The trustworthiness of medical AI systems is often undermined by their potential to violate established clinical principles, a critical issue in self-supervised learning (SSL) frameworks where explicit domain knowledge is typically disregarded. To address this trust deficit, we propose the Logic-Constrained Joint Embedding Predictive Architecture (LC-JEPA). LC-JEPA is a novel framework that bridges neural and symbolic learning by seamlessly integrating predefined medical rules into the SSL process via differentiable logic programming. It utilizes a dedicated Differentiable Constraint Satisfaction Layer (D-CSL) to formally encode clinical constraints, enabling end-to-end joint optimization of both predictive accuracy and logical consistency. We introduce the Constraint Satisfaction Rate (CSR) as a specialized metric to quantify adherence to medical knowledge. Comprehensive evaluation on the MIT-BIH Arrhythmia Database demonstrates that LC-JEPA achieves high accuracy (97.5%) while significantly improving domain adherence (CSR of 93.9%), substantially outperforming unconstrained baselines (e.g., Vanilla JEPA’s 86.4% CSR). Extensive sensitivity analysis confirms that the logic constraint is essential: removing it causes accuracy to drop from 97.5% to 32.1%, while the specific constraint weight λ ∈ [0.1,2.0] has minimal impact on performance. By successfully uniting representation learning with symbolic reasoning, LC-JEPA provides a robust and generalizable solution for building safer, more reliable, and clinically deployable trustworthy AI systems for ECG arrhythmia analysis.

Accurate and interpretable blood glucose prediction is critical for optimizing insulin therapy and preventing adverse events in type 1 diabetes (T1D) management. However, existing deep learning models often lack transparency, limiting their clinical adoption. We introduce a novel neural-symbolic framework that integrates Fourier neural operators (FNOs) with clinical rules to predict 30-minute-ahead blood glucose levels with high accuracy and interpretability. FNOs capture periodic glucose patterns, such as circadian rhythms and meal responses, while differentiable clinical rules encode medical knowledge, providing actionable insights like hypoglycemia risk alerts. Evaluated on the OhioT1DM dataset, our model achieves a mean absolute error (MAE) of 10.2 mg/dL, meeting ISO 15197:2013 standards, and 99.8% Clarke Error Grid A+B coverage, ensuring clinical safety. Leave-one-subject-out cross-validation across 12 subjects demonstrates robust generalization (MAE = 8.85 ± 1.33 mg/dL), addressing concerns about sample size limitations. The model exhibits strong robustness to continuous glucose monitoring (CGM) measurement noise (maintaining ISO compliance at 10% noise) and missing data (5.4% performance degradation at 40% data loss). By generating interpretable outputs, such as trend analyses and risk warnings, our framework supports clinical decision-making, enhances patient trust, and facilitates personalized diabetes care. This work advances medical informatics by combining data-driven and knowledge-driven methods, offering a scalable solution for real-time glucose monitoring and insulin dosing in T1D.

Cardiovascular disease (CVD) is the leading cause of global health issues which requires swift and accurate risk assessment techniques. Present-day risk calculators together with machine learning (ML) models have performed inadequately when it comes to reliability and comprehension of results. We produced a complex deep learning (DL) framework which builds Extended Long Short-Term Memory networks using an exponential gating mechanism (xLSTMeg). The new architecture design controls feature variations while extracting long-term dependencies through increased performance. A Random Forest Regression (RFR) evaluated the most important features while focusing on the top influential biomarkers, which eliminates redundant features. Performance is evaluated using cross-validation protocol and scientifically validated using DL explainability paradigm. The exponential gating mechanism achieved superior performance in CVD risk stratification by at least 6% compared to conventional Long Short-Term Memory (cLSTM) and other ML models. Analysis showed that RFR allows the system to benefits most from the 64% most significant features. The system yielded better results by employing DL algorithms which SHapley Additive exPlanations (SHAP) analysis proved essential features for risk prediction. Our study concludes that our proposed xLSTMeg architecture with optimized features for CVD risk assessment achieves better robustness and enhanced accuracy.