Search Results (356)

Artificial intelligence (AI) refers to a set of computational techniques, including machine learning, natural language processing, deep learning, and generative AI, that enable systems to perform tasks traditionally requiring human intelligence, such as prediction, pattern recognition, and decision-making. The rapid diffusion of AI into business organizations is transforming how information is processed, decisions are made, and knowledge-intensive work is performed, creating both new opportunities for economic value and new challenges for human judgment, organizational governance, and accountability. The growing adoption of AI across accounting, finance, and management makes it increasingly important to understand not only what AI can do, but also how and under what conditions it affects individuals, organizations, and markets. This paper provides a comprehensive review of the rapidly growing literature on artificial intelligence across these three disciplines. We synthesize existing research to examine how AI is transforming information processing, decision-making, governance, and organizational performance. In accounting, AI enhances auditing, financial reporting, and fraud detection while raising concerns regarding transparency and professional judgment. In finance, AI improves asset pricing, risk assessment, and trading strategies by leveraging large-scale structured and unstructured data. In management, AI reshapes organizational design, human capital, strategic decision-making, and innovation through increasingly sophisticated human–AI collaboration. Across these disciplines, we organize the literature around several unifying themes, including information asymmetry, automation versus augmentation, decision quality, interpretability, and governance. We further identify important research gaps concerning whether AI’s predictive and analytical advantages translate into meaningful economic and organizational outcomes, how AI reshapes human judgment and skills, the emerging risks, and the need for stronger research designs. By integrating evidence across three major business disciplines, this review provides a unified framework for understanding AI’s transformative role in organizations and offers a roadmap for future interdisciplinary research on the economic, behavioral, organizational, and governance consequences of AI. Full article

This scholarly article focuses on a specific application of discrete optimization methods in the emergency services. The search for the optimal deployment of service centers is one of the strategic decisions made in the field of urgent pre-hospital healthcare management. Since the consequences of the decisions are important for everyone and can directly affect the availability of the emergency medical service, different opinion groups are often taken into account when formulating a mathematical model. If there are two or more different conflicting objectives, the Pareto front of solutions usually needs to be constructed. It may serve as a good basis for finding the final system design. Since the construction of the exact Pareto set is very time-consuming and requires large computing resources, the efforts of many experts are focused on the development of efficient algorithms enabling the approximation of the original Pareto frontier in a short time. This paper introduces one of such heuristics. Even if the proposed algorithm of gradual refinement follows the idea of sequential processing of the current set of non-dominated solutions item by item inspecting the neighborhood of each element for possible extension of the Pareto front approximation, it can be simply adjusted and generalized making use of several parameters. Such an adjustment naturally raises the question of their optimal settings. Therefore, we gradually tried several procedures, from simple experimental verification of suitable values up to the development of sophisticated tuning of parameters based on machine learning methods. In this way, we created a complex advanced algorithm with elements of artificial intelligence. A series of numerical experiments are carried out utilizing real-world benchmarks that have their Pareto fronts applied in order to quantify and measure the efficacy of the proposed heuristic method. Full article

Lung cancer is the leading cause of cancer-related mortality worldwide, yet the majority of cases are diagnosed at advanced stages when a cure is rarely achievable. Low-dose computed tomography (LDCT) screening in high-risk populations reduces lung cancer mortality by 20–24% in randomised trials, but fewer than 20% of eligible adults in the United States undergo annual screening. We conducted a structured narrative review of PubMed, Embase, and the Cochrane Library for studies published between January 2002 and March 2026, supplemented by review of current clinical guidelines from the USPSTF, NCCN, ACS, CHEST, and ERS. Evidence from the National Lung Screening Trial (NLST) and the NELSON trial establishes the mortality benefit of LDCT screening, though both trials have important methodological limitations that affect generalisability. The NLST predominantly detected non-small cell lung cancer (NSCLC), particularly adenocarcinoma and squamous cell carcinoma, while small cell lung cancer (SCLC) was infrequently screen detected and showed no survival benefit from early detection. Guideline eligibility criteria have progressively broadened, and multivariable risk model-based selection using the PLCOm2012 now demonstrates superiority over categorical smoking thresholds in prospective validation cohorts, with the added benefit of reducing racial and ethnic eligibility disparities. Overdiagnosis estimates have declined substantially with extended follow-up, reaching 7% when observation exceeds five years. Implementation remains critically deficient: patient stigma, provider knowledge gaps, structural barriers, and inadequate electronic health record infrastructure collectively account for screening uptake below 20%. Integrating smoking cessation into screening encounters is evidence-based and cost-effective. Artificial intelligence tools show promising performance in nodule detection and risk prediction, but lack the prospective external validation required for routine clinical deployment. The field has established efficacy; the urgent challenge is now effectiveness at scale. Transitioning to risk model-based eligibility, expanding access to underserved populations, and mandating cessation integration represent the three highest-priority actions. A research agenda addressing never-smoker screening, personalised intervals, and robust AI validation must proceed in parallel. Full article

