and model names are dated quickly. We release the corpora, gold standard, and prototype. Model choice should weigh cost and runner acceptance over judged quality.

AI doi: 10.3390/ai7090358

Authors: Montaser N. A. Ramadan Hasan Saygin

Smart systems, the Industrial Internet of Things, and cyber-physical networks increasingly make decisions on the devices where data is generated, on nodes short of memory, compute, energy, and bandwidth, and exposed to real adversaries. Two research currents have grown to meet this: one makes artificial intelligence small and distributed (quantization, pruning, distillation, TinyML, federated and split learning), the other makes it safe (defenses against poisoning, backdoors, inversion, and evasion). This review argues that the two are entangled rather than parallel. Operators that shrink a model or scatter it across nodes also redraw its attack surface, each carrying a security dividend and a security liability, and because a node’s resources are finite and shared, model capacity and defense strength compete for one multi-dimensional budget. We formalize this as an efficiency-decentralization-security (EDS) design tension, explicitly a tension and not an impossibility, and show with published measurements that the coupling is non-monotonic. Around this thesis we build three artifacts, following an explicit design-science research process: an evidence-graded scoring matrix that separates each operator’s security dividend from its liability across seven axes and reports the direction of every effect separately from the confidence in the evidence behind it; a resource-aware threat model that judges attack and defense feasibility against a tiered device, gateway, network, and server budget with stated units; and a co-design framework whose decision workflow terminates in a defense-selection program and a verification step under adaptive attack. We work the framework through an industrial predictive-maintenance scenario with the resource arithmetic computed line by line, and evaluate it retrospectively against six published edge-AI systems. The result is a decision-support guide for building edge AI that is efficient, decentralized, and secure at once.

AI doi: 10.3390/ai7090357

Authors: Yu-Lung Chang Yen-Ching Chang

Precise segmentation of the retinal microvasculature is vital for early detection and longitudinal monitoring of systemic cardiovascular and ophthalmic diseases. However, contemporary deep learning models struggle to simultaneously preserve fine-vessel branches and suppress background noise, often leading to severe under-segmentation or false-positive artifacts. To address this persistent challenge, we propose a novel, highly efficient Triple-Path Kolmogorov–Arnold Network (KAN) optimized via Deep Mutual Learning (DML). First, the proposed architecture synergistically integrates U-Net, HaarNet, and SegNet to explicitly decouple global contextual feature extraction from high-frequency spatial edge detection. Second, we introduce field-of-view spatial constraints to strictly eliminate background interference. Third, traditional linear bottlenecks are replaced with KAN blocks, which leverage dynamic, Chebyshev polynomial-based activation functions to robustly model highly nonlinear, chaotic vascular topologies. Finally, a zero-cost DML strategy is employed during training, maximizing the network’s representational capacity without introducing any computational burden during inference. Extensive evaluations demonstrate that our proposed model successfully overcomes the traditional sensitivity-specificity trade-off. Specifically, it achieves a peak sensitivity of 72.38%, a specificity of 98.66%, and an overall Dice score of 72.48%. While maintaining exceptional computational efficiency, this tightly coupled framework establishes a robust, accurate, and clinically viable solution for automated point-of-care retinal diagnostics.

AI doi: 10.3390/ai7090356

Authors: Andreas Spilz Heiko Oppel Michael Munz

Home-based physiotherapy is performed without supervision, which leads to incorrect execution and motivates systems that assess movement automatically from inertial measurement units (IMUs). Such systems assign each repetition to a category, yet a relevant share of repetitions fall near a class boundary, where even trained raters disagree. Classifiers trained with one-hot labels collapse these borderline repetitions onto a single class and discard this ambiguity. To address this, we build on label distribution learning, which represents each repetition as a distribution over classes instead of a single label. We introduce a way to construct such distributions without a large rater pool by perturbing the thresholds of a rule-based evaluation procedure to simulate rater disagreement. We train a network to reproduce these distributions with a Kullback–Leibler objective, which we call the ambiguity approach, and compare it against a one-hot cross-entropy baseline on four IMU exercise datasets. From the predicted distribution we then determine whether a repetition is ambiguous and which classes are relevant to it. The ambiguity approach matched or exceeded the baseline classification on all four datasets and detected ambiguity and the relevant classes more reliably. Representing the label distribution in the training target therefore adds information about ambiguity at no cost to classification.

AI doi: 10.3390/ai7090354

Authors: Rakesh Ranjan Gajanan S. Kothawade Kata Sharrer Scott Tsukuda Christopher Good

The You Only Look Once (YOLO) has been widely adopted in aquaculture monitoring and management due to its real-time performance and deployment flexibility. The recently introduced YOLO26 architecture incorporates Non-Maximum Suppression (NMS)-free end-to-end inference and is optimized for deployment on resource-constrained CPU-based devices, making it particularly relevant for edge deployment in commercial aquaculture applications. Nevertheless, its performance, operational efficiency, and deployment suitability compared with previous YOLO generations remain largely unvalidated in aquaculture-specific scenarios. This study benchmarks YOLO26 against three Ultralytics predecessors (YOLOv5u, YOLOv8, and YOLO11) across nano, small, and medium model scales for the detection of fish mortality, a critical indicator of fish population health and welfare, in recirculating aquaculture systems (RAS). Twelve model variants were evaluated for detection accuracy, training efficiency across seven dataset sizes, and inference performance on both high-performance NVIDIA A100 GPUs and the resource-constrained, CPU-only Raspberry Pi 5 edge device. All models achieved comparable performance on the full dataset, with mAP50 varying by only 1.25 percentage points across three independent training runs, indicating minimal influence of architectural generation on final mortality detection accuracy when sufficient training data are available. However, notable differences emerged in data efficiency and deployment performance. YOLOv8 demonstrated the strongest training efficiency, achieving 90% mAP50 with only 400 training images, whereas YOLO26 nano and small variants required 1000 images to reach comparable accuracy. In contrast, YOLO26 exhibited advantages during edge deployment, with YOLO26n achieving the highest inference speed on the Raspberry Pi 5 at 7.84 ± 0.13 FPS across three benchmark sessions, while YOLOv5mu outperformed all contemporary medium-scale architectures on CPU-based hardware. These results demonstrate that architectural novelty alone is an insufficient criterion for model selection. The findings support a deployment-oriented framework in which training data availability, target hardware, and inference requirements collectively inform model selection for aquaculture applications.

AI doi: 10.3390/ai7090355

Authors: Daniele Giansanti

Background: Smartphones and generative artificial intelligence (GenAI) are increasingly integrated into everyday life, influencing information access, digital interaction, and cognitive tasks. While problematic smartphone use and cognitive offloading are well studied, GenAI introduces newer forms of cognitive delegation that warrant comparison. Objective: This narrative review examines convergences and divergences between smartphone- and GenAI-mediated cognitive processes, focusing on cognitive offloading, attention, reliance, self-regulation, and human agency. Review approach: Interdisciplinary literature from psychology, artificial intelligence, digital technologies, and education was narratively reviewed and compared. The aim was to integrate relevant evidence rather than provide a systematic or exhaustive synthesis. Synthesis: Smartphones primarily support connectivity, information access, and digitally mediated attention, whereas GenAI extends technological assistance toward reasoning, synthesis, and content generation. Despite important differences, both may redistribute cognitive effort between individuals and digital systems, with potential implications for self-regulation and human agency. Conclusions: Smartphones and GenAI are distinct technologies with partly overlapping but also technology-specific effects. Further empirical research is needed to determine which forms of cognitive offloading and reliance are shared across technologies and their implications for human agency.

AI doi: 10.3390/ai7090353

Authors: Jindae Kim

Test-time refinement aims to improve generated programs through additional inference, but its value after an initial candidate has been produced remains unclear. We conduct a controlled evaluation of one-round Self-Refine and Self-Debug across seven models and three Python code-generation benchmarks. For each model and task, both methods refine the same initial candidate, allowing us to measure refinement gain without variation in initial generation. Self-Debug produces positive refinement gain in 16 of the 21 model–benchmark combinations and no change in the remaining five, whereas Self-Refine reduces correctness in 16 combinations and improves it in only four. Decomposing refinement gain into repairs and regressions clarifies this contrast. After a diagnostic pass, Self-Debug’s candidate preservation leaves 34.1% of initial final-test failures unaddressed, but 99.95% of initially correct candidates remain correct. With regressions nearly absent, repairs after diagnostic failures translate directly into positive gain. Self-Refine also repairs initial failures, but its regression count is more than three times its repair count overall, producing predominantly negative gain. Resource analysis shows that Self-Refine uses more tokens while generally reducing correctness, whereas Self-Debug provides a more favorable gain–overhead balance, although its inference overhead per net additional pass varies across models and benchmarks. These results show that one-round refinement is not inherently beneficial and should be applied only when its expected gain justifies the additional computation and cost.

AI doi: 10.3390/ai7090352

Authors: Alkis Gounaris George Kosteletos

This chapter offers a conceptual and philosophical analysis of the ontological, epistemological, and ethical issues associated with the design, development, and use of Agentic AI in moral and legal decision-making. Its aim is to map, classify, and bring into focus the principal philosophical problems raised by the autonomous operation of artificial agents in contexts of moral and legal judgement. The chapter contributes to the existing literature by distinguishing between two levels of AI autonomy. Advisory systems possess limited autonomy and provide recommendations that remain subject to human evaluation, whereas regulatory systems exercise a greater degree of autonomy and are entrusted with final decision-making authority. Adopting a sceptical perspective, the analysis examines the conceptual fragility and epistemological difficulties surrounding proposals to employ Agentic AI in significant domains. Particular attention is given to the risk that ostensibly advisory systems may become tacitly regulatory in practice, especially under the influence of widespread assumptions concerning the objectivity, accuracy, and effectiveness of AI. The inquiry proceeds through a step-by-step branching structure organised around three central questions concerning the moral values on the basis of which such systems should be designed and developed, the extent to which they can understand the concepts of ethics and justice, and the attribution of responsibility for their decisions and outputs. Each question opens onto alternative lines of analysis, thereby providing a systematic map of the principal philosophical challenges involved. Its purpose is to provide a synthetic overview of the relevant philosophical issues while highlighting both their breadth and their plurality. To organise these issues systematically, the chapter adopts the schema of an algorithm, offering a clear and ordered account of the possible interconnections among the problems examined and of the ways in which alternative answers generate distinct yet interrelated lines of philosophical inquiry.

AI doi: 10.3390/ai7090351

Authors: Zainab A. Jawad Ahmed J. Obaid

The detection of audio deepfakes has emerged as a significant problem in the field of voice biometrics systems, aiming to distinguish real human voices from those generated by Artificial Intelligence (AI). With synthetic voice becoming increasingly high-quality, it is more likely that such a voice will be abused for illicit purposes like identity theft and impersonation. The dual-branch CNN with shared weights architecture presented here is augmented with self-attention modules to detect audio deepfakes with greater efficiency. Convolutional operations and dual branches are used to extract complex characteristics from raw audio signals in our module in order to directly compare the unprocessed original audio with the modified audio. Afterward, residual connections improve network performance. Designed alongside these fundamental layers, self-attention modules are trained in a layered manner to detect multi-headed attention within audio frames. This feature helps the network distinguish between original and modified audio and improves feature extraction compared with the standard method. A range of audio modifications have been analyzed to assess the effectiveness of the method, and comprehensive testing across all possible audio manipulation situations has been conducted on the Controlled Singing Voice Deepfake Detection Challenge (CtrSVDD) dataset to assess its resilience. Both deep learning (DL) and machine learning (ML) models were outperformed by the proposed dual-branch CNN with shared weights. With an accuracy of 97.26%, precision of 99%, recall of 99.27%, and an F1 score of 98.88%, this model has achieved a remarkable performance.

AI doi: 10.3390/ai7090350

Authors: Stavros Kalogiannidis Dimitrios Syndoukas Konstantinos Spinthiropoulos Dimitrios Parris George Konteos

The rising number and the complexity of crises in Europe, including the COVID-19 pandemic, migration, and cybersecurity threats, have accelerated the adoption of digital transformation and artificial intelligence (AI) in public administration. This paper is a systematic literature review (SLR) of 58 academic sources examining the effect of digital transformation and AI on crisis response outcomes in European public administration. Records were identified in Scopus and Web of Science, complemented by supplementary sources, and critically appraised through a two-step procedure based on the Mixed Methods Appraisal Tool. The review integrates major themes such as the evolution of digital governance, AI-enabled decision-making, performance in crisis management, and institutional resilience. Gains in efficiency, coordination, and responsiveness are widely reported, but the appraisal shows that they rest on a thin evidential base—one quasi-experimental study, no longitudinal designs, and twenty non-empirical records—while governance, accountability, ethical, and institutional-capacity problems persist. The article reveals significant gaps in research and suggests a broad research agenda in the future with emphasis on theoretical, methodological, and policy aspects. The paper contributes to the emerging digital-era governance literature and offers recommendations for policymakers and researchers.

