Stress is a widespread concern that impacts human health with its silent progression, causing significant public health burdens and economic loss globally. Non-invasive wearable technology empowered by physiological signal monitoring can enable early warning systems for stress, alleviating some of the burdens, allowing on-time interventions, and thus significantly improving quality of life. This study used the heart rate and respiratory rate data from 34 participants. It evaluated the performance of the hybrid deep learning CNN-Transformer model and benchmarked it against deep learning convolutional neural networks (CNNs) and Transformer models, extreme gradient boosting (XGBoost) and random forest (RF) machine learning models, comprising a total of five AI models. To mitigate data imbalance and observe the efficacy of deep learning data augmentation techniques in physiological signals for stress monitoring, two generative adversarial network (GAN) models: conditional tabular GAN (CTGAN), copula GAN (CopGAN) and variational autoencoder (VAE) based model tabular VAE synthesiser (TVAES) had been employed. The modelling performance significantly improved when applying CTGAN and CopGAN, demonstrating the usefulness of synthetic data. The CNN-Transformer achieved an average accuracy of 77%, a precision of 87% and an AUC of 83%. The study applied leave-one- subject- out (LOSO CV) to prove the CNN-Transformer hybrid’s robustness for generalizability to perform well for unseen subjects. The study integrated explainable AI models, Shapley values (SHAP), and local interpretable model-agnostic explanations (LIME), as well as Monte Carlo Dropout for uncertainty quantification to bring confidence, trust and transparency to AI systems, taking a step closer to real-world deployment. Similar studies can also help in the detection of other disorders, such as anxiety and depression.

Vision Transformers (ViTs) advance medical image classification but cannot reveal the nonlin-ear transformations learned in their feedforward layers, and the post-hoc gradient-based saliency methods used to interpret them, principally Gradient-weighted Class Activation Mapping (Grad-CAM), were designed for convolutional feature maps and require adaptation for transformer backbones. We propose FlexiKAN-XAI, a multi-scale Kolmogorov-Arnold Network (KAN) Vision Transformer in which each feedforward edge learns a unique nonlinear transformation parameterised as a B-spline curve. The architecture tokenises Magnetic Resonance Imaging (MRI) at four resolutions (16×16, 28×28, 32×32, 56×56) and supports a two-level explainability framework: Level 1 is an intrinsic readout of the trained weights as B-spline activation curves, revealing what each neuron learned; Level 2 is an architecturally native, forward-pass-only spatial attribution via scale-decomposed attention rollout, revealing where the model attends. We evaluate on three brain MRI benchmarks: Kaggle binary (98.24% ± 0.89, 10-fold cross-validation (CV)), 4-class MRI (94.03%, single split), and Figshare 3-class (85.67%, single split). On Kaggle, 5×10-fold repeated CV (n = 50) shows no significant accuracy difference between KAN and parameter-matched Multilayer Perceptron (MLP) baselines (p = 0.666, Wilcoxon; rank-biserial r rb = +0.071). Native attention rollout achieves 81.0% Pointing Game localisation, exceeding Grad-CAM on ResNet-50 (71.0%) and the architecturally matched MLP-ViT (7.0%, under naive Grad-CAM-for-ViT; see Sections 6.1 and 8). Perturbation faithfulness (Area Under the Perturbation Curve 0.250 vs 0.303 for Grad-CAM, lower is better) confirms rollout identifies more decision-relevant pixels. To our knowledge, FlexiKAN-XAI is the first to combine multi-scale tokenisation, KAN-augmented transformer encoding, and quantitatively validated two-level explainability for brain tumour MRI.

Transvaginal ultrasound (TVUS) is widely used for diagnosing ovarian endometriosis but remains limited by significant operator dependency. This systematic review and meta-analysis evaluated the diagnostic accuracy of ultrasound-based artificial intelligence (AI) models for ovarian endometriosis and directly compared their performance with that of human readers. We conducted a comprehensive search of five databases (PubMed, Embase, Scopus, Web of Science, and Cochrane Library) up to 5 December 2025 to identify studies reporting diagnostic metrics for AI models, compared with human readers, for detecting ovarian endometriomas. Pooled sensitivity, specificity, and area under the curve (AUC) were calculated using a bivariate random-effects model. Seven studies with 2737 patients (6061 images) were included. AI models demonstrated a pooled sensitivity of 91