The increase in the number and complexity of interconnected systems requires new methods to identify potential threats in today’s hyperconnected world. This trend affects systems ranging from smart homes and Internet of Things (IoT) devices to critical infrastructure which must be equipped with the corresponding cyber defense methods. Given the new landscape, it is more difficult for classical cybersecurity systems to stay up to date with novel threats, as well as to keep track of all interconnected devices defined by various protocols and behaviors. Artificial intelligence (AI) represents a strong candidate to complement traditional cyber defense methods due to its adaptability to variation and capability to identify complex data patterns, which has led researchers to develop state-of-the-art anomaly detection systems. The current critical review aims to analyze the scientific literature on three dimensions including used algorithms and datasets, domain challenges hindering AI deployment in productive environments, and the capability of explainable artificial intelligence (XAI) to support cyber security experts with insights into the model’s inner workings and decision rationale. Compared to existing scientific reviews, this paper moves beyond algorithmic comparison by providing a methodological interpretation of AI anomaly detection landscape, demonstrating how data availability, learning paradigms, and explainability collectively influence the evolution of cyber defense research towards operational deployment. This approach revealed that AI development for cyber defense is highly heterogenous, and that the available datasets strongly influence the algorithm of choice, rather than the models being chosen methodologically based on proven performance. The analysis further indicates that operational deployment remains challenging, as the literature continues to report substantial limitations related to data quality, computational requirements, and model interpretability. Full article

Low-precision floating-point computation has become central to efficient artificial intelligence, yet its behavior for compact transformer-based hyperspectral imaging (HSI) models remains underexplored. In this work, we present a controlled comparative study of fully quantized training (FQT) and post-training quantization (PTQ) for pixel-wise foreign plastic object (FPO) classification in poultry hyperspectral data. Using a fixed state-of-the-art spatial–spectral transformer backbone, a common mixed-precision strategy, and identical training and inference protocols, we evaluate FP32, FP16, BF16, FP8, and NVFP4 across predictive performance, model compression, training efficiency, and inference efficiency. The results show that mixed-precision FQT remains highly robust across the tested precision spectrum, with all reduced-precision configurations staying within 0.63 percentage points of the FP32 baseline in overall accuracy while consistently outperforming PTQ at matched precisions. Across the evaluated formats, BF16 provides the closest accuracy to FP32, whereas FP8 offers a particularly favorable balance between accuracy preservation and reduced precision, while model compression increases progressively to $3.69\times$ under NVFP4. The computational benefits, however, are strongly workload dependent. Native FP8/FP4 hardware support does not automatically improve training throughput for this compact model at moderate workloads, and the larger training batches required to better utilize low-precision hardware can degrade predictive performance. In contrast, large-batch inference can effectively exploit FP8 and NVFP4 without affecting predictive accuracy. An ablation study further shows that selective retention of numerically sensitive modules in FP32 is essential for stable ultra-low-precision operation. Overall, the findings demonstrate that low precision is a viable but workload-dependent design choice for compact HSI transformers, with FQT providing greater accuracy robustness than PTQ and FP8, offering a favorable overall accuracy–efficiency trade-off. Full article

