Search Results (190)

Online business reviews encode rich customer sentiment that is critical for commercial decision making, yet accurately predicting star ratings from free text remains a challenging five-class classification problem. Classical ensemble methods—Soft Voting, Weighted Voting, and Stacking—aggregate complementary base-model outputs to improve predictive performance, but they produce opaque decisions that are unintelligible to business stakeholders. This paper proposes using a large language model (LLM), specifically unsloth/LLaMA-3.3-70B-Instruct, as an Explainable AI (XAI) ensemble aggregator: the LLM receives the predictions and confidence scores of four heterogeneous base models (Logistic Regression, Support Vector Machine, Naïve Bayes, and BERT-base-uncased) and reasons over them to produce both a final star-rating prediction and a natural-language explanation. We evaluate the full pipeline on 10,000-sample balanced and natural-distribution test sets derived from the Yelp Academic Dataset, with additional cross-lingual validation on Spanish Amazon Reviews. The LLM aggregator (LLAMA_AGG) achieves the highest macro-F1 on both pipelines (0.6800 on balanced; 0.6720 on natural) and the best ordinal calibration (QWK = 0.9111 on balanced; 0.9337 on natural), outperforming all classical aggregators and base models. A detailed Explainable AI analysis reveals that the LLM revises 28.07% of its standalone predictions after observing the ensemble outputs, improving the accuracy by +22.2 percentage points on the revised cases. The aggregator corrects severe polar bias in the standalone LLM (±0.35 recall improvement on mid-range star classes) and produces longer explanations when evidence is conflicted—a quantitative signal of deliberative reasoning. A formal human evaluation with two judges confirms high explanation faithfulness (4.47/5) and readability (4.82/5). Model scale ablation shows an 8B parameter variant achieves 90.8% agreement with the 70B model, enabling practical deployment. These findings demonstrate that Explainable AI can be achieved through LLM-based ensemble aggregation, establishing a principled approach for business-review sentiment analysis. Full article

Low-resource hate speech detection remains a challenging task for Indonesian social media due to limited labeled annotations, highly informal linguistic expressions, and substantial lexical variability. Under such conditions, purely supervised transformer models often suffer from unstable semantic generalization, while conventional pseudo-labeling methods are vulnerable to noisy unlabeled sample propagation. To address these limitations, this study proposes a hybrid feature-enhanced IndoBERT framework integrated with a controlled semi-supervised learning strategy. The proposed model combines contextual IndoBERT embeddings with abusive lexicon cues, handcrafted linguistic indicators, and TF-IDF–SVD statistical representations through a lightweight concatenation–projection feature fusion mechanism, while unlabeled data are incorporated via adaptive confidence thresholding and class-balanced pseudo-label selection to improve pseudo-label reliability. Extensive experiments were conducted under realistic low-resource supervision settings using only 5%, 10%, and 20% labeled data, and the proposed framework was systematically compared against representative baselines, including sparse lexical machine learning models, shallow neural architectures, multilingual transformers, IndoBERTweet, naive pseudo-labeling, and LLM-based prompting. The results show that model effectiveness is strongly supervision-dependent. Under the most extreme low-resource setting, compact statistical augmentation provides the most stable complementary signal, whereas under moderate low-resource supervision, the full hybrid representation combined with controlled semi-supervised learning yields the strongest and most consistent gains. The proposed Hybrid IndoBERT + controlled SSL framework outperforms all baselines at the 20% labeled setting, reaching an accuracy of 0.8654, Macro-F1 of 0.8633, and ROC-AUC of 0.9334. Additional analyses of pseudo-label reliability, calibration behavior, computational efficiency, and qualitative error patterns further show that the proposed framework improves low-resource robustness while maintaining comparable inference-time efficiency. These findings demonstrate that low-resource hate speech detection benefits most from the staged integration of contextual semantic modeling, interpretable linguistic cues, global lexical–statistical structure, and carefully regulated unlabeled data exploitation. Additional experiments using GPT-4o-mini and Llama-3.1-8B further demonstrate that the proposed framework remains competitive against general-purpose large language model prompting approaches under low-resource Indonesian hate speech detection scenarios. The proposed framework provides a practical and reproducible direction for hate speech detection in annotation-constrained social media environments. Full article