Brain tumor segmentation from magnetic resonance imaging (MRI) supports diagnosis and treatment planning, but models trained on benchmark datasets often fail when applied to data from different institutions. This study aims to improve segmentation accuracy and cross-dataset generalization by distributing state-space modeling across all network stages. We present the Progressive State-Space U-Net (PSS-UNet), which propagates learned state representations across all encoder and decoder stages rather than placing state-space components only at the network bottleneck. The design combines squeeze-and-excitation attention with selective state-space modeling at each stage, connected through cross-scale state bridges. We trained PSS-UNet on the BraTS 2021 glioma dataset (N=1,000) and evaluated generalization on both a held-out test set (N=251), PSS-UNet achieves a Dice coefficient of 0.935±0.067, outperforming baseline V-Net (0.926±0.080) with statistical significance (p<0.001, paired t-test). Ablation experiments show that progressive state propagation contributes +0.88% improvement over bottleneck-only designs. We evaluated cross-dataset generalization on BraTS-Africa(N=145) without fine-tuning. PSS-UNet maintains 0.896±0.124 Dice, a 3.9% drop from in-distribution performance. Counterintuitively, distributing state-space computation across all stages proves more efficient than concentrating it at the bottleneck: PSS-UNet runs 2.4× faster than bottleneck-only SSM designs (79.9ms vs. 194.8ms per volume) while adding only 7.3% parameters and 4.3% inference time over the convolutional baseline, enabling real-time throughput of 12.5 volumes per second on a single GPU. These results suggest that propagating state information across network stages, rather than isolating it at the bottleneck, improves both accuracy and generalization for brain tumor segmentation.

RATIONALE AND OBJECTIVES:To develop and externally validate ultrasound nomograms combining BI-RADS features and quantitative morphometric characteristics, and to compare their performance with expert radiologists and large language models in biopsy recommendation and malignancy prediction for breast lesions. METHODS:In this multi-center, multi-national study, 1747 women with breast lesions underwent ultrasound across three centers in Iran and Turkey. A total of 10 BIRADS and 26 morphological features were extracted from each lesion. Three nomograms based on BI-RADS, morphometric, and both feature sets were constructed. Three radiologists (one senior, two general) and two ChatGPTs including ChatGPT-o3 and o4-mini-high interpreted de-identified breast lesion images. Diagnostic performance for biopsy recommendation and malignancy prediction was assessed across all cohorts. RESULTS:According to the pooled results, although the difference between the fused nomogram and the BI-RADS version was not statistically significant, the fused version consistently outperformed all models in biopsy recommendation and malignancy prediction (AUCs of 0.901 and 0.853, respectively) compared to BI-RADS nomogram (AUCs of 0.898 and 0.834), morphometric nomogram (AUCs of 0.825 and 0.708), radiologist1 (AUCs of 0.820 and 0.729), radiologist2 (AUCs of 0.605 and 0.719), radiologist3 (AUCs of 0.728 and 0.699), ChatGPT-o3 (AUCs of 0.729 and 0.689), and o4-mini-high (AUCs of 0.713 and 0.695). CONCLUSIONS:The proposed BI-RADS-morphometric nomogram outperforms standalone nomogram models, LLMs, and radiologists in guiding biopsy decisions and predicting malignancy. The proposed novel fused nomogram has the potential to reduce unnecessary biopsies and enhance personalized decision-making in breast imaging.

Background and Objective: Speech-based Parkinson’s disease (PD) detection has been considered a potential tool that is non-invasive and affordable to aid PD diagnosis both clinically and remotely. However, centralized training requires sensitive voice samples to be collected at a single point of origin, thus causing considerable privacy and data sharing issues.Methods: This paper proposes a privacy-preserving federated learning framework using a cloud-assisted approach, which is referred to as Fed-CNN-BiLSTM for speech-based PD detection. This architecture employs one-dimensional CNN layers to discover local temporal relationships from speech samples with BiLSTM layers to recognize sequence dependencies. Speech samples were preprocessed and windowed to distribute among the federated clients. Ablation studies were carried out to identify the influence of several aspects of the framework, including preprocessing, temporal modeling, class weighting, client balancing, and privacy mechanisms. Moreover, patient-level explainability maps of the temporal development of PD probabilities during speech recording were created.Results: For the IPVS data set, the model had 96.32% hold-out accuracy and 97.83% LOSO cross-validation accuracy. On the NeuroVoz data set, it reached 90.10% accuracy in cross-validation using subject-level 10-fold cross-validation. From the privacy analysis perspective, clipping provided competitive results, while introducing noise had a negative impact on PD prediction. From the explainability perspective, Healthy Control subjects showed lower PD probability levels than PD subjects, where PD participants had higher PD probability levels.Conclusions: The outcomes suggest that the Fed-CNN-BiLSTM architecture offers a reliable, privacy-preserving, and interpretable model for PD identification using speech.