The posterior communicating artery (PComA) aneurysm is a challenging microsurgical problem due to multiple factors. These include the technical complexity of the surgery itself, potential injury to blood vessels during surgery, recovery of cranial nerve function, and overall neurological outcome after the operation. These are all related to the area where the PComA aneurysm is located, but they have fundamentally different biological determinants. Prior methods of describing aneurysms do not adequately describe how the relationships of the internal carotid artery (ICA) and PComA/P1 configuration influence the proximity of the aneurysm to other important structures such as the perforating arteries, the anterior choroidal artery (AChA), cranial nerve III (CN III), and the surgical corridor. We created PComA-CORE, an artificial intelligence-based framework designed to evaluate whether the elements of experienced microsurgeons’ thought processes can be measured individually while still maintaining temporally valid predictions, human interpretability, and explicit estimates of predictive uncertainty. Methods: Using a highly detailed database of clinical, radiographic, anatomical, intraoperative, and longitudinal data from 687 adult patients who underwent microsurgical clipping of PComA aneurysms over the period 1997–2026, we applied PComA-CORE to predict separately: Microsurgical Complexity (C); Oculomotor Recovery (O); Neurovascular Preservation Risk (R); and Expected 90-Day Functional Outcome (E). The models used cases from 1997–2020 (n = 564) for development and cases from 2021–2026 (n = 123) for temporal evaluation. Several architectures, including penalized regression, machine-learning techniques, interpretable machine learning, and ensembles, were compared using nested cross-validation, discrimination metrics, calibration metrics, decision-curve analysis, explainability measures, uncertainty-aware prediction, inter-observer reproducibility, and model-to-score distillation. Results: Four discrete predictive architectures were identified by PComA-CORE. PComA-C was found to be highly dependent upon anatomy because the specific geometric characteristics of individual vascular segments and the presence or incorporation of branches around the aneurysm strongly influenced temporal predictions. PComA-R was found to behave as a distributed susceptibility phenotype based on neurovascular attributes rather than a deterministic injury model and achieved a temporal AUC of 0.703. Among patients with preoperative CN III palsy, PComA-O identified that recovery primarily depended upon the time course of neurological dysfunction and structural deformation of the affected nerve. Temporal validation was not feasible given the small number of recent non-recovery events. Conversely, PComA-E showed that global functional outcome continued to depend predominantly upon clinical neurological severity, with a temporally evaluated penalized model achieving an AUC of 0.878. Uncertainty-aware predictions indicated that some cases would benefit from greater caution in interpretation. High-resolution anatomical phenotypes demonstrated good inter-observer reproducibility. Score distillation demonstrated that simplification preserved predictive information, but did so differently depending on the endpoint. Conclusions: The problem of predicting the consequences of clipping a PComA aneurysm is multidimensional and does not exist as a single “risk” prediction problem. Technical complexity, neurovascular vulnerability, neural recovery, and global disability each exist within distinct predictive spaces and require different levels of anatomical detail and/or computational complexity. PComA-CORE establishes a human-supervised framework to transform expert microsurgical thought processes into explicit, reproducible, uncertainty-aware, and clinically interpretable representations. While prospective multicenter validation will be needed prior to clinical use, it has the potential to establish a basis for explainable AI, precision cerebrovascular neurosurgery, anatomy-informed risk stratification, and clinically interpretable decision-support systems in complex aneurysm surgery. Full article

Artificial intelligence is increasingly shaping the preservation, interpretation, management, and reactivation of musical cultural heritage, especially in domains affected by technological obsolescence such as computer music, electroacoustic works, and analog audio archives. This perspective and case-study article examines the research trajectory of the Centro di Sonologia Computazionale (CSC) of the University of Padua through three analytically selected research directions: AI-assisted musical co-creation, multimodal generative systems informed by psychophysical research, and AI-based preservation of musical cultural heritage. The cases were selected because they represent complementary relationships between human expertise and computation (creative dialogue, perceptually grounded generation, and historically informed preservation) and because they include both completed studies and ongoing implementations. The article distinguishes previously published results, current implementations, and prospective components. It argues that the CSC case contributes a historically grounded model in which AI augments rather than replaces musical, archival, and interpretative expertise. Case studies include generative models for Disklavier performance, multimodal systems connecting sound with taste and emotion, a custom OCR and RAG workflow for Music V, and AI modules for analog-audio preservation within the MPAI/IEEE-CAE ARP framework. Full article