Information cascades in online social networks arise from stochastic interactions among user behavior, temporal activation, and graph-structured exposure. Early prediction of cascade outcomes remains difficult because only a short diffusion prefix is observable, while future propagation depends on sparse user-level transitions across a heterogeneous social network. This study develops a graph-conditioned stochastic modeling framework for early Twitter cascade prediction. Retweet cascades are formulated as history-dependent stochastic processes over a finite user vocabulary, and a causal dual-head Transformer is used to infer cascade virality and logarithmic final size from short observed prefixes. To incorporate social-network structure, user embeddings pretrained from the follow graph are introduced as external structural priors. A controlled ablation design separates the effects of random embeddings, graph-pretrained embeddings, frozen structural priors, and handcrafted feature fusion. Experiments on Higgs Twitter retweet cascades show that direct full-vocabulary next-user prediction is statistically fragile under sparse short-prefix observations, motivating macro-level cascade outcome prediction. Among the evaluated configurations, the frozen graph-pretrained Transformer achieves the strongest overall balance, reaching an AUC of 0.819, a Brier score of 0.151, and an RMSE of 0.192, while the causal Transformer without a graph prior already surpasses logistic regression and approaches Random Forest; however, gains over competitive baselines are modest and statistically significant only in selected pairwise comparisons. Calibration analysis, bootstrap confidence intervals, and paired statistical tests confirm that graph-derived user priors provide more reliable improvements than sequence modeling alone under short-prefix sparse observations. These findings indicate that graph-conditioned structural priors offer a promising complement to causal sequence modeling for early Twitter cascade prediction. Full article

This study explains how reinforcement learning (RL) contributes to congestion mitigation in inland freight terminal operations by testing a serial process model in which RL strengthens operational learning stability (OLStab), OLStab improves logistics efficiency, and logistics efficiency lowers congestion. Rather than presenting RL as a stand-alone congestion-reduction instrument, the paper examines a distinct inland-terminal application in which congestion emerges from interacting gate, yard, transfer, and dispatch frictions. Using a simulation-based explanatory design calibrated to a realistic macro-logistics context, and interpreting the results as simulation-informed evidence rather than direct field proof, the study analyzes 500 episode-level observations representing complete terminal runs under varying control conditions. The results show that RL positively affects OLStab, OLStab positively affects logistics efficiency, and logistics efficiency negatively affects congestion. The serial indirect pathway from RL through OLStab and logistics efficiency to congestion is statistically significant, whereas the direct effect of RL on congestion becomes non-significant once the mediators are introduced. Decision latency sensitivity does not significantly moderate the RL-to-OLStab relationship, suggesting that latency-related boundary conditions are more context-specific than the main capability pathway. The article contributes by offering a cautious simulation-based and mechanism-centered explanation of RL-enabled congestion mitigation in inland terminals, by treating OLStab as a simulation-grounded intermediate operational stability index, and by showing that the empirical pattern is better explained by theory-ordered simulator-level mechanism than by a residual direct RL effect. Full article

Background/Objectives: Inheritance classification in rare diseases remains challenging because curated knowledge is incomplete, heterogeneous, and imbalanced across inheritance categories. Disease-level inheritance modeling can support knowledge organization, annotation review, and hypothesis generation in rare-disease resources. This paper introduces RareFusion-Net, a multimodal benchmark framework for disease-level inheritance classification, and evaluates whether integrating ontology-enriched disease text, structured epidemiological metadata, and gene-association information improves prediction in curated rare-disease knowledge bases. RareFusion-Net is intended for knowledge modeling, not individual patient diagnosis. Methods: We developed RareFusionBalanced, a gated multimodal fusion model that combines biomedical disease descriptions, structured metadata, and gene-related information using auxiliary supervision. Ontology-enriched disease text was treated as the dominant semantic modality, while tabular and gene modalities were incorporated as complementary evidence when available. Robustness was improved using balanced regularization, selective transformer fine-tuning, dropout, weight decay, label smoothing, early stopping, and prediction aggregation across random seeds. Evaluation included accuracy, macro-F1, micro-F1, macro-AUC, mean average precision, calibration metrics, class-wise analysis, statistical testing, and ablation experiments. Results: RareFusionBalanced achieved 0.7382 test accuracy, 0.6284 macro-F1, 0.7382 micro-F1, 0.9183 macro-AUC, and 0.6686 mean average precision. Calibration was favorable, with an expected calibration error of 0.0395 and a Brier-OVR of 0.0528. The multimodal model slightly outperformed TextOnly-TransformerBalanced, but improvement over the best TF-IDF baseline was not statistically significant. Ablation showed ontology-enriched text as the strongest modality, with gene associations adding complementary value. Conclusions: RareFusion-Net provides a practical benchmark for ontology-aware rare-disease inheritance modeling. Results suggest selective multimodal benefit while highlighting minority-class difficulty, limited statistical superiority, need for external validation, and improved biological interpretability. Full article