ABSTRACT Depression is a common and devastating mental health illness with serious personal and societal consequences. Despite advancing treatment techniques, there are still hurdles in the effective diagnosis and treatment of depression, such as prompt diagnosis, personalized medication, and continuous monitoring. In recent years, artificial intelligence (AI) has emerged as a potential tool in mental health treatment, providing novel solutions to these difficulties. This systematic study aims to comprehensively assess the existing AI systems for depression detection and treatment. The paper presents a systematic and comprehensive review of the last decade for depression detection, prediction, and treatment. One hundred eighty journal articles fulfilling preset inclusion criteria were found and analyzed using Preferred Reporting Items for Systematic Reviews and Meta‐Analyses from major academic databases. This review used a variety of detection modalities (physical, physiological, repetitive transcranial magnetic stimulation, and pharmacological treatment response) and AI approaches, including machine learning (ML) and deep learning (DL), to address various areas of depression care, including detection, diagnosis, prediction, and treatment. Key findings demonstrate that AI offers tremendous promise in boosting depression care across the continuum, from early identification to individualized therapy optimization and remote monitoring. ML and DL models demonstrate promising accuracy in predicting depression onset, severity, and treatment response based on diverse data sources, including electroencephalogram, electrocardiogram, photoplethysmography, electrodermal activity, electronic healthcare records, facial, speech, text, and pharmaceutical data. The paper highlights the important research challenges in current automated depression decision‐making models. Finally, we emphasize the prospects for developing effective and robust AI‐based depression models incorporating data and model fusion, the model's trust, portability, privacy preservation, and security features. This article is categorized under: Fundamental Concepts of Data and Knowledge > Explainable AI Technologies > Machine Learning Technologies > Artificial Intelligence

OBJECTIVES:The 2022 update of the Ovarian-Adnexal Reporting and Data System (O-RADS) improves risk stratification of adnexal lesions; however, radiologist interpretation remains subject to inter-observer variability and conservative diagnostic thresholds. Concurrently, deep learning (DL) models demonstrated promise in ovarian mass characterization. This study evaluates radiologist performance applying O-RADS version 2022 (v2022), compares it to convolutional neural network (CNN) and vision transformer (ViT) models, and investigates diagnostic gains from hybrid human-artificial intelligence (AI) frameworks with emphasis on explainable DL approaches that could enhance clinical applicability. METHODS:In this retrospective study, a total of 512 ultrasound images from 227 patients (110 with at least 1 malignant lesion) were analyzed. Sixteen DL models, including DenseNets, EfficientNets, ResNets, VGGs, Xception, and ViTs were trained and validated. For each model, a hybrid framework integrating radiologist-assigned O-RADS scores with DL-predicted malignancy probabilities was constructed. RESULTS:Radiologist-only O-RADS assessment achieved an area under the curve (AUC) of 0.683 and an accuracy of 68.0%. CNN models yielded AUCs of 0.620-0.908 and accuracies of 59.2-86.4%, while ViT16-384 reached the best performance, with an AUC of 0.941 and an accuracy of 87.4%. Hybrid human-AI frameworks significantly enhanced most CNNs (9 out of 12 CNNs, p < .05) and ViTs (3 out of 4 ViTs, p < .05). CONCLUSIONS:DL models outperform radiologist-only O-RADS v2022 assessment. The integration of expert radiologist scores with AI yields the highest accuracy, supporting hybrid human-AI paradigms as a promising approach to standardize ultrasound interpretation, reduce false-positive diagnoses, and improve identification of high-risk ovarian lesions.