Background/Objectives: Concern that COVID-19 vaccination causes pulmonary nodules persists because vaccination coincided with expanded computed tomography (CT), low-dose CT (LDCT) screening, post-COVID imaging and artificial intelligence (AI)-assisted detection. This review asks whether the current literature supports causality and how vaccination history should inform thoracic imaging. Methods: A focused search of PubMed, PubMed Central and the Cochrane Library was performed through 15 July 2026. Evidence was classified as direct or contextual; no PRISMA screening, risk-of-bias scoring or quantitative synthesis was performed. Results: Direct evidence remains sparse. One case report was too confounded for inference. A two-sample Mendelian randomization study found no broad lung disease risk signal, but nodules were not modeled and the heterogeneous endpoints were exploratory. An ecological study of 1,616,750 samples linked rising detection to SARS-CoV-2 infection waves and AI-assisted reading; lacking individual vaccination data, it cannot establish whether vaccination affected detection. Screening interruption produced the opposite pattern: Lung-RADS 4 nodules rose from 8% to 29%. The most reproducible post-vaccination thoracic finding is regional lymph-node activation on [18F]fluorodeoxyglucose positron emission tomography/computed tomography ([18F]FDG-PET/CT), an expected immune response rather than a parenchymal nodule. Conclusions: Current evidence is insufficient to establish vaccination as an independent, population-level cause of pulmonary nodules. Vaccination history should guide [18F]FDG-PET/CT interpretation; CT-detected parenchymal nodules warrant standard risk stratification. Full article

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. Full article

Background: Infertility is a common reproductive health disorder that affects roughly 10–15% of couples during their reproductive period. Hysterosalpingography (HSG) is a widely utilized imaging modality for assessing uterine cavity morphology and fallopian tube patency and continues to play a central role in infertility investigations. Nevertheless, the interpretation of HSG findings may vary according to the experience and expertise of the evaluator, potentially leading to inconsistencies in clinical decision-making. Although artificial intelligence (AI) has demonstrated considerable potential in medical image analysis across various specialties, evidence regarding its application in the interpretation of HSG examinations remains scarce. Therefore, this study aimed to evaluate the level of agreement among radiologists, gynecologists, and an AI-based system in the assessment of identical HSG images. Methods: In this retrospective study, a total of 1443 HSG images obtained from 414 women who underwent hysterosalpingography as part of an infertility evaluation between January 2021 and January 2025 were reviewed. Cases with incomplete clinical records or suboptimal image quality were excluded from the analysis. All examinations were independently assessed by an experienced radiologist, a gynecologist specializing in infertility management, and a multimodal artificial intelligence system based on ChatGPT-5, with each evaluator blinded to the assessments of the others and to the patients’ clinical information. Image interpretation included the evaluation of contrast distribution, peritoneal spill, uterine cavity findings, tubal patency, and overall HSG impression, which were categorized according to predefined diagnostic criteria. The primary outcome was the degree of interobserver agreement among the evaluators. Agreement analyses were performed using Cohen’s kappa (κ) and Gwet’s AC1 coefficients. Analyses were conducted using IBM SPSS Statistics (version 30.0; IBM Corp., Armonk, NY, USA) and R statistical software (version 4.4.0; R Foundation for Statistical Computing, Vienna, Austria). Statistical significance was set at p < 0.05 (two-sided). Results: A total of 1443 HSG images obtained from 414 women were included in the final analysis. The mean age of the study population was 30.97 ± 5.59 years, and primary infertility accounted for 87.9% of cases. The average number of images acquired per examination was 3.49 ± 1.05. According to Cohen’s kappa analysis, the highest levels of agreement were observed for the assessment of image artifacts and contrast medium distribution. Agreement between the AI system and the radiologist was particularly strong for contrast medium distribution (κ = 0.757). For the overall interpretation of HSG findings, AI demonstrated substantial agreement with the radiologist (κ = 0.637), exceeding the level of agreement observed between the radiologist and the gynecologist (κ = 0.363). In contrast, concordance involving AI was lower for the evaluation of uterine abnormalities, intrauterine filling defects, and tubal patency. When agreement was reassessed using Gwet’s AC1 statistic, concordance coefficients were consistently higher than the corresponding kappa values across all evaluator pairs. Near-perfect agreement between AI and the radiologist was identified for contrast medium distribution (AC1 = 0.954), peritoneal spill (AC1 = 0.893), and patterns of peritoneal contrast passage (AC1 = 0.841). Procedures performed under local anesthesia yielded a significantly greater number of images than those conducted under general anesthesia (3.86 ± 0.86 vs. 3.08 ± 1.10, p < 0.001). No significant associations were detected between abnormal HSG findings and either infertility type or anesthetic technique. In multivariable analysis, the use of general anesthesia was independently associated with a lower image count, whereas the presence of tubal pathology emerged as an independent predictor of acquiring a greater number of images during the examination. Conclusions: Our findings indicate that AI-assisted interpretation of HSG images has the potential to complement expert assessment, showing substantial concordance in several key diagnostic domains. While the technology appears promising as a decision-support tool in infertility evaluation, further research and refinement are warranted, particularly regarding the assessment of tubal and uterine pathologies. Full article