AI doi: 10.3390/ai7090349

Authors: Joseph Geraci Bessi Qorri Christian Cumbaa Mike Tsay Christopher Alexander Marrella Seb Zappulla Paul Leonczyk Adam Gogacz Luca Pani

Small, heterogeneous clinical datasets pose a challenge for whole-cohort prediction because clinically meaningful treatment or response patterns may be diluted across biologically diverse patients. We describe and evaluate NetraAI, an interpretable dynamical-systems framework for selective subgroup discovery that uses finite-iteration contraction-inspired dynamics and long-range memory (LRM) to identify stable, outcome-linked Model-Derived Subgroups (MDS). This system can abstain by assigning No Call when a stable subgroup assignment is not supported. A large language model (LLM) Strategist is outlined only as a possible future extension; it is not evaluated here and contributes nothing to the results reported. Foundation and language models asked to perform subgroup discovery directly did not recover the structure the specialized discovery step recovered. We demonstrate this framework across three retrospective clinical trial datasets: Clinical Antipsychotic Trials of Intervention Effectiveness (CATIE) schizophrenia (olanzapine vs. perphenazine comparative treatment-preference benchmark), Canadian Biomarker Integration Network in Depression (CAN-BIND) depression (escitalopram response), and Comprehensive Molecular Characterization of Advanced Pancreatic Ductal Adenocarcinoma for Better Treatment Selection (COMPASS) pancreatic cancer (GnP vs. FOLFIRINOX observational regimen-associated response). The benchmark is not a contest between NetraAI and competing predictors: the same eight downstream methods are evaluated with and without what NetraAI discovered. Given the full feature sets and their own selection procedures, those methods were at or near chance on all three datasets, and blind de novo searches by an independent interaction model and by a pretrained tabular foundation model did not recover an equivalent signature or subpopulation. In internal downstream evaluation, given the discovered variables alone—the same patients, the same classifiers, the full cohort and no abstention of any kind—every one of the eight methods improved on every dataset, 24 of 24 method-dataset comparisons, moving from a raw-feature range of 0.46–0.62 AUC to 0.54–0.78. Restricting further to the subpopulation in which those variables hold improved all eight methods again in CAN-BIND and in COMPASS, 16 of 16 comparisons, reaching 0.66–0.83 and 0.96–1.00, respectively; in CATIE, where the called subgroups are the least outcome-homogeneous of the three, it improved only one of eight. Taking the framework as a whole, 23 of 24 method-dataset combinations improved over the raw-feature baseline. NetraAI abstains on patients without stable subgroup structure, calling 27.7% to 40.4% of each cohort. The contribution demonstrated is therefore subgroup discovery rather than downstream prediction, and its beneficiaries are the conventional methods themselves. In COMPASS, NetraAI identified a three-SNV signature associated with regimen-linked response ranking among called patients; because the cohort was observational and the permutation test was conditional on the selected signature, this finding is exploratory. The two mechanisms are complementary rather than competing: variable discovery establishes which features carry the structure, and abstention establishes in which patients it holds. Neither mechanism replaces conventional modeling. Variable discovery improved every method on every dataset, the pretrained tabular foundation model included; identifying the population in which those variables hold conferred further benefit in two of the three datasets and not in the third. These findings support NetraAI as an exploratory system for generating compact, inspectable subgroup hypotheses that may inform future enrichment strategies after external validation, and indicate that its value lies in what it contributes to other methods rather than in competing with them.

AI doi: 10.3390/ai7090348

Authors: Zhixiang Zhang Qiang Tang Zhongshen Zhang Xubin Feng Meilin Xie

Different imaging modalities exhibit inherent discrepancies in intensity characteristics and information representation, resulting in pronounced heterogeneity among multimodal features. When these heterogeneous features are directly learned and fused within a unified representation space, feature coupling may arise, leading to mutual interference between global structural information and local fine-grained details. To address the limited differentiated modeling of structural and fine-grained information in existing methods, we propose Hierarchical Dual-Perspective Collaborative Modeling for Multimodal Image Fusion (HDPF). HDPF employs a Dual-Perspective Feature Aggregation Block (DFAB) to jointly exploit convolution-based local representation and Transformer-based global contextual modeling. Building upon this dual-perspective representation, HDPF further constructs two differentiated pathways dedicated to structural information and detail information, respectively, thereby providing differentiated representations of complementary multimodal information. The extracted features are subsequently reorganized and integrated by the decoder to reconstruct the final fused image. Extensive experiments are conducted on multiple infrared–visible image fusion (IVF) datasets as well as medical image fusion (MIF) tasks. On the TNO dataset, HDPF achieves VIF and MI scores of 0.80 and 3.50, respectively. On the MRI–PET fusion task, the VIF score reaches 0.61. The experimental results demonstrate that HDPF achieves competitive and relatively balanced fusion performance across the evaluated multimodal image fusion tasks.

AI doi: 10.3390/ai7090347

Authors: Juncal Uriol Emma O’Brien Iker Hernández Roberto Viola Eneko Iradier Jon Montalbán

Agentic artificial intelligence (AI) systems are increasingly deployed across distributed cloud–edge–radio infrastructures, where autonomous agents make decisions with direct operational and economic impact. Smart contracts (SCs) enforce business rules and policies, constraining autonomous actions according to predefined operational intent. As agent autonomy expands across heterogeneous networks, ensuring accountability requires transparency, verifiability, and compliance with SC-defined requirements. To address these challenges, this paper proposes the Agent Governance Framework (AGF), which integrates SC-based governance into agentic AI systems, enabling verifiable accountability through traceable autonomous decisions. Built on European Telecommunications Standards Institute (ETSI) and TM Forum principles, AGF comprises six components: (i) a TM Forum-aligned business support system (BSS); (ii) an agentic AI-enhanced operations support system (OSS); (iii) a Network Resource Operations subsystem; (iv) an Identity and Access Role Manager; (v) a Distributed Marketplace; and (vi) a Traceability and Auditability Registry. Together, these components provide an end-to-end (E2E) traceable architecture, linking each action to the responsible agent and SC. A prototype on a local test network with the evaluation of two complementary network test cases demonstrates the framework’s feasibility, confirms its full traceability and immutability, and highlights AGF’s potential as a foundation for reliable, large-scale agent-based systems in telecommunications networks. The results demonstrate 100% traceability across the E2E governance loop and a minimal latency overhead of 2–4% due to the Traceability and Auditability registry.

AI doi: 10.3390/ai7090346

Authors: Sumayyamol Mukkil Muhammed Ismail Mobyen Uddin Ahmed Shahina Begum

Zero-day network attacks pose a significant threat because their unknown signatures evade traditional detection mechanisms. This research develops an AI-enhanced intrusion detection system that aims to detect such attacks while providing interpretable outputs for security analysts. Four machine-learning models are evaluated under strict zero-day conditions using two benchmark datasets. SHAP and LIME are applied to produce instance-level explanations, and a formal stability assessment is conducted to determine their reliability. Experimental results show that the hybrid model combining anomaly-based detection with deep learning achieves the highest zero-day Recall, with statistically significant advantages over individual models in detecting previously unseen attacks, while the standalone LSTM achieves the strongest overall balance between Precision and Recall. The generated explanations consistently reveal security-relevant features, and stability analysis confirms their robustness across conditions. The study demonstrates that integrating deep learning with stable explainable AI offers a practical and trustworthy solution for zero-day intrusion detection, contributing validated evidence to an area where explanation reliability is rarely examined.

AI doi: 10.3390/ai7090345

Authors: Alexander Ivannikov Igor Temkin Ilya Savelev

Predicting emergencies caused by uncontrolled and sometimes sudden changes in methane concentration within working and adjacent zones of coal mines remains a critical and challenging task, the solution for which can greatly enhance mining safety. This study presents a hybrid machine-learning model trained on real and synthetic data for accurate methane concentration forecasting and risk-level classification. The authors propose an ensemble method comprising staged data preprocessing, generation of physically meaningful features, and weighted ensembles for both regression and classification. The system is augmented with expert rules to correct forecasts and a built-in anomaly detection mechanism based on residual analysis. Experimental evaluation confirmed the model’s high performance: for regression, the coefficient of determination reached 0.984–0.997; the classifier achieved a recall of 92.8% for the rare “Accident” class under severe data imbalance (10:1). The ensemble approach reduced error variance by 40–60% compared to baseline models. The results indicate the feasibility of pilot application for dynamic early warning, which can substantially reduce coal mine accident risks.

AI doi: 10.3390/ai7090344

Authors: Qian Wang Jiaqi Feng Yu Zhou Miao Li Zhi Zhang Luyao Wang Wenping Song

Existing deepfake detectors often perform well on in-domain data but generalize poorly to unseen datasets or manipulation methods. This limitation is largely attributed to their reliance on dataset-specific semantic cues rather than transferable forgery patterns. To address this limitation, we propose a generalizable deepfake detection framework that combines frequency-domain enhancement with feature disentanglement. A Phase-Amplitude Frequency Enhancement (PAFE) module enhances subtle spectral artifacts introduced during deepfake generation. We then feed the enhanced representations into an asymmetric dual-branch architecture that separates content-related information from forgery-related features. The content branch models facial semantics, while the forgery branch extracts discriminative forgery features with reduced content interference. A spatial self-attention module further refines the forgery features. We optimize the framework using image-level reconstruction loss, feature-level contrastive loss, and classification loss. Together, these objectives encourage effective feature disentanglement and improve the discriminability of the learned forgery features. Extensive experiments on several widely used deepfake benchmarks show that the proposed framework achieves competitive detection performance and improved cross-domain generalization compared with existing methods.

AI doi: 10.3390/ai7090343

Authors: Stefan Alecu Gheorghe Adrian Onea

Artificial intelligence is increasingly used in sport science to classify movement, analyse tactics, estimate readiness, forecast performance and injury risk, and support rehabilitation. Predictive performance alone, however, does not establish practical usefulness, safety, or responsible use. Through purposive source selection and critical synthesis, this review integrates sport-specific evidence with guidance on prediction modelling, human–AI interaction, and AI risk management. It highlights five recurring gaps: limited data and labels; leakage-prone or temporally inappropriate validation; weak reporting of calibration and uncertainty; unclear translation from prediction to action; and insufficient attention to athlete rights and post-deployment monitoring. We propose a six-layer framework covering data and context, prediction, decision translation, individualisation, human oversight, and deployment monitoring. For each applicable layer, the framework specifies evidence to document, appraisal questions, and common failure modes. It distinguishes model development from operational decision support and links technical evaluation with workflow utility, fairness, privacy, accountability, contestability, and lifecycle governance. The framework is conceptual, not a validated scoring or certification instrument. Sport-science AI should be evaluated as a socio-technical intervention rather than an isolated algorithm. Future studies should prioritise athlete-level and temporal validation, calibrated probabilities, external and prospective evaluation, prespecified action pathways, subgroup performance, documented override procedures, and ongoing monitoring.

AI doi: 10.3390/ai7090342

Authors: Raza Hasan Vishal Dattana Salman Mahmood

The effective resolution of non-deterministic polynomial time hard (NP-hard) combinatorial optimization problems requires a delicate balance between global exploration and local exploitation. While Quantum-Inspired Algorithms (QIAs) leverage principles of superposition to explore vast search spaces, they often lack the fine-grained exploitation capabilities of classical heuristics. To address this limitation, we propose the Quantum-Memetic Hybrid Algorithm (QMHA), a component-based framework that synergistically integrates qubit-based global search with adaptive classical refinement. The QMHA architecture explicitly coordinates five distinct algorithmic components: (1) quantum rotation gates for exploration, (2) a problem-aware memetic operator for immediate solution refinement, (3) an adaptive learning rate schedule, (4) periodic local search, and (5) a stagnation-based population reset for diversity management. We rigorously evaluate the framework against nine established metaheuristics, including Genetic Algorithms (GA), Differential Evolution (DE), Particle Swarm Optimization (PSO), Simulated Annealing (SA), Ant Colony Optimization (ACO), MAX-MIN Ant System (MMAS), Memetic Algorithms (MA), Quantum Evolutionary Algorithm (QEA), and Harmony Search (HS), across a comprehensive benchmark suite comprising six NP-hard problem families: constrained combinatorial (Knapsack), graph-based (Max-Cut), permutation-based (TSP), constraint satisfaction (Graph Coloring), bin optimization (Bin Packing), and scheduling (Flow Shop Scheduling), as well as real-world machine learning (Feature Selection) problems and the continuous Congress on Evolutionary Computation (CEC) 2022 benchmark. Statistical analysis using Friedman tests and Nemenyi post hoc comparisons confirms that QMHA achieves a statistically significant performance advantage (p<0.004) and superior average rank (1.5) compared to component baselines and state-of-the-art competitors. Comprehensive analyses include computational complexity profiling, parameter sensitivity mapping, scalability testing up to D=2000, noise robustness evaluation, variable correlation degradation analysis, a six-component ablation study, exploration–exploitation dynamics tracking, integration mechanism comparison across five architectures, and a multi-objective extension feasibility study. The proposed framework offers a robust, verified approach to hybrid optimization without relying on biological metaphors.