This study introduces NeuroStrata, a connectivity-aware deep representation learning framework for EEG-based mental stress analysis using Time-Varying Partial Directed Coherence (TV-PDC). Unlike conventional EEG classification approaches based on static features, NeuroStrata models the temporal evolution of frequency-specific directed connectivity across distributed brain regions. EEG signals from the 32-channel SAM 40 dataset recorded during mental arithmetic tasks were used to generate TV-PDC connectivity maps. These maps were processed using pretrained Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to extract deep connectivity embeddings, which were subsequently classified using lightweight machine learning models. Experimental results demonstrate that beta-band connectivity provides the highest discriminative capability, achieving a peak accuracy of 97.3

Reliable automatic detection of drowsiness from electroencephalogram (EEG) recordings is critical for safety-sensitive applications, yet manual annotation remains slow and prone to inter-rater variability. We present a dual-branch late-fusion framework that integrates a channel-bagging neural network with an interpretable Random Forest (RF) classifier. This design enhances transparency by revealing which channels and features inform each prediction, thereby supporting clinical trust. EEG signals first undergo bandpass and notch filtering, common average referencing, and train-only z-score normalization. Then the neural branch learns per-channel representations and averages channel-wise probability estimates to mitigate the influence of noisy electrodes, while the RF branch operates on physiologically meaningful features, including Hjorth parameters and Higuchi fractal dimension (HFD). Finally the predictions are obtained by averaging the probabilities from both branches. Explainability is provided through temporal occlusion and attribution maps, as well as RF feature importances, highlighting when and where the model focuses during decision-making. Subject-wise evaluations show a wide but interpretable range of performance: accuracy spans 65.14–96.13%, specificity 37.84–95.86%, sensitivity 74.22–96.38%, and F1-score 72.60–96.13%, with higher performance on less challenging recordings and a clear sensitivity–specificity trade-off on difficult cases. When pooling five subjects, the system achieves 72.65% accuracy, 73.77% specificity, 71.53% sensitivity, and a 72.22% F1-score. Future work will focus on improving cross-subject generalization, enhancing robustness to artifacts and missing channels, and optimising the framework for low-latency deployment in noisy clinical and real-world environments.

BACKGROUND AND OBJECTIVE:Brain disorder diagnosis and prediction remain challenging because neuroimaging, electrophysiological, behavioral, and multimodal data are high-dimensional, noisy, heterogeneous, and limited by small clinical cohorts. This systematic review synthesised applications of quantum artificial intelligence (QAI) for brain disorder diagnosis, prediction, detection, and monitoring. METHODS:Following PRISMA guidelines, studies published from 2016 to 13 January 2026 were retrieved from Scopus, Web of Science, and IEEE Xplore. After screening, 36 studies met the eligibility criteria and were qualitatively analysed according to disorder category, data modality, QAI method, implementation setting, validation strategy, and performance. RESULTS:At the broader disease-group level, neurodegenerative disorders were the most frequently investigated, followed by mental health and psychiatric disorders. At the individual level, Parkinson's disease and schizophrenia were the leading applications, followed by depression, anxiety, Alzheimer's disease, and stress-related tasks. MRI-based modalities were the most frequently used data source, followed by multimodal data and EEG. Methodologically, primary QAI approaches were dominated by quantum neural and QDL architectures, followed by quantum-inspired optimization or feature-selection methods and quantum-kernel/conventional QML classifiers. Qiskit/IBM Quantum and PennyLane were the most frequently reported quantum software frameworks. However, most studies relied on simulators, classical quantum-inspired implementations, or unclear implementation settings, with limited real-hardware evaluation. CONCLUSIONS:QAI shows emerging potential for brain disorder analysis, particularly through hybrid quantum-classical learning, quantum neural architectures, quantum-kernel methods, and quantum-inspired optimization. Nevertheless, current evidence remains preliminary and requires larger datasets, subject-level and external validation, fair classical benchmarking, noise-resilient circuits, real quantum hardware evaluation, explainability, and clinical validation.