Early non-invasive correction of neonatal auricular deformities is highly dependent on timely and precise diagnosis. However, clinical practice is often compromised by the subjectivity of visual assessments and the lack of objective tracking metrics, which frequently leads to missed optimal treatment windows. To address these challenges, we developed an interpretable deep learning-based diagnostic system for the automated screening and fine-grained classification of these deformities. Methodologically, a large-scale, multi-source dataset (n = 4644) was curated to support model training. The system pairs an automated object detector (YOLOv11) for background-reduced region-of-interest isolation with a cascaded classification pipeline optimized via ConvNeXt-Tiny. Crucially, we introduced a supervised contrastive learning module to project high-dimensional morphological features into a continuous severity score, enabling quantitative longitudinal tracking of therapeutic efficacy. To evaluate generalization and robustness, the framework underwent rigorous evaluation across three independent real-world cohorts and one controlled synthetic stress test. The system achieved 88.2% accuracy (Area Under the Curve (AUC): 0.949) in binary screening and 87.4% accuracy (macro-AUC: 0.976) in multi-class subtyping on the internal baseline. To enhance interpretability and build clinical trust, Gradient-weighted Class Activation Mapping (Grad-CAM) was utilized to explore the spatial distribution of the model’s attention, which frequently aligned with key anatomical landmarks. Furthermore, the learned severity scores robustly quantified post-intervention improvements (p = 0.0004), effectively capturing subtle anatomical normalization. While validation for rare subtypes remains underpowered, and the severity score currently functions mainly as a learned morphological similarity index requiring future clinical calibration, this study ultimately provides an objective and standardized web-based tool to facilitate the early intervention and precision management of neonatal auricular anomalies. Full article

Student dropout remains an important challenge in higher education because it affects degree completion, institutional resource efficiency, workforce preparation, and students’ long-term socioeconomic opportunities. This requires not only accurate predictions but also decision support that is both reliable and aware of uncertainty. This study posits that the amalgamation of probabilistic modeling, uncertainty quantification, and graph-based refinement can augment both predictive reliability and decision support for the early detection of dropouts. A reliability-centered predictive framework is presented, integrating Educational Competition Optimization (ECO)-based feature selection, probabilistic Support Vector Classification (SVC), isotonic regression for probability calibration, and split conformal prediction for distribution-free uncertainty quantification. In addition, a similarity-driven Graph-based Fuzzy Cellular Automata (Graph-FCA) refinement mechanism is developed, where student relationships are modeled using a k-nearest neighbor graph with radial basis function similarity. Entropy-based confidence weighting is used to control uncertainty-aware propagation. An Explainable Artificial Intelligence layer based on SHAP provides both global and local interpretability, and fairness-aware evaluation assesses consistency across demographic groups. The suggested framework maintains predictive performance while improving probabilistic reliability. The Graph-FCA refinement achieves an accuracy of 0.7503, which is close to the calibrated ECO–SVC baseline (Accuracy = 0.7537; Macro-F1 = 0.6704) and also reduces the Brier score. The conformal prediction layer achieves empirical coverage close to the desired confidence level, ensuring reliable uncertainty estimates. The ECO–SVC–Conformal–GraphFCA framework transforms traditional classification into a reliable, understandable, and uncertainty-aware early warning system, enhancing its usefulness for ethical and informed decision-making in engineering education. Full article