Background/Objectives: Colorectal carcinoma (CRC) remains a leading cause of cancer-related morbidity and mortality worldwide, with increasing incidence in younger populations. Despite advances in imaging, conventional approaches remain limited by subjective interpretation and insufficient characterization of tumor heterogeneity. Radiomics, particularly in a multiparametric framework integrating computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET), has emerged as a promising tool to enhance diagnostic and prognostic performance. This review aims to critically evaluate methodological strategies for optimizing multiparametric radiomics in CRC. Methods: A narrative review was conducted based on a literature search of PubMed/MEDLINE, Scopus, and Web of Science up to March 2026. Studies focusing on CT-, MRI-, and PET-based radiomics in CRC were included. Key methodological aspects analyzed included imaging acquisition and standardization, tumor segmentation techniques, radiomic feature extraction, feature selection methods (e.g., LASSO and PCA), and model validation approaches. Results: Multiparametric radiomics models integrating CT, MRI, and PET consistently demonstrated superior diagnostic accuracy compared to single-modality approaches, particularly in T-staging and lymph node involvement prediction. PET-derived metabolic features further enhanced characterization of tumor biology and improved prognostic stratification, including prediction of progression-free survival (PFS) and overall survival (OS). However, methodological heterogeneity, small sample sizes, and variability in imaging protocols and segmentation practices remain significant limitations affecting reproducibility and generalizability. Conclusions: Multiparametric radiomics represents a powerful advancement in precision oncology for CRC, enabling improved tumor characterization, risk stratification, and personalized treatment planning. Standardization, multicentric validation, and integration with artificial intelligence are essential for successful clinical translation. Full article

Background/Objectives: Health literacy and patient education play a key role in supporting informed decision-making, treatment adherence, and disease self-management. While educational interventions have traditionally focused on improving knowledge and understanding, emotions are increasingly recognized as factors associated with learning processes and healthcare experiences. This scoping review mapped the role of emotions in health literacy and patient education by identifying the emotions investigated, assessment methods, technologies employed, educational outcomes, and opportunities for emerging technologies. Methods: A scoping review was conducted following PRISMA-ScR and Joanna Briggs Institute guidance. Searches were finalized in Scopus and PubMed on 6 June 2026, with no publication-date restrictions; only English-language journal articles and reviews were eligible. Two reviewers screened records, extracted data, and classified studies, resolving disagreements by consensus and consulting methodological or clinical co-authors when needed. Results: A total of 135 studies were included. Anxiety and emotional distress were the most frequently investigated emotional categories, whereas emotional support, reassurance, and emotional engagement received less attention. Interviews, surveys, questionnaires, and standardized scales were the predominant assessment methods. Emotional well-being and understanding were the most frequently reported educational outcomes, followed by adherence, communication, and emotional support. Educational programs and educational materials were the most commonly employed approaches, while digital health solutions appeared less frequently. Limited use of artificial intelligence and objective emotion-recognition methods was identified. Conclusions: Emotional factors were associated with several patient-education outcomes, but the heterogeneous and predominantly self-reported evidence does not establish causality. Emerging technologies warrant further evaluation before their effectiveness, feasibility, and safety in patient education can be established. Full article