AI doi: 10.3390/ai7090341

Authors: Fars Samann Thomas Schanze

The standard 12-lead ECG is the gold-standard tool for cardiac diagnosis and monitoring; however, its multi-electrode configuration limits its use in prolonged monitoring and wearable devices. Existing portable ECG devices typically record only a single lead, reducing their diagnostic value. To overcome this limitation, this study proposes novel Multi-Input Multi-Output neural networks (MIMO-Nets) that reconstruct a standard 12-lead ECG from a single raw ECG lead without requiring any feature extraction techniques. Four deep learning architectures were investigated: MIMO-Conv-Net, MIMO-U-Net, Conv-Net, and U-Net. Each model was trained and evaluated using all 12 standard ECG leads with different length L as candidate inputs to determine the optimal single-lead acquisition strategy. Results show that Lead II with L = 256 consistently provides the highest reconstruction performance across all architectures, making it the most informative single-lead input. Using Lead II as input, MIMO-U-Net yielded the highest reconstruction accuracy on a processed clinical dataset (ST-Petersburg INCART), with an average correlation of 0.864 at L = 256, closely followed by U-Net (0.862). On raw, unprocessed recordings from the PTB-XL database, both models achieved lower but still competitive correlations (0.740 and 0.732, respectively). These findings demonstrate the feasibility of single-lead-to-12-lead ECG reconstruction and highlight the role of signal preprocessing in reconstruction fidelity, supporting the development of wearable ECG systems for continuous cardiac monitoring.

AI doi: 10.3390/ai7090340

Authors: Bhagiradh Kantheti Carlos A. Paz De Araujo

Transformer architectures, originally developed for language modeling, have been increasingly applied to time-series forecasting, but deploying them on sensitive data such as patient vital signs, financial records, or energy use raises privacy risks at two points: inference (the server sees raw inputs and outputs) and training (the model can memorize individual records). This survey organizes privacy-preserving methods for Transformer forecasting along a two-axis taxonomy: protection stage (inference, training, and data-level) and protection locus/mechanism (cryptographic wrapper, gradient, input, feature, output, or data path). For inference, we review secure multi-party computation (CrypTen through MPCFormer, SecFormer, Iron, BOLT, PUMA, BumbleBee, and CipherPrune) and homomorphic encryption (CKKS, THE-X, hybrid HE-MPC, and automated FHE compilation via Orion), focusing on how each handles the Transformer non-linearities (softmax, GELU, and LayerNorm). For training, we organize differential-privacy mechanisms by where noise enters the gradient path (DP-SGD, DP-FedAvg, DP-FTRL, PATE, and adaptive clipping), the input path, and the more recent feature path and analyze the privacy, utility, and dimensionality trade-offs. We further survey privacy in large language and foundation models, federated Transformers, time-series privacy attacks (membership inference, gradient inversion, and attribute inference), differentially private synthetic data, and applications in energy, health care, and finance, closing with open problems and future directions. Throughout, we show that the surveyed mechanisms provide non-comparable privacy guarantees and should be chosen by threat model instead of headline utility.

AI doi: 10.3390/ai7090339

Authors: Wei Zhang Hui Fang Jiajun Bu Qiming Zhong Yueyao Yu

Urban rail transit systems require reliable decision-support methods to assist operators in handling complex fault events. Although large language models provide strong reasoning and knowledge integration capabilities, their open-ended generation may produce responses that are difficult to align with operational procedures. To address this issue, we propose a role-aware agentic retrieval-augmented generation framework for transit fault management. The framework integrates keyword-based retrieval, role-aware multi-agent reasoning, and constrained action generation within a unified architecture. Operational procedures and historical cases retrieved from transit knowledge repositories provide evidence for agent reasoning, while constrained outputs ensure that generated instructions remain consistent with predefined operational actions. Experimental results on a real-world urban rail transit emergency exercise dataset demonstrate that the proposed framework improves the semantic score from 0.893 to 0.965 and increases Role F1 score from 0.824 to 0.915 compared with an Agentic Retrieval-Augmented Generation (RAG) baseline while maintaining comparable inference efficiency. These results demonstrate the potential of the proposed framework for decision support in urban rail transit fault scenarios.

AI doi: 10.3390/ai7090338

Authors: Priyanka Vyas Sheetal U. Bhandari Pramod R. Sonawane

Solar energy makes a considerable contribution to global power generation, necessitating photovoltaic (PV) fault detection to ensure high yields in solar power systems. In real-world solar installations, operational data are geographically dispersed, heterogeneous, and sensitive, which imposes privacy restrictions. Existing PV fault-detection techniques have relied on centralized training, requiring raw image data from multiple plants to be collected at a single server, which has led to privacy risks, bias from heterogeneous datasets, and scalability issues. To overcome these challenges, this research provides a novel FL-based PV fault-detection model called Federated Convolutional Transformer Network (Fed-CVTNet), which combines Convolutional Neural Network (CNN) and Vision Transformer (ViT) architecture within the FL framework. Initially, the raw images are pre-processed to enhance the input quality, and the Region of Interest (ROI) is identified via a pretrained YOLO model. Then, the proposed Fed-CVTNet facilitates networked learning among many geographically dispersed clients by only sharing model updates using federated averaging (FedAvg). The experimental findings illustrate that the proposed FL model shows substantial quality improvements compared to the customized CNN-ViT models trained on a dataset of 5600 images and the CNN-ViT models that lack federated aggregation. The highest accuracy achieved by the proposed technique is 98.99%; the proposed Fed-CVTNet has better sensitivity (98.15%) and specificity (99.21%), and much lower false positive and false negative rates than its centralized counterparts. The comparative analysis establishes that federated weight aggregation outperforms centralized baseline models by 2.3% and is effective in reducing the data privacy risk.

AI doi: 10.3390/ai7090337

Authors: Cátia Simões João Viana João Couvaneiro Luís Proença Vanessa Machado Fang Hua Naichuan Su José João Mendes João Botelho

Artificial intelligence (AI) is increasingly being incorporated into oral healthcare and dental education, yet the quality and effectiveness of randomized evidence supporting these interventions remain uncertain. We systematically reviewed randomized controlled trials (RCTs) evaluating AI-based interventions in oral healthcare and dental education. A comprehensive literature search was conducted in PubMed, Embase, ClinicalTrials.gov, and OpenGrey from database inception to 30 June 2026. Twenty-eight published RCTs involving 2306 participants and 26 registered unpublished RCTs were identified. Educational applications represented the largest category (53.6%), followed by diagnosis and clinical decision support, treatment planning, and disease monitoring. Across the included RCTs, AI-based interventions were more frequently associated with improvements in educational performance, diagnostic accuracy, treatment planning, clinical decision-making, and periodontal monitoring than conventional approaches, although findings were heterogeneous, several studies reported no significant differences, and a small number favored conventional methods. No serious adverse events were reported, but safety reporting was inconsistent. Most published RCTs presented some concerns regarding risk of bias (82.1%), primarily related to randomization, deviations from intended interventions, and selective reporting. APPRAISE-AI demonstrated predominantly moderate methodological quality (median score 44%), with recurrent weaknesses in robustness, reproducibility, and data quality. The identification of 26 registered RCTs without published results also suggests a potential risk of dissemination bias. These findings indicate that AI-based interventions show promise across multiple applications in oral healthcare and dental education, but improvements in trial methodology, reporting quality, and transparency are needed to strengthen the evidence base and support clinical implementation.

AI doi: 10.3390/ai7090336

Authors: Benyamin Mirab Golkhatmi Mohammad Hossein Moattar

Accurate localization of colon polyps plays a crucial role in the early diagnosis of colon cancer. In this paper, we propose a novel network architecture, X-Net, which integrates the Pyramid Vision Transformer (PVT) and EfficientNet-B0. We selected EfficientNet-B0 due to its optimized design and computational efficiency, enabling it to effectively capture fine-grained local features. Meanwhile, we use PVT V2-b1 (Pb1) because of its hierarchical structure and self-attention mechanism, which enhance the model’s ability to preserve multi-scale features. By concatenating these two networks in part of the decoder, the proposed model effectively leverages both local and global features. Additionally, the X-Net architecture introduces three blocks: Shape-Aware Enhancement Block (SAEB), Multi-Scale Feature Block (MSFB), and Multi-Aspect Attention Block (MAAB). The SAEB utilizes differences in initial-layer features extracted from the encoder to filter noise in polyp images. The MSFB extracts and enhances multi-scale features from the backbone, effectively integrating semantic information, which improves polyp segmentation accuracy and enhances robustness against variations in polyp size and shape. Finally, the MAAB employs high-level features from the outputs of the MSFB and SAEB to attend to the foreground, background, and boundaries. To address the class imbalance problem, we propose a hybrid loss function combining Tversky Loss and Binary Cross-Entropy Loss. For evaluation, we trained the proposed network on four polyp datasets: Kvasir-SEG, CVC-ClinicDB, CVC-T, and ETIS. The results demonstrate that the model performs well across different datasets with varying image quality and characteristics.

AI doi: 10.3390/ai7090335

Authors: Yu-Lin Zhang Ming-Yang Qian Chen-Yang Wang Zhan-Heng Chen

Computational drug repurposing increasingly integrates chemical, biological, omics, network, text, and clinical data through deep learning. This structured narrative review examines how such modalities are encoded, aligned, and fused. We organize representative studies into four mechanism-centered families: heterogeneous-graph neural networks, multimodal knowledge-graph embeddings, pretrained language/sequence model-based cross-modal alignment, and multi-view or reconstruction-based fusion. Direct drug–disease association and repurposing studies form the core evidence; drug–target interaction, drug–drug interaction, target-identification, molecular-pretraining, and drug–microbe studies are treated as adjacent methodological evidence. We compare architectures, evaluation settings, failure modes, and evidence levels across oncology, neurology, infectious, and rare diseases. Practical guidance covers leakage-aware random, cold-start, temporal, and cluster-based evaluation; an actionable reproducibility checklist; and a scenario-based model-selection framework. We distinguish computational prioritization, docking, preclinical, retrospective clinical, and prospective evidence, and examine data sparsity, uncertain negatives, missing or noisy modalities, interpretability, and translational limitations. Future priorities include temporal and causal evaluation, external and multi-center validation, federated learning, and emerging therapeutic modalities. Multimodal fusion can improve complementary representation, but its value depends on task definition, data quality, evaluation design, and independent validation.

AI doi: 10.3390/ai7090334

Authors: Li Lin Minghao Shi Tenglong Wang

Long-horizon robotic manipulation requires a policy to bridge task-level semantic reasoning with metric three-dimensional interaction geometry. Existing vision–language–action policies usually acquire geometry implicitly from visual tokens or introduce deterministic intermediate variables only in the image plane, which rely on expensive human annotations and training cost. This article presents a spatial Gaussian-guided hierarchical framework that uses ordered 3D Gaussian interaction regions as an explicit planning interface between vision–language reasoning and action generation. The proposed framework enables efficient adaptation of a pretrained vision–language model for robotic manipulation tasks. First, an automatic geometric enhancement pipeline converts raw robot demonstration videos into near-, mid-, and late-stage Gaussian supervision through foreground extraction, metric depth estimation, stable camera aggregation, end-effector localization, 3D lifting, and temporal grouping, without requiring manual 3D interaction annotation. The generated Gaussian representations provide structured spatial guidance, where their covariance characterizes interaction-region extent and variability rather than fully calibrated physical uncertainty. Second, a shared vision–language backbone predicts structured subtasks and Gaussian interaction regions, while a conditional diffusion executor generates future action chunks under these semantic and geometric conditions. A trajectory-to-Gaussian likelihood objective explicitly encourages consistency between generated motions and the predicted spatial interaction plan. Experiments on a mixed real-robot dataset derived from LHManip and RH20T show that our method improves trajectory tracking success from 55.7% to 70.8% over a same-backbone direct VLA baseline. Closed-loop simulation evaluation on LIBERO with 80% backbone parameter frozen achieves 85.3% average task success, demonstrating the effectiveness of explicit 3D interaction representations for spatial reasoning and long-horizon manipulation.