Friction at the cold rolling interface is affected jointly by the surface roughness, lubrication state, local pressure, and relative sliding. A constant friction coefficient is therefore insufficient to describe its non-uniform distribution along the contact arc. Accordingly, this study proposes a macro–micro two-scale mixed-lubrication and dynamic friction model based on the measured roll roughness. First, the measured roll roughness profile was represented within a finite effective scale interval by a scaled and truncated Weierstrass–Mandelbrot (W–M) function. The parameters D and G were obtained as finite-scale W–M roughness parameters and were introduced into a mixed-lubrication load-sharing model to calculate the local mixed-lubrication friction coefficient. The pressure distribution along the contact arc was calculated using the Karman equation, and the local macroscopic pressure was mapped to a representative microscopic contact load. Finally, the mixed-lubrication friction coefficient was used to calibrate the dynamic friction factor separately in the forward-slip and backward-slip zones, and the friction stress distribution along the contact arc was calculated. For the selected effective scale interval and preprocessing procedure, the fitted W–M roughness parameters were D = 1.528 and G = 9.15 × 10⁻⁸ m. The W–M parameter D had a more significant influence on the mixed-lubrication friction coefficient and load-sharing behavior than the scale parameter G. Increasing the rolling speed strengthened the oil-film load-carrying effect and reduced the equivalent interfacial friction coefficient. The friction stress was positive in the backward-slip zone and negative in the forward-slip zone, with a direction reversal near the neutral point. Field forward-slip inversion showed that both the simulated and measured equivalent friction coefficients decreased with increasing rolling speed, with a difference of approximately 0.009~0.017. The proposed model can capture the main trend of cold rolling interfacial friction with variations in the rolling speed and contact state. Full article

Zero-shot remote sensing is attractive for scene classification because new regions, sensors, and label taxonomies often appear before sufficient annotated data are available for supervised adaptation. We present OATS-RS, an inference-centric framework that keeps a remote sensing vision–language model (VLM) backbone frozen and improves zero-shot decisions through ontology-aware prompt construction, hierarchical and contrastive scoring, adaptive multi-view aggregation, unlabeled transductive refinement, ambiguity-aware local re-ranking, and selective prediction. The method targets the common remote sensing regime in which neighboring classes such as annual crop, permanent crop, forest, pasture, herbaceous vegetation, river, and sea or lake overlap strongly in red–green–blue (RGB) appearance, meaning that they require more than a single class-name prompt. On the supplied final EuroSAT RGB evaluation with a GeoRSCLIP Contrastive Language–Image Pre-training (CLIP)-family Vision Transformer Base with 32 × 32-pixel patches (ViT-B-32) backbone, the complete pipeline obtains top-1 accuracy of 0.522, balanced accuracy of 0.522, macro-averaged F1 score (macro-F1) of 0.535, and top-3 accuracy of 0.887. The strongest classes are industrial area, residential area, river, highway, and pasture, whereas the weakest classes remain herbaceous vegetation and several fine-grained vegetation categories. Selective prediction increases accepted-example accuracy to 0.538 at 0.934 coverage, but the expected calibration error (ECE) remains high at 0.384. These results support a qualified conclusion: ontology-guided zero-shot inference can already recover useful semantic shortlists for structured remote-sensing scenes, but fine-grained natural-class disambiguation, calibrated confidence, multi-dataset transfer, component-level ablations, and measured runtime remain essential before dependable deployment claims can be made. Full article

Modern distribution networks undergo frequent topology reconfiguration, volatile bi-directional flows, and noisy measurements, making five-class line-loss anomaly classification both valuable and challenging. In this study, PC-LossGNN is proposed—a physics-consistent spatiotemporal graph neural network for edge-level classification into Normal, Infrastructure, Documentation, Metering, and Theft. A static topology prior is fused with a measurement-adaptive graph and confidence-aware multi-source features; power-flow physics is injected via residual-guided attention using active/reactive balance, voltage-drop, and ohmic-loss residuals. A dual-path decoder is employed to yield calibrated probabilities and interpretable class evidence, trained under an uncertainty-weighted curriculum objective. On six months of real utility data, macro-F1 of 0.8503 and accuracy of 0.9915 are achieved, surpassing XGBoost, LSTM, GCN, STGCN, and two recent physics-aware spatiotemporal GNN baselines including ST-RGNN and PA-STGCN. Ablation indicates that physics-consistent regularization is pivotal, while adaptive topology and interactive temporal encoding further improve performance. Robustness tests with injected Gaussian noise show more graceful degradation than baselines. These results suggest that PC-LossGNN provides accurate, physically plausible, and interpretable five-way line-loss diagnostics suitable for real-world operations. Full article