Diabetic retinopathy (DR) and diabetic macular oedema (DME) are two of the most significant preventable contributors to blindness in the adult population worldwide, yet current automated screening systems typically address each condition in isolation and rely on a single imaging modality. In this study, we propose a deep learning model that simultaneously grades DR severity and detects DME by fusing paired colour fundus and optical coherence tomography (OCT) images acquired from the same eye during the same clinical visit. Our architecture employs two parallel EfficientNet-B0 backbones pre-trained on ImageNet, one for each modality, whose 1280-dimensional feature vectors are concatenated into a 2560-dimensional joint representation. This fused representation passes through a shared fully connected block before branching into a three-class DR classification head and a binary DME detection head. We train and evaluate the model on a private dataset of 425 paired fundus and OCT eye images (850 images). The proposed architecture adopts feature-level fusion, in which modality-specific deep features are independently extracted from fundus and OCT images using separate convolutional backbones and subsequently concatenated to form a joint representation for multi-task learning. On the held-out test set (n= 85), the fusion model achieves 82.4% DR accuracy (area under the receiver operating characteristic curve [AUC] = 0.929, macro sensitivity = 0.81, macro specificity = 0.905) and 97.6% DME accuracy (AUC = 0.999, sensitivity = 0.833, specificity = 1.000). The fusion model detects 10 of 12 DME-positive eyes compared with only 7 of 12 for either the fundus-only or OCT-only baselines, representing a 43% relative improvement in DME sensitivity. Stratified five-fold cross-validation (n = 425 aggregated predictions) corroborates these findings, with the fusion model reaching 87.1% DR accuracy (AUC = 0.978) and 99.1% DME accuracy (AUC = 1.000). Gradient-weighted class activation mapping visualisations confirm that the fundus branch attends to clinically relevant macular lesions, whereas the OCT branch highlights retinal layer disruptions and subretinal fluid, providing interpretability. To the best of our knowledge, the proposed MultiRetNet is the first lightweight, task-specific multimodal architecture to jointly grade DR severity and detect DME from paired same-eye, same-visit fundus and OCT images through explicit feature-level fusion within a single end-to-end multi-task framework, distinct from recent generalist ophthalmic foundation models, supporting the value of multimodal fusion for comprehensive diabetic eye screening pending external validation.

Psychological stress has a significant impact on human well-being, behaviour and health that incurs significant economic costs. This study is designed to develop predictive models that can potentially assist in the diagnosis and management of psychological stress through artificial intelligence (AI). We propose an explainable-AI deep learning Transformer-based Attentive Interpretable Tabular Learning (TabNet) model that captures multimodal imbalanced data comprised of the derivative of heart rates (HR) and electrodermal activities (EDA) of 99 subjects from four publicly available datasets (i.e., WESAD, SWELL, NEURO and UBFC-Phys). The proposed model creates the generalizability of the representative dataset, with transfer learning used as a key to identifying the stress with HR and EDA variables. The proposed stress classification model performed relatively well in detecting the stress with 76% accuracy, 77% precision, 76% recall and a F1 score of 74% and the leave one subject out (LOSO) cross-validation strategy. The proposed model also considered the variability of subject-wise stress with the capability to automate the feature selection process using the attention mechanisms. Finally, the model has been tested using a synthetic dataset with a Synthetic Minority Over-sampling Technique (SMOTE) for the imbalanced dataset and the study has used SHapley Additive exPlanations (SHAP) and local interpretable model-agnostic explanation (LIME) for model explainability that can bring greater trustworthiness of the proposed model for its future exploration in clinical healthcare environments.

This study introduces NeuroStrata, a connectivity-driven deep representation learning frameworkfor electroencephalogram (EEG)-based mental stress analysis using Time-Varying PartialDirected Coherence (TV-PDC). Unlike conventional EEG classification pipelines that rely onstatic features, proposed framework models temporal evolution of frequency-specificdirected connectivity across distributed brain regions. EEG signals from the 32-channel SAM40 datasets, recorded during mental arithmetic tasks, were used to produce frequency-specificTV-PDC connectivity maps. These maps were processed using pretrained Convolutional NeuralNetworks (CNNs) and Vision Transformers (ViTs) to extract deep connectivity embeddings,which were subsequently classified using lightweight machine learning (ML) models. Experimentalresults show that beta-connectivity exhibits highest discriminative capability,achieving a peak accuracy of 97.3% using the LAION-CLIP-ViT-L14 backbone with a SupportVector Machine (SVM) classifier, while alpha-connectivity provides consistently stableperformance across model configurations. Connectivity feature importance analysis revealsprominent frontal-driven alpha influences and centrally integrated beta connectivity patterns,reflecting the engagement of regulatory and sensorimotor networks during stress. Temporalevaluation further indicates that classification performance stabilizes in mid-to-late temporalwindows, suggesting the progressive consolidation of stress-related connectivity signatures.From an artificial intelligence perspective, the framework implements a connectivity-aware deeprepresentation learning strategy that integrates time-varying effective connectivity modellingwith deep embedding extraction using pretrained CNN and ViT architectures to encode dynamicdirected brain network structure. From an engineering application perspective, it enablesstructured evaluation of stress-modulated connectivity dynamics through interpretable directednetwork analysis, temporal window stability characterization, and a unified multi-stage EEGconnectivity processing pipeline for automated inference of task-evoked cognitive stress states.