AI doi: 10.3390/ai7090333

Authors: László Vancsura

The cost–performance relationship in mutual funds is a longstanding open question in financial economics, particularly when costs are assumed to exert a single, linear effect on returns. This study proposes an integrated causal machine learning framework to revisit this question using a panel of Hungarian open-ended public investment funds across all major asset classes—equity, bond, absolute yield, misc, money market, real estate, and commodity—covering 2017–2024. Six machine learning algorithms are benchmarked for return prediction, and Double Machine Learning, with fund-level cluster-robust inference and year fixed effects, is applied to estimate the effect of the Total Expense Ratio (TER) on next-year returns, under the identifying assumptions stated in the paper, while flexibly controlling for a set of observed fund-level confounders (size, NAV dynamics, volatility, past and cumulative performance, and fund age) without imposing a linear functional form. To move beyond average effects, a Causal Forest model—tuned using an out-of-fold, effect size-neutral selection criterion—estimates heterogeneous treatment effects across funds, and SHAP-based interpretation uncovers the mechanisms underlying this heterogeneity. The results show that, once the outcome is measured in the year following the one in which TER is observed and panel dependence is properly accounted for, the average TER effect is not robustly different from zero at the full-sample level; where a statistically robust effect emerges, it is negative rather than positive, concentrated in equity and absolute-yield funds, and largely confined to the period after 2022, which coincided with the war in Ukraine, rising interest rates, and heightened market volatility, although the research design does not identify which, if any, of these developments drove the change. Average-effect models are shown to conceal this heterogeneity, and the results are further shown to be sensitive to two methodological choices that might otherwise appear secondary—the timing convention linking cost and return, and the criterion used to select among competing heterogeneous-effects specifications—underscoring the importance of making such choices explicit. These findings demonstrate the added value of combining predictive and causal machine learning, together with identification-robust and panel-robust inference, for uncovering heterogeneity that conventional econometric approaches overlook and offer a transferable methodological template for causal machine learning applications in finance and other high-dimensional decision-making domains.

AI doi: 10.3390/ai7090332

Authors: Xinyuan Zhang Rui Zhao Tianyue Sun Yonghong Xu

Driving-fatigue-induced attentional decline and response retardation are critical contributors to traffic accidents. However, stably and precisely identifying fatigue states from noisy electroencephalogram (EEG) signals remains a challenging issue in intelligent driving safety. To address the dual deficiencies of traditional methods in fatigue feature extraction precision and noise robustness, this paper innovatively constructs a collaborative recognition framework that integrates an Improved Strip Attention Mechanism (ISAM) with a Deep Residual Shrinkage Network (DRSN). The core innovations of this framework are twofold: ISAM achieves precise localization and focused enhancement of fatigue-related rhythmic bands in EEG signals via row–column separable adaptive pooling and channel-wise attention augmentation; concurrently, the DRSN module introduces an improved soft-thresholding function, which adaptively generates filtering thresholds through channel attention to effectively suppress noise and artifact interference in physiological signals. The deep fusion of these two modules forms a closed-loop optimization chain of “targeted feature reinforcement–adaptive noise suppression,” enabling the model to stably extract highly discriminative fatigue representations from complex non-stationary EEG signals. Validation on two public datasets, SEED-VIG and SADT, demonstrates that the proposed method achieves recognition accuracies of 98.86% and 97.38%, respectively, outperforming mainstream methods such as the convolutional spatial-frequency network and multi-scale convolutional neural network by 17.38% and 17.76%. These results confirm the significant advantages of the proposed dual-module collaborative architecture in precise fatigue characterization and anti-interference capability, offering a highly reliable technical solution for real-time driver mental fatigue monitoring in real-world road scenarios.

AI doi: 10.3390/ai7090331

Authors: Ali Akbar ForouzeshNejad Alexander Gegov

Agile outcome models commonly map Jira variables directly to a retrospective label and then explain the prediction through fragmented feature attributions; they rarely separate domain concepts, team clustering, unresolved work, and concept-label coupling. This study evaluates a domain-informed, concept-bottleneck-style explainable AI architecture for retrospective diagnosis of Agile Epic outcomes. A frozen Jira export of 10,000 unique issue-level records was linked to a pre-specified analytical cohort of 180 Epics across 14 teams. Six experts rated efficiency, effectiveness, sustainability, and contextual risk, while outcomes were recorded as Successful, Challenged, or Unsuccessful. Because the outcome labels and concept ratings were informed by the same Jira evidence, the models estimate consistency with an expert labelling procedure, rather than independent project success. Under five-fold group-aware cross-validation, the fixed-configuration flat LightGBM achieved macro-F1 = 0.864 ± 0.053 and the fixed-configuration HMXAI/CBM-style model achieved 0.843 ± 0.084. These descriptive primary scores are not a joint nested-model-selection comparison. The proposed method, therefore does, not demonstrate a performance improvement; its contribution is an inspectable diagnostic structure. Performance fell materially on the resolved-only subset (LightGBM macro-F1 = 0.645), and model-specific nested, leave-one-team-out, calibration, uncertainty, correlation, and intervention analyses further bound the claims. Concept interventions were not uniformly monotone, so the concept layer is domain-interpretable in form but not yet user-validated as actionable. The study contributes a transparent audit of when concept-level diagnosis can complement flat classification and when circularity, censoring, and shortcut learning restrict interpretation.

AI doi: 10.3390/ai7090329

Authors: Dalibor Nikolić Ivica Djalović Ivan Vitezović Dejan B. Stojanović Sara Pavkov Rastislav Stojsavljević Mlađen Jovanović

Benchmarking deep learning forecasters against classical and physically based numerical baselines remains uncommon in the time-series forecasting literature. Meteorological station networks offer an under-exploited evaluation environment, uniquely providing a physically based climate-model comparator alongside standard baselines. We evaluated eight forecasting approaches—climatology, SARIMA, Random Forest, and five deep learning architectures (TFT, N-HiTS, PatchTST, TiDE, xLSTM)—against bias-corrected output from a five-member CMIP6 ensemble, on 100 meteorological stations across four Western Balkan countries (monthly temperature and precipitation, 1961–2020), using non-parametric significance testing, a rolling-origin backtest (five windows, 2011–2020), and a five-seed robustness check. For temperature, all five deep learning architectures achieved lower MAE than the classical baselines (p < 10−99), though PatchTST’s advantage over climatology was not significant; the best-performing architecture varied across seeds and evaluation windows, so we characterise a leading cluster (N-HiTS, TFT, TiDE, PatchTST) rather than a single winner. The primary temperature advantage was geographically broad-based, while the comparison against the physical-model baseline was robust to the choice of comparator GCM. For precipitation, by contrast, a simple climatological-mean baseline outperformed all five deep learning architectures with no exception across all five rolling-origin windows. The deep learning advantage over classical and physical baselines is thus variable-specific rather than universal. Meteorological station networks, combined with a physically based climate-model comparator, constitute a well-suited evaluation environment for the broader time series forecasting community.

AI doi: 10.3390/ai7090330

Authors: Hongbin Dai Chen Wu Qinqin Zhang Huibin Zeng

The reliable monthly forecasting of fine particulate matter (PM2.5) requires artificial intelligence (AI) models that are accurate, temporally valid, and transparent. We develop a history-only rolling-decomposition framework with lightweight Mamba-inspired selective state-space backbones for 475 city-level administrative units in China. Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and three alternative decomposition strategies use only the PM2.5 data available before each decomposition cutoff, thereby avoiding future-information leakage and yielding inspectable multi-scale predictors. In the primary seven-model benchmark, CEEMDAN-DenseMamba achieved the lowest mean root mean squared error and mean absolute error (6.931 and 4.833 μg m−3, respectively). Equal-optimization reruns, paired moving-block bootstrap intervals, component-count sensitivity, city-wise diagnostics, and a parameter-matched gated recurrent unit baseline were then used to examine performance attribution. Under common optimization, the dense-connection contrasts showed paired error reductions with confidence intervals below zero, whereas the incremental CEEMDAN effect within a fixed backbone was smaller and its paired confidence intervals crossed zero. The recurrent baseline remained competitive. These findings support transparent, temporally valid multi-scale forecasting while limiting inference to future-month prediction for the known cities and the present experimental setting.

AI doi: 10.3390/ai7090328

Authors: Alaa Mohasseb Ronel Beukman Andreas Kanavos

The integration of Artificial Intelligence (AI) in higher education is reshaping institutional decision-making, governance, policy development, and educational innovation. Effective leadership in AI-enabled environments requires more than technical competence, extending to strategic awareness, ethical judgement, and the ability to critically evaluate the broader implications of AI technologies. Drawing on survey and interview data from 52 academic leaders across UK higher education institutions, this study examines current levels of AI literacy and explores how AI capability relates to institutional readiness and leadership practice. The findings reveal variation in participants’ self-reported AI literacy and engagement, identify technical, strategic, ethical, and organisational capability gaps, and highlight structural barriers including limited time, fragmented professional development, and insufficient institutional support for institution-wide AI adoption. In response, the paper presents the AI Literacy Leadership Framework (AILLF), which conceptualises AI literacy as a multidimensional leadership capability comprising technical, strategic, ethical, and applied dimensions that support four interconnected leadership domains: innovation, decision-making, ethical governance, and policy development. Informed by leadership theory, international AI governance frameworks, and the study’s exploratory empirical findings, the AILLF is accompanied by a proposed capability progression model and role-differentiated leadership competency guide to provide implementation guidance for higher education institutions. The study contributes an empirically informed conceptual framework that integrates AI literacy, leadership theory, and AI governance, providing a foundation for future research and institutional approaches to AI leadership within higher education.

AI doi: 10.3390/ai7090327

Authors: Raghu Raman Prema Nedungadi

Claims of AI sovereignty are increasingly invoked but operational control remains uneven. Claims to control are made through national models, sovereign clouds, data localization mandates, and procurement rules; however, whether such claims translate into demonstrable control over how AI systems are run, inspected, and contested remains poorly understood. This paper introduces credible sovereignty, the gap between declared and demonstrable control in deployment, as a conceptual lens for analyzing AI governance to examine how this gap is opened and closed across infrastructure, data, and model supply chains. Using a PRISMA-guided social-science corpus and machine learning-based BERTopic modeling, validated through topic diversity and topic separation diagnostics and triangulated through close reading, the analysis identifies four governance logics through which sovereignty is contested: data infrastructure and legitimacy frameworks; techno-bloc diplomacy and infrastructure politics; European regulatory sovereignty; and community-driven sovereignty in the Global South. Across these logics, sovereignty is enacted less through national capabilities than through proxy mechanisms—certification regimes, procurement clauses, cloud governance, and deployment architectures—each carrying trade-offs between autonomy, dependence, and accountability. Rereading the corpus through an Antecedents–Decisions–Outcomes lens yields a testable research agenda: antecedents that push actors toward sovereignty seeking; design and governance choices that translate ambition into implementation; and outcomes—resilience, inclusion, accountability—against which sovereign AI programs should be assessed. This paper reframes sovereignty as a layered operational capability rather than a discursive claim and links computational synthesis to a normative construct that applies across jurisdictions and scales.

AI doi: 10.3390/ai7090326

Authors: Nguyen Thi Thanh Binh

This review synthesizes theoretical and empirical insights from 1055 peer-reviewed articles on artificial intelligence (AI), corporate governance, and ethics. Situated in the corporate governance and accounting literature, it develops a computational framework to identify thematic patterns and conceptual links among AI, transparency, accounting, governance, and ESG. Using latent Dirichlet allocation, co-occurrence network analysis, sentence-level semantic similarity, and exploratory regression, the study identifies three recurring configurations of conceptual association: (1) Ethics, Governance, and Transparency; (2) Machine Learning, Finance, Blockchain, and Accounting; and (3) Corporate, ESG, and Accounting. The findings indicate that these themes are repeatedly connected within the scholarly literature.

AI doi: 10.3390/ai7080325

Authors: Tomi Božak Shivalika Goyal Marc Langheinrich Martin Gjoreski Gašper Slapničar

Eye tracking is a valuable behavioral signal for human-centered AI, yet the reliability of feature-based machine-learning models for inferring task type and workload across users and tasks remains uncertain, because experimentally defined workload labels may reflect task type and visual structure as much as cognitive demand. The practical problem is that designers of gaze-adaptive systems need to know which inferences are dependable enough to act on, and reported accuracies alone do not answer this, because the choice of prediction target and validation split can determine the result. This study systematically evaluates feature-based machine-learning models with post hoc explainability across three prediction targets: task type, binary load-versus-rest, and three-level workload. Eye-movement features derived from fixations, saccades, pupils, and blinks were extracted from short temporal windows collected from 54 participants performing attention, visual-spatial, and memory tasks under rest, easy, and difficult conditions, and evaluated using leave-one-subject-out (LOSO) and leave-one-group-out (LOGO) validation. Task type was classified most reliably (85.9% LOSO, 83.4% LOGO), binary load-versus-rest showed moderate, validation-sensitive robustness (81.4% LOSO, 63.9% LOGO), and three-level workload classification was substantially more challenging (56.4% LOSO, 44.3% LOGO). SHAP and statistical analyses consistently identified fixation dispersion, pupil-related measures, and subject-normalized features as the strongest contributors across all three targets. These findings show that prediction target definition, validation strategy, and post hoc explainability jointly determine what can be reliably inferred from gaze-based machine-learning models. Eye tracking alone therefore appears promising for task-type recognition and may support coarse engagement-related inference when the deployment task family is represented during model development, whereas task-independent fine-grained workload estimation remains unsupported by the present evidence.