To establish a quantitative mapping relationship between macro- and micro-parameters in the PFC^2D parallel bonding model, and in view of the inherent complexity of the mutual validation process between laboratory experiments and parameter calibration, this paper takes uniaxial compression tests as the mechanical reference. By combining orthogonal experimental design, Pearson correlation analysis and multivariate analysis of variance, this study systematically investigates the effects of 10 micro-parameters on 6 macro-mechanical indicators (modulus of elasticity E, Poisson’s ratio ν, uniaxial compressive strength ${\sigma }_c$, friction-to-cohesion ratio FCR, crack initiation strength ${\sigma }_{ci}$ and crack damage stress ${\sigma }_{cd}$). To reduce the coupling dimension between cohesion and internal friction angle in the calibration of PFC macro–micro parameters, this paper defines the Friction-to-Cohesion Ratio (FCR) as the ratio of the equivalent macroscopic angle of internal friction to the equivalent macroscopic cohesion, and systematically conducts sensitivity analyses of uniaxial compression simulations. The results indicate that the elastic modulus E is primarily governed by $E^{\ast }$, ${\overline{E}}^{\ast }$, ${\overline{k}}^{\ast }$ and $R_f$; the Poisson’s ratio ν is mainly influenced by $E^{\ast }$, $k^{\ast }$, ${\overline{E}}^{\ast }$, ${\overline{k}}^{\ast }$ and $R_f$; the uniaxial compressive strength ${\sigma }_c$, the crack initiation strength ${\sigma }_{ci}$ and the crack damage stress ${\sigma }_{cd}$ are primarily regulated by ${\overline{\sigma }}_c$ and $R_f$; whilst the Friction-to-Cohesion Ratio (FCR) is mainly affected by ${\overline{\sigma }}_c$, $\overline{\phi }$, $R_f$, $\overline{c}$ and $\beta$; Elasticity parameters and strength parameters are governed by different micro-mechanisms, reflecting the fundamental decoupling of stiffness and strength in the PFC model. This study established a progressive ‘screening–validation–quantification’ sensitivity analysis framework, revealing the directional regulation patterns of various micro-parameters on macroscopic responses, thereby providing a theoretical basis for the targeted optimisation and efficient calibration of micro-parameters in PFC discrete element simulations. Full article

Seagrass meadows support critical coastal ecosystems, but corresponding species-level remote sensing data remain limited, particularly in the Arabian Gulf, where field spectral data for dominant taxa are extremely limited. We present the first multi-species spectral characterisation of three dominant seagrass species in the Kingdom of Bahrain—Halophila stipulacea (n = 46 spectra, 25 stations), Halodule uninervis (n = 34, 19 stations), and Halophila ovalis (n = 17, 8 stations)—measured with an ASD FieldSpec^® 4 Hi-Res spectroradiometer (Malvern Panalytical, Malvern, UK; 350–2500 nm) from samples collected across 29 geographic stations (52 species–station sampling units). All sample counts reported here underwent quality control. Kruskal–Wallis tests with Benjamini–Hochberg (BH) correction, Jeffries–Matusita (JM) distance, Hedges’ g, and linear discriminant analysis (LDA) were used to characterise inter-species differences. H. ovalis was clearly distinguished from both co-occurring species: the Hd. uninervis–H. ovalis pair showed a discriminating window of 692–1394 nm (mean |g| = 1.31, BH q = 0.000046), and that for the H. stipulacea–H. ovalis pair was 700–1376 nm (mean |g| = 1.21, BH q = 0.000285); the JM distances were 1.60–1.67. A secondary shortwave-infrared discriminating window (1607–1755 nm; mean |g| = 0.90, BH q = 0.006) was also identified for the Hd. uninervis–H. ovalis pair. The H. stipulacea–Hd. uninervis pair showed meaningful geometric separation (JM = 0.994) but no individually significant wavelengths at the available sample size. ASentinel-2-proxy LDA achieved 85.6% overall accuracy (balanced accuracy = 87.3%; macro area under the curve = 0.917), outperforming a Landsat-proxy model by 20 percentage points. For each species, both a best-overall index and a visible-range alternative optimised for submerged satellite remote sensing are reported. The primary indices achieved balanced accuracies of 0.877–0.924; the visible-range alternatives achieved 0.818–0.907. Performance degraded substantially under noise (σ ≥ 0.002: −7.5 percentage points [pp]) and wavelength misregistration (±2–3 nm shifts caused losses of 5.5–15.7 pp), calling for stringent calibration requirements. These results constitute the first multi-species spectral library for Kingdom of Bahrain seagrasses, supporting Sentinel-2-based species mapping in the Arabian Gulf. Full article