Despite its transformative potential in healthcare, the adoption of artificial intelligence (AI) in clinical practice remains constrained by a persistent trust deficit among clinicians and patients. To address this, we conducted a systematic comparative review of 112 peer-reviewed studies published between 2015 and 2025, following the PRISMA guidelines for study selection. Articles were sourced from major scientific databases, focusing on methodological innovations and clinical evaluations to enhance AI trustworthiness. Using a novel Composite Human-Centered Trustworthiness Score (HCTS), we systematically evaluated and compared the contributions of relevant studies. Our analysis identified four human-centered pathways: explainable AI (XAI), comprising intrinsic interpretable models and post-hoc techniques (e.g., SHAP, LIME) to support error analysis and stakeholder communication; human-in-the-loop (HITL) frameworks that leverage clinician expertise via active learning and interactive visualization to improve model reliability and usability; hybrid neuro-symbolic architectures that integrate symbolic reasoning with deep learning to achieve robustness in complex or data-sparse settings; and uncertainty quantification (UQ) methods (e.g., Bayesian inference, Monte Carlo dropout, and ensemble techniques) that provide confidence estimates that are critical for high-stakes clinical decisions. We found that integrated strategies, including XAI-driven HITL loops and XAI + UQ frameworks, yield the greatest gains in transparency, human oversight, and computational capability. Addressing technical challenges (data heterogeneity, system interoperability), ethical and regulatory imperatives (fairness, accountability), and advancing multimodal and continual-learning paradigms are essential for ensuring the safe, transparent, and sustainable deployment of AI in clinical practice.

This study addresses the challenge of automatic sleep stage classification using a single-channel approach. It introduces an effective method that leverages the photoplethysmography (PPG) signal for sleep stage clas- sification.We propose a deep learning framework, Pleth-SleepNet, based on a custom architecture that combines a parallel convolutional neural network and long short-term memory (CNN+LSTM). The model was trained and evaluated using PPG signals from two datasets: the DREAMT (Dataset for Real-Time Sleep Stage Estimation Using Multisensor Wearable Technol- ogy) and the cyclic alternating pattern (CAP) dataset, focusing on patients with obstructive sleep apnea (OSA), excessive daytime sleepiness (EDS), snoring, insomnia, and related conditions. Pleth-SleepNet achieves remarkable performance with an accuracy of 98.81% for OSA, 99.79% for EDS, and 99.65% for snoring using the DREAMT dataset, and 99.92% for insomnia, 100% for narcolepsy, 99.14% for rapid eye movement (REM) behavior disorder (RBD) and 99.77% for periodic leg movement (PLM) for CAP dataset. The findings confirm the effectiveness of using a single PPG channel for robust and accurate sleep staging, particularly in patients with sleep disorders. The Pleth-SleepNet model not only provides a near-perfect classification approach but also offers a cost-effective and accessible solution, highlighting its strong potential for clinical and wearable healthcare applications.

ABSTRACT Gastrointestinal (GI) diseases remain among the leading causes of global mortality, with early detection directly linked to survival outcomes. While previous reviews have focused on single imaging modalities, this systematic review uniquely examines artificial intelligence applications across endoscopic, radiological, and histological approaches, reflecting actual clinical diagnostic pathways. This systematic review analyzes 76 high‐quality studies (2016–2024) and provides the first comprehensive assessment of how AI performs across different imaging techniques for GI abnormality detection. This multi‐modal perspective is particularly timely as healthcare systems move toward integrated diagnostic workflows. Our analysis reveals endoscopy as the most widely used modality ( n = 44), particularly for Helicobacter pylori , colorectal polyps, and ulcerative colitis detection. Histological analysis emerges as the second most common approach ( n = 25), especially for celiac disease and ulcerative colitis, while CT imaging ( n = 10) primarily supports colorectal polyp detection. Deep learning methods significantly outnumber traditional machine learning techniques (68 vs. 8 studies), consistently achieving 90%–99% diagnostic accuracy across multiple disease categories. However, these systems face significant implementation barriers to clinical adoption. Most validation is still conducted in controlled, single‐center settings using curated datasets that poorly reflect clinical complexity. Future studies must prioritize multicenter validation, standardized imaging protocols and preprocessing pipelines, and the integration of interpretable AI models capable of providing transparent diagnostic rationale. This review maps the current technical landscape while highlighting critical translational challenges that must be addressed to enable real‐world impact. This article is categorized under: Technologies > Data Preprocessing Technologies > Artificial Intelligence