AI doi: 10.3390/ai7080324

Authors: Chenxuan Zhang Peixiao Fan Siqi Bu Yuxin Wen

The transition towards smart megacities has deeply integrated Artificial Intelligence (AI) with power–transport networks. While AI empowers complex operations like multi-network coordinated dispatch and emergency rescue, current algorithm-centric perspectives largely ignore its massive physical energy costs. Accordingly, this Perspective examines the dual role of AI, considering it not only as an intelligent decision-support tool but also as a potential source of additional stress on physical infrastructure. First, through a structured synthesis of the representative literature, we deconstruct the functional dependencies between algorithms and physical infrastructures, identifying how AI reshapes the operational paradigms of power, ground transport, and aerial networks under routine and emergency scenarios. We then introduce the concept of the “Computation–Energy Paradox.” Integrating conceptual analysis with a quantitative case study of a typical community, we illustrate a plausible failure mechanism: during extreme disasters, intensified AI invocation for emergency management generates surging computational loads, which paradoxically exacerbate power shortages and reduce the operating margin of already weakened systems. In addition, we analyze core engineering bottlenecks, including spatiotemporal computation–energy mismatches and physical constraints in extreme edge environments. To address these challenges, we outline a prospective roadmap encompassing lightweight emergency AI and computation–power-coordinated offloading mechanisms. Finally, the sustainable development of such systems suggests a paradigm shift: AI must evolve from a purely virtual algorithm into a physical component of an integrated compute–power–transport system.

AI doi: 10.3390/ai7080323

Authors: Cesar Agustin Corona-Patricio Carolina Reta Jose Antonio Cantoral-Ceballos

Electroencephalography-based visual decoding has important applications in brain–computer interfaces and cognitive neuroscience, yet the relative effectiveness of different feature extraction methods for sustained visual paradigms remains unclear due to the absence of standardized, multi-dataset comparative evaluations. This study systematically compares eight feature extraction methods across three public EEG datasets: MindBigData MNIST, MindBigData MNIST-8B for digit recognition, and MSS for natural image classification. The methods include coherence, Granger causality, directed transfer function, partially directed coherence, transfer entropy, discrete wavelet transform, empirical wavelet transform (EWT), and wavelet scattering transform. Two deep learning architectures, EEGNet and EEG Conformer, were trained using two pre-processing pipelines, with and without artifact removal. EWT achieved the highest classification accuracy, reaching 97.83% for digit-vs-blank and 77.10% for within-session natural image classification. Connectivity-based methods consistently underperformed, with the best connectivity method (coherence) reaching up to 91.67%, suggesting that spectral power information is more discriminative than inter-channel relationships. Cross-subject generalization remained challenging, with best accuracies near 68%. The findings establish wavelet-based adaptive spectral decomposition as a strong baseline for EEG visual decoding and highlight the need for domain adaptation techniques to address cross-subject variability.

AI doi: 10.3390/ai7080322

Authors: Christopher Valdez-Cantú Jose Antonio Cantoral-Ceballos Joanna Alvarado-Uribe

Large Language Models (LLMs) are increasingly integrated into agentic workflows that require extended reasoning, persistent state management, coordinated tool use, and controlled execution. As this operational scope expands, a central question emerges: whether probabilistic generation alone can reliably support coherent behavior across interacting system components. This paper addresses that question through a structural diagnostic review of contemporary agentic systems. Starting from LLM-based tutoring as an analytically demanding entry point and extending toward structurally related agent architectures, the paper draws on a five-phase review of N=145 research records. The analysis is organized through the Agentic Structure Taxonomy (AST), which structures the literature across four dimensions: Cognition, Interaction, Orchestration, and Governance. The review identifies five recurrent empirical problem patterns and uses them as abductive diagnostic cues for formulating seven cross-dimensional transition gaps that capture recurrent discontinuities at the boundaries between reasoning, state, control, and execution. From these gaps, fourteen structural constraints are derived across three control domains: state isolation, control alignment, and execution governance. These constraints are interpreted not as prescriptive design mandates, but as analytically derived conditions associated with reducing error propagation across subsystem transitions. The paper argues that reliability in agentic systems is shaped not only by model performance or prompt design, but also by whether the boundaries linking probabilistic reasoning to persistent state, orchestration, and execution are governed by explicit structural conditions.

AI doi: 10.3390/ai7080321

Authors: Maksim V. Ulizko Aleksandr V. Chernikov Ivan V. Tomilov Natalia F. Gusarova Aleksandra S. Vatian

Normative AI assistants are increasingly used in domains governed by duties, permissions, prohibitions, exceptions, priorities, and institutional policies. Existing retrieval-augmented generation (RAG) and legal AI benchmarks evaluate answer accuracy, retrieval quality, citation grounding, natural-language inference, clause extraction, or general legal reasoning ability. These dimensions are necessary but insufficient when supplied evidence is incomplete, mutually inconsistent, or defeasible. The objective of this study is to introduce ParaTraceBench, a paraconsistent trace-based benchmarking framework for post-retrieval normative reasoning over fixed evidence packages. Each scenario contains a query, evidence fragments, extracted facts, defeasible rules, typed attack edges, priority relations, an expected conclusion status, and a gold diagnostic trace. The formalism uses evidence-grounded arguments, a single edge-based attack representation, explicit attack-licensing rules, acyclic priority bases with a transitive closure, grounded argument labeling, trace-normal-form alignment, and deterministic scoring. The operational NER metric is explicitly interpreted as inconsistency-conditioned unsupported-conclusion avoidance rather than proof of logical non-explosion. We evaluated the framework using 140 scenarios, external validation on 567 anonymized Russian-language cases from Russian Federation and EAEU-related materials, reasoning-oriented baseline adaptations, five-run prompt-fairness and stability controls, and a deterministic component-dependency audit. On the full external set, the trace-based configuration reached 85.7% answer-status accuracy, 85.5% contradiction-localization accuracy, 94.2% operational NER, 84.1% priority-handling accuracy, and 83.7% belief-revision accuracy. These results indicate that contradiction-aware trace evaluation provides diagnostic information beyond final-answer accuracy under the evaluated fixed-evidence conditions, while not establishing causal architectural superiority, logical non-triviality, or end-to-end RAG performance.

AI doi: 10.3390/ai7080320

Authors: Weimao Ke Lixiao Yang Mengyang Xu

Retrieval-Augmented Generation (RAG) has become a practical approach for question answering over external corpora, particularly when answers should be grounded in source documents rather than generated only from model parameters. While recent large language models can process increasingly long contexts, they do not remove the need for selecting, organizing, and auditing evidence, especially when systems rely on smaller local models for privacy, cost, or deployment constraints. In this paper, we frame retrieval granularity as an evidence-design variable for answer grounding in small-model RAG question answering. After a brief exploratory NewsQA phase that motivates the error categories, the main study uses the HotpotQA distractor validation split with 7405 hard multi-hop questions and sentence-level supporting-fact annotations. With Qwen3-8B as the fixed generator, we compare closed-book, fixed-budget whole-context, retrieved-context, gold-document, and gold-supporting-fact conditions while varying retrieval granularity, retriever type, and context budget. Retrieved context substantially outperforms closed-book answering and the 1024-token fixed-budget whole-context condition but remains below gold-document and gold-supporting-fact upper bounds, indicating that retrieval, generation, and evaluation limitations should be analyzed separately. Sentence-level retrieval under-recovers multi-hop evidence, especially for questions with three or more supporting facts, while paragraph-level and moderate token-level chunks recover substantially more complete evidence. In the full condition matrix, hybrid retrieval with 256-token chunks and no overlap achieves an F1 of 0.6816 with a supporting-fact recall of 0.9609, compared with an F1 of 0.6166 and supporting-fact recall of 0.7801 for BM25 sentence retrieval. Additional ablations show that fixed-budget whole-context performance is strongly affected by truncation, that overlap has little practical effect under the tested 1024-token budget, and that a stronger BGE dense retriever improves the best retrieved-context F1 to 0.7027. These results align with a diagnostic perspective on chunking: using evidence at a task-appropriate level of granularity can improve grounding, auditability, and answer quality, but the observed patterns should be interpreted within the HotpotQA distractor setting, fixed generator, and tested context budgets.

AI doi: 10.3390/ai7080319

Authors: Neethika Alluri Pardha Saradhi Varma Gottumukkala Hemalatha Indukuri

Large language models (LLMs) now generate fluent natural language and source code, creating challenges for authorship attribution, academic integrity, and software supply-chain security. Most existing detectors for AI-generated content are evaluated separately on natural language or source code, often under matched train–test conditions that can overestimate real-world reliability. We present a paired-prompt benchmark for human-versus-machine detection across English text, Python code, and mixed text–code documents. The benchmark includes 22,141 instances from HC3, CodeSearchNet, MBPP, and HumanEval across training, validation, and test partitions, plus Mix-Eval, a mixed-content set of 997 Jupyter-notebook-style samples. We evaluate RoBERTa-large for text, GraphCodeBERT and CodeBERT-base for code, a unified RoBERTa-base detector trained on both modalities, and zero-shot baselines. Fine-tuned detectors achieve near-perfect in-distribution performance, with AUROC 1.0000±0.0000 and accuracy above 99.5%. Across five instruction-tuned generator families of varying size (3.8B–7B) and architecture, with the human and problem distributions held fixed, cross-generator transfer causes negligible degradation (AUROC spread 0.0002; drops of at most 0.0003). In contrast, domain shift is the main failure mode: on MBPP+HumanEval, GraphCodeBERT drops to 0.85±0.02 AUROC and CodeBERT-base to 0.67±0.02. On Mix-Eval, the unified detector outperforms a routed text–code pipeline by 21 AUROC points (0.96 vs. 0.75), largely because of router failures on mixed inputs. Training-time augmentation improves low-false-positive performance, while legacy supervised detectors show systematic class inversion on modern LLM outputs. These results show that reliable deployment requires cross-domain evaluation, mixed-content testing, and calibration beyond in-distribution accuracy.

AI doi: 10.3390/ai7080318

Authors: Samir Elouaham Fatima Ezzahra Bouaaza Ilyas Ait Ichou Boujemaa Nassiri

Diabetic retinopathy (DR) is a common complication of diabetes and one of the leading causes of preventable vision loss worldwide. Because the manual grading of color fundus images is slow and depends on the availability of trained specialists, automated screening tools are needed. This study proposes a lightweight channel-wise refinement strategy for automatic five-class DR grading, built on a ResNet50 backbone. Two custom blocks are evaluated: CBR, which applies a 3 × 3 convolution, batch normalization, and a ReLU activation to make the channel representation more compact, and CBS, which applies a 3 × 3 convolution, batch normalization, and a SiLU activation to reinforce local spatial features. On the Diabetic Retinopathy Balanced dataset, the baseline ResNet50 reached an accuracy of 90.77%, a precision of 90.60%, a recall of 90.79%, and an F1-score of 90.64%. In the ablation study, the best configuration was ResNet50 + CBR, with an accuracy of 91.85%, a precision of 91.75%, a recall of 91.88%, and an F1-score of 91.76%. The full CBR-CBS Hybrid ResNet50 was close behind, with an accuracy of 91.81% and an F1-score of 91.71%. The CBR block accounts for most of this improvement, which suggests that channel-wise refinement helps the model separate subtle lesion patterns. These results establish lightweight channel-wise refinement (CBR) as an effective, compact, and interpretable enhancement of ResNet50 for automated five-class DR grading, delivering a consistent multi-metric gain over the baseline and accuracy competitive with the literature, which makes it a promising solution for large-scale screening.

AI doi: 10.3390/ai7080317

Authors: Alexandru Dima Maria-Elena Mihăilescu Darius Mihai Mihai Carabaș Mihai Dascalu

IT help desks at large organizations face a high volume of recurrent, well-documented user requests that nevertheless require human-written replies, creating a persistent staff workload that is repetitive in content but non-trivial in tone and procedural correctness. We present FroLineR, short for Front-Line Response, a system that drafts the initial staff reply to such tickets in the login and account-activation category and integrates into a human-in-the-loop ticketing workflow on a Romanian-language ticketing platform. The generator is an unmodified instruct model augmented with retrieval from a small set of hand-curated guide documents, using a Romanian system prompt refined over several rounds of staff review. To evaluate and refine the prompt without manual labeling, we cluster the first user message of every historical thread with both BERTopic and Semantic Signal Separation (S3), score configurations along coherence and lexical-diversity axes, and extract a 200-message evaluation set from the winning model. Prompt convergence was certified by several rounds of manual review by support staff. The production system is quantized to Q4_K_M GGUF, served through llama-cpp-python behind a small Flask API, and deployed with GPU offloading on the target server, reducing end-to-end per-answer latency from approximately 830 s on the server’s CPU to roughly 61 s once layers are offloaded to the GPU, with no observable degradation in answer quality.