Multi-access edge computing (MEC) enables computation-intensive perception and decision-making tasks in vehicular networks to be offloaded to nearby edge servers. Existing approaches usually fix the artificial intelligence (AI) inference model, overlooking how model selection jointly affects latency, energy consumption, and service reliability. We propose a hierarchical model selection and control (HMSC) framework based on deep reinforcement learning (DRL) for MEC-assisted vehicular networks. The framework couples a vehicle-layer MAPPO component that provides a communication interface representation for subchannel assignment and energy accounting with a centralized MEC-layer soft actor-critic (SAC) agent that, under SDN orchestration, adaptively selects lightweight or high-fidelity AI models and allocates computational resources. Accordingly, the core contribution of this paper lies in MEC-side model-aware computation control under an explicitly defined subchannel-contention abstraction, rather than in physical-layer transmit-power optimization. Both layers are guided by a composite objective that integrates normalized end-to-end (E2E) latency, normalized energy consumption, and a deadline-violation penalty. Using a discrete-time simulation framework, HMSC reduces E2E latency compared with static inference and non-hierarchical DRL baselines and sustains a higher deadline satisfaction ratio (DSR) under constrained uplink throughput and varying traffic loads. The learned policy is load-aware, favoring high-fidelity inference under light load and lightweight inference under congestion; a post hoc analysis using YOLOv5-family accuracy reference further quantifies the inference-quality implications of this adaptive selection behavior. These results show that coordinated MEC-side control of AI model selection and computation, under a shared deadline-aware objective, provides a robust latency–energy trade-off for MEC-assisted vehicular networks. Full article

Hydrogen (H₂) is expected to play a critical role in modern industry, particularly in ammonia synthesis, petroleum refining, and low-carbon transportation. The safe storage of H₂ remains a major challenge due to its low volumetric density under ambient conditions. Metal–Organic Frameworks (MOFs), highly porous crystalline materials, have emerged as promising H₂ storage candidates owing to their high surface areas and tuneable pore structures. Molecular simulations such as grand canonical Monte Carlo or density functional theory are costly and limited in exploring large material spaces, motivating efficient predictive tools to accelerate discovery. Here, Machine Learning (ML) techniques are compared to an explainable artificial intelligence (XAI) approach using symbolic regression (SR), trained on 10,123 experimentally measured H₂ adsorption datapoints from real-world MOFs. The best performing model achieved a goodness of fit of 0.9986 with lower computational demand, but reduced interpretability, addressed using XAI analysis and clustering. SR achieves a lower goodness of fit of 0.914 but produces a physically meaningful equation highlighting structural features driving high gravimetric efficiencies. These results demonstrate strong ML capability for predicting how MOF properties and environmental conditions affect H₂ uptake. This offers engineers and researchers a practical means of screening potential MOFs for H₂ storage applications, with the XAI analyses providing additional confidence in the predictions. They allow researchers to understand the physical reasoning behind each output, assess the reliability of individual predictions, and make fully informed decisions, enabling predictive models to be acted upon with confidence in real-world contexts. Full article

Background/Objectives: Obstructivesleep apnea (OSA) is one of the most widespread forms of sleep disease, affecting over 936 million adults globally. The health consequences of obstructive sleep apnea (OSA) are well documented; however, it remains largely underdiagnosed because the current gold-standard diagnostic method, polysomnography (PSG), is often costly, time-consuming, and unavailable in many healthcare settings. To address these challenges, this study presents a novel explainable deep learning (DL) framework for automated multi-class OSA severity classification using single-lead electrocardiogram (ECG) signals. Methods: The proposed framework integrates a hybrid CNN–BiLSTM architecture with explainable artificial intelligence (XAI) techniques to generate clinically meaningful predictions and explanations across four OSA severity classes: Normal, Mild, Moderate, and Severe. The framework was evaluated using the publicly available PhysioNet Apnea-ECG dataset (70 recordings) together with an institutional ECG dataset (150 recordings), resulting in a combined cohort of 220 recordings. Results: The proposed framework achieved an overall classification accuracy of 94.7%, with sensitivity and specificity values of 92.3% and 96.1%, respectively. Furthermore, the proposed model consistently outperformed conventional machine learning algorithms, including Support Vector Machine (SVM), Random Forest, and XGBoost, by 5.5%, 4.2%, and 2.9%, respectively. To enhance transparency and clinical trust, SHAP (SHapley Additive exPlanations) was employed to identify the most influential physiological predictors driving model decisions. Heart rate variability features, particularly RMSSD and pNN50, emerged as the strongest indicators of OSA severity. Moreover, computational efficiency analysis revealed that the model required only 0.23 s to process a 60 s ECG epoch on a standard computing platform, supporting its suitability for real-time deployment. Conclusions: The findings demonstrate that explainable deep learning applied to ECG signals can provide accurate, interpretable, and computationally efficient assessment of OSA severity. The proposed framework may support OSA screening, clinical triage, and early intervention, particularly in resource-constrained healthcare environments. Full article