Accurate skin cancer diagnosis is vital for early treatment and improved patient outcomes. Deep learning models have shown promise in automating skin cancer classification, yet challenges remain due to data scarcity and limited uncertainty awareness. This study presents a comprehensive evaluation of deep learning-based skin lesion classification with transfer learning and UQ on the HAM10000 dataset. We benchmark several pre-trained feature extractors (including Contrastive Language-Image Pre-training (CLIP) variants, ResNet50, DenseNet121, VGG16, EfficientNet-V2-Large, and ConvNeXt Large) combined with traditional classifiers such as SVM, XGBoost, and logistic regression. Multiple PCA settings (64, 128, 256, 512) are explored, with LAION CLIP ViT-H/14 and ViT-L/14 at PCA-256 achieving the strongest baseline results. In the UQ phase, Monte Carlo Dropout (MCD), Ensemble, and Ensemble Monte Carlo Dropout (EMCD) are applied and evaluated using uncertainty-aware metrics (UAcc, USen, USpe, UPre). Ensemble methods with PCA-256 provide the best balance between accuracy and reliability. Further improvements are obtained through feature fusion of top-performing extractors at PCA-256. Finally, we propose a feature-fusion-based model trained with a Predictive Entropy (PE) loss function, which outperforms all prior configurations across both standard and uncertainty-aware evaluations, advancing trustworthy deep learning-based skin cancer diagnosis.

Phonocardiogram (PCG) analysis is an inexpensive and non-invasive technique for the automatic diagnosis of heart valve diseases. In the clinical domain, PCG recordings are typically corrupted by noise, inter-subject variability, and overlapping signal characteristics, necessitating uncertainty-based decision support. To solve this problem, this paper presents an uncertainty-aware deep multimodal early fusion network (DMEFNet) that jointly integrates one-dimensional(1D) temporal signals and two-dimensional (2D) time-frequency image representations for multiclass PCG classification. Four main uncertainty quantification (UQ) methods, namely Monte Carlo (MC) dropout, Bayesian Neural Networks (BNNs), Deep Ensembles (DE), and Dirichlet-based Evidential Deep Learning (EDL), are used for predictive uncertainty estimation. Extensive experimental evaluations on the public HVD dataset demonstrated that uncertainty estimates are well-calibrated, scoring low predictive uncertainty when samples are correctly classified and higher uncertainty for ambiguous or noise samples. To the best of our knowledge, this work is among the first few attempts to analyze UQ in PCG-based heart valve disease classification. The introduced framework increases clinical trust, reliability and enables risk-aware decisions, thereby promoting the development of PCG-based diagnostic systems for real-world clinical applications.

Background Fuzzy logic techniques have gained significant prominence in healthcare, primarily due to their ability to address and manage the inherent imprecision and uncertainty in healthcare data analysis. We conducted a comprehensive review investigating how fuzzy techniques have developed and been applied in healthcare between 2017 and 2025. Methods We conducted a systematic literature review following PRISMA guidelines, analyzing 91 papers from major medical and engineering databases. Our analysis focused on three distinct methodological streams: classical fuzzy systems, combined fuzzy-machine learning approaches, and emerging fuzzy-enhanced deep learning frameworks. We evaluated each paper’s methodology, implementation details, and clinical relevance. Results The distribution of research approaches showed a balanced landscape across methodologies, with traditional fuzzy systems comprising 30.1%, hybrid approaches 34.4%, and fuzzy-deep learning implementations 33.3% of studies. Medical imaging dominated the application domains, led by MRI studies (36.3%) and CT applications (12.1%). Biosignal analysis also showed strong representation, particularly in EEG (22%) and ECG (7.7%) applications. Performance analysis revealed that both deep learning and conventional feature engineering methods achieved comparable accuracy rates of approximately 96.5%, with some variations in consistency across different applications. Conclusions This research area has undergone significant evolution, particularly since 2023, with an increased emphasis on incorporating fuzzy techniques into deep learning frameworks. This transition shows that fuzzy approaches, originally designed as standalone solutions, are now becoming critical components of modern healthcare AI systems, providing unique benefits in dealing with medical data uncertainty.