AI doi: 10.3390/ai7080316

Authors: Yongqiang Gong Ruiqi Ma Xicheng Wang Ruixi Li Han Dong Yijin Liu Xi Peng Quanle Guo Yin Liu

Background: Large language models (LLMs) show considerable promise for medical question answering and reasoning. Their use in medical microbiology, however, remains constrained by limited domain-specific knowledge and the risk of hallucinated outputs. Objective: To develop and evaluate Smart Medical Microbiology (SmartMM), a specialized LLM for accurate, reliable, and context-aware responses in medical microbiology. Methods: SmartMM integrates domain-adaptive continual pretraining, supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), knowledge distillation, and retrieval-augmented generation (RAG). We constructed a high-quality microbiology corpus from textbooks, clinical guidelines, the scientific literature, case reports, and other authoritative sources. Model performance was assessed using objective examinations, subjective generation tasks, expert review, and real-world user preference evaluation. Results: SmartMM achieved accuracies of 0.897 and 0.563 on true-or-false and fill-in-the-blank questions, respectively. In subjective generation tasks, it obtained the highest ROUGE-L score (0.265) and BERTScore F1 score (0.771) among all compared models. Expert assessment showed excellent inter-rater reliability, with all ICC(C,3) values exceeding 0.970. In a user evaluation involving 20 participants and 100 real-world questions, SmartMM received the largest number of first-place rankings (33), placing it among the top-performing systems overall. Conclusions: SmartMM showed strong domain adaptability in medical microbiology knowledge organization, semantic generation, and retrieval-augmented reasoning. These findings support its potential use in educational support, infectious disease knowledge assistance, and retrieval-enhanced medical question answering.

AI doi: 10.3390/ai7080315

Authors: Vijayalakshmi Sankaran Paramasivam Alagumariappan Sumendra Yogarayan Thayananth Caran Varshana Balaguru Ramana

Globally, respiratory diseases such as asthma, chronic obstructive pulmonary disease (COPD) and pneumonia affect populations significantly, requiring early and accurate diagnosis for effective clinical management. Manual auscultation and expert interpretation are the common shortcomings in conventional diagnostic approaches, as they lead to time-consuming and inconsistent analysis. To address these limitations, an artificial intelligence-driven framework for respiratory disease classification using multi-spectral feature extraction and deep learning architectures is proposed to classify four different respiratory conditions: Asthma, COPD, Pneumonia and Healthy. The dataset is collected from Kaggle’s respiratory sound database and the COUGHVID V3 database, which together contain 322 Asthma signals, 746 COPD signals, 323 Pneumonia signals and 174 Healthy signals. Subsequently, the features are extracted using four different feature extraction techniques—Constant Q Transform (CQT), a Gammatone spectrogram, Mel-Frequency Cepstral Coefficients (MFCC) and Perceptual Linear Prediction (PLP)—and these extracted spectral representations are provided as inputs to various deep learning models such as a Deep Convolutional Neural Network (Deep CNN), a Temporal Attention Network (TAN) and an Autoencoder for automated feature learning and disease classification. The proposed framework is evaluated using several performance metrics, and the experimental results clearly indicate that the performance of the proposed classification framework strongly depends on the selection of spectral feature extraction techniques and deep learning models. Among all the evaluated combinations, it is evident that the Autoencoder model integrated with CQT features exhibited the best classification performance, with an accuracy of 98.72%, precision of 98.74%, recall of 98.72%, Matthews correlation coefficient (MCC) of 98.11%, Cohen’s kappa value of 98.10% and the least log loss of 0.025. The proposed artificial intelligence (AI)-enabled respiratory disease classification framework has demonstrated the ability to produce a reliable computer-aided diagnostic system which is suitable for smart healthcare applications and automated pulmonary disease screening.

AI doi: 10.3390/ai7080314

Authors: Sheikh Mohammed Shariful Islam Alan W. Gemmill Yafit Hirshler Michaela Pascoe Jeannette Milgrom

Perinatal mental health remains a critical global challenge, with maternal mortality, preterm birth, and persistent disparities in care contributing to adverse outcomes for mothers. In addition, mental health difficulties in the perinatal period are associated with poorer developmental outcomes for young children and impose an economic burden on societies. Addressing these issues requires innovative approaches that can complement traditional clinical practices. Artificial intelligence (AI) has emerged as a powerful tool with the potential to transform perinatal care by enabling early risk prediction, personalised interventions, and scalable support systems. However, there are no existing reviews on use of AI across different stages of perinatal mental health. We conclude with a call to action for clinicians, researchers, policymakers, and technology developers to collaborate on a consensus framework that ensures ethical, safe, and equitable integration of AI into perinatal care.

AI doi: 10.3390/ai7080313

Authors: Zhifan Jin Lin Zhou He Wang Sijia Chen Liman Liu Wenbing Tao

Crowd counting estimates the number and spatial distribution of people in images and videos, supporting smart city management and public safety. Existing methods often rely on intra-level feature refinement and simple cross-scale fusion, such as concatenation or addition, which limits interaction between fine-grained spatial details and high-level semantic representations. In addition, the limited receptive field of convolutional networks restricts global context modeling in scenes with heavy occlusion and extreme scale variation. To address these challenges, we propose a Hierarchical Context-Aware Multi-Scale Attention Network (HCMA). Its bidirectional cross-level interaction is realized through two complementary top-down decoding streams, where an attention-gating stream provides spatial guidance for the counting-oriented representations carried by a density-feature stream. HCMA includes three modules: the Selective Context-Aware Attention Module (SCAM), which performs context-dependent multi-scale filtering; Dynamic Positional Pooling (DPP), which introduces an image-level mean token and stochastic global-relation aggregation; and the Multi-Scale Enhancement Attention Module (MSEA), which refines high-level semantic features under scale variation. Experiments on ShanghaiTech, UCF-QNRF, and NWPU-Crowd show competitive counting accuracy across scenes with different density ranges, scale variation, and occlusion. In particular, HCMA achieves an MAE of 73.2 on NWPU-Crowd, 17.2% lower than that of DM-Count.

AI doi: 10.3390/ai7080312

Authors: Rashed Bahlool Nabil Hewahi

As regulatory requirements increasingly shape automated lending decisions, fairness remains a critical challenge in high-stakes domains, particularly credit scoring. Although artificial intelligence models can achieve strong predictive performance, they may also reproduce biased outcomes that reduce financial inclusion or transfer harm to overlooked protected groups. Existing fairness interventions commonly operate at a single stage of the decision-making pipeline, despite bias often propagating across representational and decision layers. This study proposes iCert-Fair, a two-layer framework for technical fairness assessment and harm recovery in credit scoring. The first layer adopts a fairness-through-explainability paradigm, using SHAP-based explanations to identify direct and proxy dependence on protected attributes and guide structural dataset repair, while the second layer applies targeted threshold-policy adjustments to recover residual harm while preserving decision utility. Experiments on the German and Taiwanese credit datasets show that fairness gains are model- and dataset-specific and may be collective, concentrated, transferred, or recovered unevenly across protected attributes. The direct comparison with representative pre-processing, in-processing, and post-processing methods revealed that baseline methods targeting one protected attribute at a time frequently transferred residual harm to other monitored attributes. In contrast, the fairness-focused recommendations generated by iCert-Fair achieved larger collective fairness improvements across all considered protected attributes while avoiding residual harm. These gains were obtained while preserving predictive utility on the German dataset and with utility degradation remaining below 5% across the evaluated performance metrics on the Taiwanese dataset, alongside consistently lower false-negative risk. The empirical findings support the use of complementary structural and policy-level interventions and demonstrate the importance of jointly evaluating aggregate disparity, worst-case attribute-level harm, cross-attribute transfer, and predictive utility.

AI doi: 10.3390/ai7080311

Authors: Boris Galitsky

Large language models can generate fluent chain-of-thought (CoT) reasoning that appears coherent while exhibiting systematic distortions in evidence weighting and hypothesis comparison. This paper studies hallucinated CoT as a discourse-structural phenomenon, not only a factual one. We introduce a diagnostic reasoning benchmark with paired grounded and hallucinated explanations, where traces differ in how they organize evidence, alternatives, and defeaters. We extract discourse tree features that summarize evidence allocation, contrast preservation, commitment timing, and evidence integration, and combine them with the Joint Knowledge–Reasoning Hallucination Measure (JKRHM). Experiments on the synthetic diagnostic dataset and preliminary external validation on HaluBench suggest that discourse structure provides an interpretable signal for detecting reasoning hallucinations and complements existing factuality and uncertainty-based hallucination detectors. Because the HaluBench reasoning rationales are generated as an intermediate representation, these results should not be interpreted as definitive external proof of generalization. The results support a cautious conclusion: discourse analysis does not replace factual verification, but it helps expose reasoning paths that are structurally unsupported even when they are fluent and persuasive.

AI doi: 10.3390/ai7080310

Authors: Yepeng Shi Shengliang Fang Shunhu Hou Yuhai Li You Fu Qichen Wang

Radio Environment Map (REM) construction is frequently constrained by sparse and unevenly distributed spectrum measurements. While existing completion methods primarily target Power Spectral Density (PSD) data under random missing patterns, the reconstruction of Reference Signal Received Power (RSRP) maps under structured data loss remains underexplored. This study addresses this gap by proposing a fully convolutional densely connected autoencoder(AE) for RSRP map completion. The encoder stacks dense blocks and transition layers, a bottleneck preserves the latent representation, and the decoder restores spatial resolution through transposed convolution. Both global and local skip connections are incorporated to fuse large-scale structure with fine-grained details. A composite loss function supervises observed and missing regions separately, which preserves the fidelity of known measurements while improving inference over unobserved grid points. Experiments on the public DeepREM dataset under random, spatial, and strip-wise missing patterns show that the method achieves the best or comparable completion accuracy in most tested settings, with the most pronounced performance gains over mainstream baselines under the challenging spatial block-missing case.

AI doi: 10.3390/ai7080309

Authors: Gali Chelouche-Dwek Peter Fonagy

Background: Teachers routinely encounter children whose behaviour reflects emotional distress and dysregulation, yet they have limited opportunities to practise the relational skills required to respond effectively. These challenges are particularly pronounced in Alternative Provision (AP), which serves children who frequently present with histories of trauma, neurodevelopmental differences, and complex emotional and behavioural needs. Mentalization, the capacity to understand behaviour in terms of underlying mental states, is central to effective relational practice in such contexts. Conversational Artificial Intelligence (AI) may offer a scalable means of supporting this form of skills development, but its feasibility as a teacher-training modality remains largely unexplored. Methods: This mixed-methods proof-of-concept feasibility study evaluated MentaClassAI, a novel AI-based training tool in which educators engaged in simulated voice conversations with AI child characters portraying classroom dysregulation and subsequently received individualised, mentalization-informed feedback. Eleven staff members from a single AP school (four teachers and seven teaching assistants) completed a single training session and were allocated to either a psychoeducation video condition (n = 6) or a no-video condition (n = 5). The video condition received a brief introduction to mentalization and epistemic trust prior to engaging with the simulation. Pre- and post-engagement measures included the Reflective Functioning Questionnaire (RFQ-8) and a Teacher Self-Efficacy Scale. Post-engagement measures included an 18-item acceptability questionnaire, a Technology Acceptance Model scale, and open-ended questions analysed using thematic analysis. Results: Acceptability was high, with 84.8% of questionnaire responses falling within the positive range (overall M = 5.60/7). Feedback accuracy (M = 6.55) and clarity (M = 6.36) received the highest ratings. Participants reported higher teacher self-efficacy after the session than before (d = 1.20, p = 0.003), with 10 of 11 participants demonstrating improvement. Self-reported hypomentalizing was lower after the session (d = −0.86, p = 0.017). Between-condition differences (video versus no-video) were not statistically significant. The video condition scored numerically higher on the directional indicators. Qualitative analysis identified five themes: the value of consequence-free rehearsal; the specificity and usefulness of feedback; appreciation of the focus on the child’s emotional experience; limitations in the ecological diversity of AI child characters; and a desire for more naturalistic interaction. Conclusions: These findings provide preliminary support for the feasibility and acceptability of AI-based mentalization practice for AP staff. The principal value of the tool appears to lie not only in the simulation itself but in the quality of the reflective feedback generated. Although based on a small sample, the observed pre–post changes provide an encouraging signal that may justify a controlled trial. The contribution of pre-session psychoeducation to training outcomes remains an important question for future research.

AI doi: 10.3390/ai7080308

Authors: Laura Acosta García Julen Cestero Portu Ander García Gangoiti Marco Quartulli

Warehouse operations involve dynamic item flows, fluctuating demand, and heterogeneous layouts, making adaptive decision-making essential for efficient storage and order fulfillment. In this context, reinforcement learning (RL) provides a promising approach for learning adaptive warehouse control policies under stochastic environments. However, evaluating RL-based solutions in real warehouse settings is often costly and time-consuming, motivating the need for realistic and reproducible simulation environments. In this paper, we introduce a configurable warehouse simulation environment modeling stochastic item arrivals, order generation, and internal logistics operations across diverse layouts and workload conditions. Based on this environment, we construct a reproducible experimental testbed composed of multiple scenarios ranging from low-load to highly congested settings. The testbed is publicly released to support reproducible research and comparative evaluation within the research community. We formulate the warehouse management problem as a Markov decision process (MDP) and apply a Maskable Proximal Policy Optimization (Maskable PPO) agent to learn adaptive control policies. The RL-based approach is evaluated across the defined scenarios and compared against heuristic baseline strategies. Experimental results show that the proposed solution achieves performance comparable to a strong greedy first-in, first-out (FIFO) heuristic while improving order fulfillment by up to 13.5 percentage points under challenging workload conditions. These results demonstrate the ability of RL to learn robust warehouse control policies that adaptively optimize performance and maintain operational stability across a wide spectrum of distinct scenarios.