Rapid and non-contact surface inspection is essential for quality control in modern production. Optical coherence tomography (OCT) can image a surface without contact, but turning those images into roughness parameters usually requires specialized processing software. This study examined whether a multimodal large language model (LLM) could estimate roughness parameters directly from OCT B-scans as a screening tool. The study was designed as a controlled macro-scale proof of concept using periodic, analytically defined phantoms rather than as validation on stochastic industrial micro-roughness. Five test surfaces with exactly known geometries were designed, 3D-printed, and scanned with a spectral-domain OCT system. For each surface, roughness values were computed from the theoretical shape, extracted from the OCT image using MATLAB, and also estimated by the LLM from the same image. The repeatability of the LLM was checked by running the same prompt ten times per surface. On a sawtooth profile, the LLM estimates varied by 3.8% for Ra, 4.2% for Rq, 3.5% for Rp, 2.8% for Rv, and 3.1% for Rt. Across all five surfaces, the variation in Ra and Rq was around 3–5%, and for Rt, it stayed below 5%. The results show that a generative AI approach can produce repeatable roughness estimates that are useful for comparative screening. This method offers a flexible option for surface comparison and AI-assisted quality control when calibrated measurements are not required. Full article

Autonomous unmanned aerial vehicles (UAVs) must decide when to trust onboard perception, when to request edge support, and when to avoid acting under poor visual or communication conditions. This study develops a risk-aware edge-assisted UAV perception framework that combines calibrated visual confidence with next-window service-level agreement (SLA) feasibility. The local branch uses MobileNetV3-Small for fast onboard color recognition, while the edge branch uses ResNet-18 for stronger remote inference. Low-confidence samples are offloaded only when the SLA predictor estimates that the wireless link is feasible; otherwise, the system enters fallback, meaning that the current prediction is not treated as immediately actionable. The evaluation follows a hard cross-illumination split: indoor and fluorescent light samples are used for training and validation, and indoor night and sunlight samples are reserved for testing. Under this setting, the local model achieves 76.89% accuracy and 73.25% macro-F1, while the edge model achieves 81.26% accuracy and 77.58% macro-F1. The SLA predictor, trained on enhanced telemetry features while preserving the original target label, achieves 85.74% accuracy, 85.57% macro-F1, 0.9420 ROC-AUC, and 0.9585 PR-AUC on temporally held-out records. The joint policy achieves 93.23% coverage and 79.90% success over active decisions, using local inference for 82.76% of the samples, edge offloading for 10.47%, and fallback for 6.77%. These results indicate that the framework is best understood as a tunable risk management layer for UAV perception rather than a pure accuracy maximization classifier. It avoids blind offloading and reduces forced decisions when both visual confidence and communication feasibility are weak. Full article

Background/Objectives: Accurate bone lesion classification on radiographs is critical for clinical decision-making and forensic identification. Existing deep learning approaches treat radiographs as whole images, neglecting available spatial annotations and clinical metadata. To develop an ROI-guided deep learning framework integrating clinical metadata for three-class (Normal, Benign, Malignant) bone lesion classification and to assess its clinical safety profile. Methods: Using the BTXRD (3746 radiographs: 1879 Normal, 1525 Benign, 342 Malignant), an EfficientNetV2-S backbone was combined with an 11-dimensional metadata MLP trained on ROI-cropped regions. Training employed Focal Loss with adaptive class weighting, Mixup/CutMix augmentations, Stochastic Weight Averaging, and Test-Time Augmentation. Five-fold stratified cross-validation with bootstrap confidence intervals (n = 2000) and probability calibration metrics were used. Results: The framework achieved 96.05% accuracy (95% CI: 95.41–96.66%), 93.94% balanced accuracy, 92.62% macro F1-score, and 99.21% macro-AUC (95% CI: 98.89–99.42%). Critically, near-zero Malignant-to-Normal misclassifications occurred (1/342, 0.29%; 95% Clopper–Pearson CI: 0.01–1.62%) across all 3746 predictions. The minority Malignant class attained F1 = 83.53% despite comprising only 9.1% of the dataset. Conclusions: ROI-guided deep learning with metadata fusion achieves state-of-the-art bone lesion classification with clinically safe error patterns and probability outputs whose calibration was explicitly quantified, supporting its potential as a decision support tool in diagnostic radiology and forensic anthropology, pending external validation on independent cohorts. Full article