AI doi: 10.3390/ai7080307

Authors: Oleksandra Topal Inna Novalija Joao Pita Costa Dumitru Roman

This study deploys a scalable machine learning pipeline: combining a transformer-based classifier applied to 2.01 million English-language AI-related news headlines (July 2022–July 2024) with large-language-model and human-annotator validation (three annotators, Fleiss’ κ=0.80) on stratified subsamples, to extract six interpretable, bias-linked discourse indicators computed at the AI-domain level: evaluative orientation (valence), loss salience, narrative drift, exposure-adjusted sentiment, cross-source divergence, and novelty-phase framing. Each operationalizes an established cognitive-psychology construct as a computable property of the information environment associated with biased risk–benefit reasoning. Results show systematic variation across domains: technical and methodological areas such as deep learning and natural language processing exhibit gain-salient framing, while safety-critical topics such as deepfakes (loss-to-gain headline ratio = 3.17) and facial recognition show strongly loss-salient profiles. Cross-model validation using an LLM on a stratified sample of 1000 headlines confirms that domain-level indicator rankings are robust to classifier choice (Spearman ρ=0.83; p<0.001), establishing the rank stability of pipeline outputs independently of the specific classification architecture. As a contextual application, domain-level profiles are mapped to European Union AI governance instruments, documenting parallels between discourse patterns and regulatory risk tiers. The framework provides a scalable, reproducible methodology for monitoring evaluative conditions in technology news across domains, sources, and time.

AI doi: 10.3390/ai7080306

Authors: Genaro Herrera Cano Michal Dyrda Youssef Beshay David Baltrusaitis Mitch Paro Rafael Olivieri-Ortiz Grigoriy Androsov Antonio Medina Luna Michael Baldwin

Traumatic cervical spine fractures (CSFxs) require timely diagnosis due to associated morbidity. Artificial intelligence (AI)-based decision support systems have been proposed to improve imaging workflow efficiency. However, their performance in non-trauma settings remains unclear. This study evaluated the diagnostic performance of the AIDOC decision support system (DSS) for detecting CSFxs in a non-trauma academic community hospital using a retrospective analysis of 1812 cervical spine CT scans, with radiologist interpretation as the reference standard. Sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) were calculated for fracture detection and heatmap-based localization. The AI system demonstrated a sensitivity of 72.2% and specificity of 98.1%, with an accuracy of 97.9%. In the context of low fracture prevalence (0.99%), PPV was low (27.7%), while NPV was high (99.7%). Heatmap-based localization showed reduced sensitivity (43.8%) despite high specificity (97.5%). These findings demonstrate high specificity and NPV, with lower sensitivity for localization and low PPV in a low-prevalence setting. Prospective multi-institutional studies are required to further validate these diagnostic performance metrics and assess generalizability across diverse clinical settings and imaging protocols.

AI doi: 10.3390/ai7080305

Authors: Madhini Balasundaram Supraja Perumal

Waste management on university campuses is complicated by the constant change in packaging types, which existing waste-sorting systems cannot recognise unless they are retrained. Open-vocabulary object detectors can identify objects from text descriptions instead of a fixed list of categories, offering a possible solution to this problem. However, it is not known how well this ability survives when such a detector is fine-tuned and deployed on low-power hardware. This paper presents EcoSortBin, a waste-sorting system built on the YOLOE-26 detector and deployed on a Raspberry Pi 4. YOLOE-26 was first fine-tuned on a 1330-image campus waste dataset covering seven classes, with masks generated using the Segment Anything Model, establishing a baseline called WasteYOLOE26-S with 74% top-1 accuracy on known classes; however, this fine-tuning reduces the model’s ability to recognise the same seven classes when they appear in a different dataset or setting. RLPA (RepRTA-Compatible LoRA Prompt Adapters) addresses this by adapting only the text-embedding component of the model using a small set of additional parameters (16,384 parameters, rank 16), leaving the rest of the network unchanged; this restores cross-domain generalisation but reduces top-1 accuracy on known classes to only 15%, which is too low for practical use. To recover this accuracy without losing cross-domain generalisation, frozen-backbone neck fine-tuning (NeckFT) was added, which fine-tunes the feature-combining layers of the network while keeping the main backbone frozen, preserving its pretrained visual–text alignment. Combining RLPA with NeckFT achieved the best balance of the three approaches, with 73.5% top-1 accuracy and a Cross-Domain Generalisation Ratio (CDGR) of 0.2435. To test whether this ability extends to genuinely new categories, the model was further tested on 28 novel categories not seen during training, totalling 840 images. RLPA + NeckFT showed consistent zero-shot generalisation to novel objects with container-like shapes, such as bottles and jars. After quantisation for edge deployment, the model kept its full accuracy ranking and produced a compact 41.8 MB file suitable for the Raspberry Pi 4. These results show that RLPA + NeckFT gives a practical balance of accuracy and generalisation for campus waste detection on low-power hardware.

AI doi: 10.3390/ai7080304

Authors: Xinran Zhang Yang Sheng Sijie Shen Dongjie Fan Lizhuang Liu

Learning from high-dimensional, low-sample-size (HDLSS) data remains a persistent challenge in machine learning, as models must infer reliable patterns from limited observations while handling an excessive number of variables—a scenario particularly prevalent in biomedical applications. Such data structures render predictive modeling highly vulnerable to erratic optimization and overfitting. To address this challenge, we propose the Global Interaction Network with Dynamic Generalization Control (GINet-DGC), an artificial intelligence (AI) framework that integrates feature-wise structural priors with dynamic generalization monitoring. Rather than directly learning an unconstrained first-layer weight matrix, GINet-DGC generates task-specific weights from multi-view feature descriptors, encompassing latent semantic, global distributional, local topological, and hierarchical representations. This structure-constrained weight generation strategy effectively narrows the feature-interaction search space and acts as an inductive regularizer against noise and redundant molecular features. Furthermore, we introduce an Overfitting-aware Index (OFI) to monitor the training trajectory and effectively identify the generalization saturation point for adaptive termination. Empirical evaluations on eight public real-world biomedical HDLSS gene-expression datasets, using a repeated stratified 5 × 5 cross-validation protocol, demonstrate that GINet-DGC achieves competitive and stable performance against 17 baselines. These findings support the effectiveness of the proposed framework within the evaluated public biomedical HDLSS benchmark setting.

AI doi: 10.3390/ai7080303

Authors: Camilo Aurelio Velandia Nelson Iván Bedoya Andrés Chiappe David Muñoz-Ballier

This study identifies and characterises the Socratic mediation patterns enacted by MIA, an AI-based conversational agent used in distance higher education. A deductive content analysis was conducted on 737 conversations using six categories: exploration of prior knowledge, contextual adjustment, linkage to experiences, autonomy-oriented prompts, dialogic progression, and verification prompts. The categorical framework achieved full expert content-validity agreement (S-CVI/Ave = 1.00). Contextual Adjustment (77%), Verification Prompts (76%), and Autonomy-Oriented Prompts (74%) were the most frequently observed categories. Sixty of the 64 theoretically possible category combinations occurred in the corpus, and Linkage to Experiences appeared more frequently in personal conversations (33.7%) than in academic conversations (18.3%). The distribution of categories also varied according to conversation length, with longer exchanges containing a broader range of coded dialogic moves. These findings describe the conversational repertoire through which MIA operationalised features associated with Socratic mediation. Because the study did not include independent measures of student satisfaction, learning, engagement, or self-regulation, the results should not be interpreted as evidence of educational effectiveness or causal effects. The study contributes an operational framework for analysing Socratic features in AI–student interactions and identifies directions for outcome-based research.

AI doi: 10.3390/ai7080302

Authors: Xinru Guo Linze Bai

Accurate mapping of paddy rice is essential for agricultural monitoring, yield estimation, and food security assessment. However, optical imagery is often affected by clouds and spectral confusion, while SAR imagery suffers from speckle noise and weak spatial detail representation. Simple optical and SAR feature concatenation is therefore insufficient for complex agricultural landscapes. To address these limitations, this study proposes ACBE-CroFuseNet, an optical and SAR cross-fusion semantic segmentation network for paddy rice extraction using Sentinel-1 SAR and Sentinel-2 optical imagery in Yancheng, Jiangsu Province. ACBE-CroFuseNet introduces two task-oriented designs for paddy rice mapping. First, an attention cross-fusion module is developed to adaptively model modality contributions and spatial responses between optical spectral–textural features and SAR scattering–structural features. Second, a boundary enhancement module with boundary supervision is introduced to strengthen the delineation of fragmented paddy fields and field edges. Multimodal feature aggregation and multi-scale deep supervision are further used to improve feature utilization and segmentation stability. Compared with UNet++, Swin-Unet, CroFuseNet, and CMFFNet under five-fold cross-validation, ACBE-CroFuseNet achieves the best overall performance. The extracted paddy rice area in Yancheng in 2025 demonstrates the applicability of the proposed method for large-scale crop mapping.

AI doi: 10.3390/ai7080301

Authors: Sofia I. Hernandez Torres Jennifer Achay Scotty Bolleter James A. Bynum Eric J. Snider

Junctional hemorrhage has a high fatality rate due to how difficult it is to control rapid bleeding from major vessels. The available methods to stop junctional blood loss are prone to placement errors as well as failure during transport and during prolonged field care. On the battlefield, medical imaging with a portable ultrasound can be leveraged for visualization of the underlying tissue and application of compression at the anatomical junction to effectively stop blood flow. In this work, we developed AI models for anatomical landmark tracking using a perfused human cadaver model. These AI models were paired with an end-user clinical application to guide proper placement and compression, improving junctional hemorrhage control on the future battlefield. The trained U-Net semantic segmentation model demonstrated strong performance across predictions for both validation and hold-out, blind subjects. Overall pixel accuracy across the dataset was 98.9% for training and 98.6% for blind subjects. The artery and vein predictions achieved the highest class-specific training intersection-over-union scores, both at 0.73. This segmentation model trained to interpret human tissue provides evidence that ultrasound visualization can help guide compression at anatomical junctions. Future work will focus on improving blind performance for implementation of this AI model into closed-loop control of hardware prototypes, delivering real-time predictions and control.

AI doi: 10.3390/ai7080300

Authors: Mehdi Chrifi Alaoui Nour-Eddine Joudar Mohamed Ettaouil

Automatic detection of stress in social media text holds promise for supporting digital mental health, but most existing Transformer-based approaches are opaque and computationally demanding. This work presents DR-Transformer, a Dual-Regularized Transformer that combines two complementary mechanisms: (i) a group sparsity penalty (L2,1/L2 elastic net) applied to the query and key projection matrices of every attention head, which encourages whole-row sparsity, producing more concentrated and inspectable attention patterns; (ii) a supervised contrastive loss on the [CLS] projection, which organizes the latent space according to the stress label. The architecture is intentionally lightweight (six layers, eight heads, 256-dim embeddings; ∼9.5 M parameters) and runs entirely on consumer-grade hardware (NVIDIA GTX 1660, 6 GB). Experiments on the publicly available Dreaddit dataset (binary stress classification, 2838 train/715 test segments) compare DR-Transformer against Logistic Regression, BiLSTM, a Standard Transformer of identical architecture, and MentalBERT. Across five seeded runs, DR-Transformer (Full) reaches F1=0.876 (bootstrap 95% CI 0.852–0.898), outperforming the Standard Transformer (F1=0.842; McNemar p<0.001 with Bonferroni correction) and performing comparably to the much larger MentalBERT (F1=0.879; p=0.421). Sparse regularization increases the fraction of near-zero attention weights (below 0.01) from 0.215 to 0.682, while the supervised contrastive loss improves the silhouette score of [CLS] embeddings from 0.312 to 0.483. Dual regularization thus combines accuracy, efficiency, and structurally induced attention concentration in a single model which can be trained without specialized infrastructure. We use the term “interpretable” throughout in this restricted, structural sense—to refer to concentrated and inspectable attention—rather than in the sense of established causal or mechanistic faithfulness; this is only partially and indirectly supported by our token deletion analysis.

AI doi: 10.3390/ai7080299

Authors: Yuxin Wen Peixiao Fan Zhiyu Mao Fang Chi Chenxuan Zhang Yuhong Lu

Modern power grids require safer and more reliable field operations, yet conventional robots often face limitations in unstructured environments because of rigid pre-programming and weak perception–action coupling. This review examines Embodied Intelligence (EI) as an emerging direction for enhancing power-system field operations. We first evaluate the environmental adaptability of morphological carriers, including quadrupeds, humanoids, and unmanned aerial vehicles, and then define the perception–cognition–execution closed-loop architecture used in this review. Three application domains are then examined. Intelligent inspection focuses on active perception and potential open-vocabulary object detection. Live-line maintenance emphasizes Sim-to-Real methods and shared autonomy, while disaster-response applications involve heterogeneous air–ground robotic coordination. The review also discusses the potential for EI to reduce human exposure to hazardous tasks and influence labor structures, while a regional text-based proxy illustrates differences in policy attention to digital infrastructure. Finally, we analyze major constraints, including hardware endurance under extreme climates, edge-computing latency, foundation-model uncertainty and hallucination, cybersecurity, and safety certification. Overall, EI should not be interpreted as a mature replacement for current utility practice; it is a developing technological direction whose safe deployment will require field validation, standardized evaluation, cybersecurity assurance, and continued human supervisory authority.

AI doi: 10.3390/ai7080298

Authors: Tomáš Valenta Ondřej Rozinek Josef Horálek

The shift from passive predictive models to autonomous agents capable of tool use and multi-step planning moves the AI safety landscape from prediction error to control failure: small misjudgements become irreversible actions, and risks compound across long horizons and populations of interacting systems. We present a structured review and taxonomy of open scientific problems in agentic AI safety, mapped explicitly onto the EU AI Act and the NIST AI Risk Management Framework. The corpus follows a PRISMA-ScR scoping review, assembled through anchor-based citation chaining and curated reading lists across arXiv, the major machine-learning conferences, and selected security and fairness venues, with a primary March 2026 search cut-off (extended to May 2026 during revision for a small number of high-relevance governance and agentic-safety sources), explicit eligibility criteria, and an analytical distinction between open scientific problems and deployment risks. The taxonomy identifies eight problem families spanning reinforcement-learning policies and language-model planners: goal specification, inner alignment, safe learning and robustness, scalable oversight, interpretability, tool-use security, multi-agent safety, and evaluation and assurance. Mapping these onto the two frameworks shows close alignment for some families and notable absences for others, with multi-agent safety surfacing as a regulatory gap. We add a per-family research roadmap with concrete milestones and a practitioner-facing deployment-posture triage, arguing that progress on inner alignment, interpretability for deceptive-alignment detection, and multi-agent safety would most directly reduce compliance uncertainty.

AI doi: 10.3390/ai7080297

Authors: Jalal Mehdiyev Ramiz Aliguliyev

Detectors built on transformer language models report near-perfect accuracy on standard fake news benchmarks, which suggests the task is almost solved. We argue that much of this accuracy reflects a confound between writing register and veracity: in common benchmarks, the real class is human-written while the fake class is machine-generated or machine-rewritten, so a detector can separate the classes by recognizing AI writing style rather than by judging truth. To evaluate this, we designed a two-regime evaluation. Phase 1 is the laundering regime, comparing untouched human-real articles against laundered fake articles. Phase 2 is register-controlled: the real class is passed through the same cross-LLM laundering chains as the fake class, so both classes are read in one machine register, and the register cue is no longer available to the detector. We train seven detectors on three datasets and evaluate each frozen detector under both regimes. Under Phase 1, detectors appear robust; under Phase 2, detection on WELFake collapses from about 99% to about 62% AUROC and the largest models approach chance. The effect is benchmark dependent, large on WELFake, mild on IFND and near zero on GossipCop, and it is confirmed by bootstrap testing with false discovery rate control. We recommend register-controlled evaluation as standard reporting practice.

AI doi: 10.3390/ai7080296

Authors: Daniele Giansanti Andrea Lastrucci

Background: Artificial intelligence-driven digital therapeutics (AI-DTx) are rapidly emerging as a transformative paradigm in healthcare, integrating machine learning, deep learning, and generative AI into digital interventions across diverse clinical domains. Despite rapid growth, the evidence landscape remains fragmented, with heterogeneous methodologies, diverse application contexts, and limited cross-domain synthesis. Aim: This narrative view aims to provide an evidence-informed narrative synthesis of the available secondary literature on AI-driven digital therapeutics, primarily focusing on systematic reviews and meta-analyses, to identify emerging patterns, cross-cutting trends, and future directions across clinical and technological domains. Methods: A narrative synthesis of secondary evidence was conducted, focusing on 23 systematic reviews, meta-analyses, and relevant review articles addressing AI-driven digital therapeutics. The identified literature was analyzed to explore recurring themes across clinical domains, technological approaches, and implementation challenges. Findings were further contextualized through selected recent randomized controlled trials and translational studies to provide insights into emerging clinical applications and real-world perspectives. Results: Across the available literature, AI-driven digital therapeutics demonstrate a broad and rapidly evolving expansion across mental health, chronic disease management, rehabilitation, and behavioral health. The field is characterized by a progressive shift from static, rule-based interventions toward more adaptive systems supported by machine learning, deep learning, and generative AI. A key emerging theme is the role of AI as an enabling layer for personalization, adaptation, and dynamic intervention delivery rather than as a standalone therapeutic modality. Mental health represents the most extensively studied domain, particularly through conversational agents and cognitive behavioral therapy-informed interventions, while other clinical areas are progressively expanding their translational potential. Persistent challenges include methodological heterogeneity, limited long-term validation, and incomplete integration into routine clinical workflows. Discussion: The current evidence suggests a transition toward hybrid human–AI models of care, in which digital systems may support and augment clinical practice through adaptive and data-driven approaches. However, the field remains characterized by fragmented evidence, evolving evaluation approaches, and challenges related to standardization, validation, and real-world implementation. Conclusions: AI-driven digital therapeutics are evolving toward increasingly adaptive and clinically oriented healthcare solutions. Future progress will depend on improving methodological consistency, strengthening long-term evaluation, and supporting responsible integration into clinical pathways to ensure safe, scalable, and meaningful impact.

AI doi: 10.3390/ai7080295

Authors: Christian Pascal Silva Bouix Vinícius Villa e Vila Marcos Roberto Benso Sergio Nascimento Duarte Carlos Roberto Padovani Roseli Aparecida Francelin Romero Patricia Angélica Alves Marques

The escalating frequency of extreme hydrological events under environmental uncertainty poses a severe socio-economic threat to floodplains such as the Brazilian Pantanal, the world’s largest tropical wetland. Mitigating dynamic flooding and drying cycles is highly challenging due to a critical scarcity of in situ monitoring, leaving flood risks poorly understood. To address these data gaps, this study presents an advanced deep learning forecasting framework that integrates multisource environmental data, fusing satellite-derived precipitation (CHIRPS) and global land data assimilation evapotranspiration (GLDAS) data with historical river gauge telemetry. Multi-layered neural network architectures were optimized and combined with progressive moving average filters (10− and 15−day windows) to capture the complex hydrometeorological patterns of the data-scarce Miranda River Watershed. The optimal deep learning configuration, utilizing a robust two-hidden-layer topology (15 and 60 neurons), consistently outperformed standard baselines. Although purely exogenous data blocks successfully minimized satellite noise and captured seasonal trends (NSE ≥ 0.92), structural underestimation of peak flows was observed. When incorporating the previous day’s streamflow (lag t−1) as a physical anchor, this limitation was noticeably alleviated, increasing both the Nash–Sutcliffe Efficiency (NSE) and Coefficient of Determination (R2) values above 0.99. While this performance surge is driven by the strong temporal persistence inherent to the autoregressive lag, it introduces an operational trade-off by restricting the forecast to a reactive 24 h window. In this regard, an evaluation of the operational forecast horizons revealed that the exogenous deep learning blocks maximize warning lead times, providing a vital tool for proactive civil defense and disaster risk reduction. Ultimately, this multisource framework establishes a methodological foundation for automated decision support systems, providing the high-accuracy streamflow forecasting capability required to support future flood mitigation frameworks.

AI doi: 10.3390/ai7080294

Authors: Jinyin Bai Wei Zhu Xiangchen Wang Shiluo Guo Zongzhe Nie Tianjin Ni Jinji Zhou Kaiyang Kou Lingxin Xu Yihao Zhong

Post-disaster emergency communication recovery is not merely a link-repair task but a high-level planning problem constrained by service priorities, inter-object dependencies, resource budgets, and time windows. Existing restoration optimization methods generally rely on fully structured inputs, whereas direct large language model (LLM) planning may produce fluent candidates that violate encoded prerequisites, stage-order relations, budget limits, or temporal constraints. To address this challenge, we propose ICG-Restore, an intent-constrained, graph-enhanced LLM planning framework with rule-consistent minimal-edit repair. ICG-Restore transforms mixed restoration requests and structured network observations into task packages that can be checked for validator-level feasibility under an encoded high-level constraint model and evaluated by downstream abstract executors or schedulers. The framework compiles natural-language requests, structured observations, and operational rules into a task-intent object; retrieves task-relevant context from a heterogeneous scenario graph and a restoration knowledge graph; generates stage-wise restoration candidates; and applies bounded local corrections to candidates that violate encoded constraints. In this paper, “minimal-edit” is a descriptive label for a bounded local repair principle that prioritizes less disruptive corrections. Candidates accepted by the validators are evaluated and ranked by a safety-aware agent executor operating in an abstract restoration action space. Experiments on controlled abstract topologies covering three scales, four restoration tasks, and five environmental evolution modes show that ICG-Restore improves validator-level constraint satisfaction and benchmark-estimated recovery utility. Compared with Direct-LLM, it improves CSR and CRS by 1.99% and 24.56%, respectively; benchmark-specific WCTC@5 structural-alignment diagnostic increases by 38.87%.

AI doi: 10.3390/ai7080293

Authors: Gintautas Dzemyda Modestas Motiejauskas

Images may evoke different emotional responses depending on their content, style, and viewer interpretation. This is particularly important when evaluating works of art, architectural designs, or interior design choices. Visual emotion recognition (VER) models are commonly evaluated using supervised classification metrics such as accuracy, precision, recall, and macro-F1, together with general uncertainty indicators such as entropy, maximum softmax probability (MSP), and top-1/top-2 margin. However, these measures do not indicate whether the model’s strongest competing-emotion predictions remain within the same sentiment group or cross the positive–negative sentiment boundary. This paper proposes the top-2 cross-sentiment rate (CrossRate), a label-free measure for analyzing sentiment-level ambiguity in VER models. CrossRate measures the proportion of samples for which the top-1 and top-2 predicted emotion classes belong to opposite sentiment groups. The measure is evaluated on VER datasets using both standard classification metrics and uncertainty indicators. Experiments on EmoSet-118K show that varying the model’s parameters reduces CrossRate from (22.15±0.45)% to (7.81±0.62)% and increases accuracy from (79.13±0.16)% to (80.10±0.15)%. These changes are not fully reflected by entropy, MSP, or margin, indicating that CrossRate captures a complementary aspect of sentiment-level prediction behavior. The WikiArt case study further demonstrates that CrossRate can be applied when ground-truth emotion labels are unavailable. The proposed measure is applicable to any VER model whose predicted emotion classes can be mapped into positive and negative sentiment groups. The application of CrossRate is illustrated by its use in estimating the emotions of artworks. It offers even non-art experts the opportunity to form an opinion.

AI doi: 10.3390/ai7080292

Authors: Teng Shi Mingshan Xie Zhenxin Zhao Jifeng Gao

Timely and accurate identification of fish diseases, along with the provision of corresponding treatment plans, is crucial for improving fish welfare and reducing economic losses in aquaculture facilities. To address issues such as low detection efficiency and excessive reliance on manual experience in aquaculture, this paper proposes an innovative diagnostic and treatment method based on the Octopus framework for model optimization. The CBAM (Convolutional Block Attention Module) attention mechanism is introduced into the YOLOv11n backbone to enhance feature extraction, while the Focus-CIoU (Focus-Complete Intersection over Union) loss function is adopted to improve the detection performance for small-target lesions and dense fish schools. Additionally, by integrating wireless sensor networks and the OTDA (Octopus-inspired Therapeutic Decision Agent) bionic agent, an integrated ‘detection-identification-perception-recommendation’ framework is constructed. The results showed that the optimized YOLOv11n model achieved a detection precision of 98.00%, a recall of 94.86%, and an F1 score of 96.40% in detecting ulcer disease, tail rot disease, and red skin disease. Furthermore, the mAP@50 was 97.58%, and the mAP@50–95 was 85.68%. Compared with the traditional YOLOv11n model, the optimized model achieved improvements of 1.39%, 1.48%, 1.43%, 1.4%, and 2.74% in precision, recall, F1 score, mAP@50, and mAP@50–95, respectively. The treatment recommendations produced by means of the Octopus bionic intelligent agent resulted in an average system response time of no more than 53 s, with the maximum observed response time not exceeding 60 s. The identification accuracy of OTDA remained above 87.1%.