TrustECG-DG: An Explainable, Uncertainty-Calibrated and Domain-Generalized Deep Learning Framework for Reliable Cardiac Abnormality Detection across Heterogeneous ECG Datasets

While deep learning has made significant advances in cardiac abnormality detection using electrocardiograms (ECG), current achievements are based on landmark classification accuracy, which may not necessarily translate to reliable performance on heterogeneous datasets, calibrated confidence or clinically interpretable predictions. Current ECG learning methods are often domain-dependent, and they treat explainability and uncertainty estimation as separate components. To address these drawbacks, this study proposes TrustECG-DG, an explainable, uncertainty-calibrated, and domain-generalised deep learning framework for reliable cardiac abnormality detection from multi-lead ECG signals. It combines a multi-scale convolutional neural network to learn morphological features with a temporal Transformer to model long-range rhythm dependencies. Cross-attention between morphological and temporal representations is used bidirectionally. The framework jointly learns domain-invariant features, class-aware contrastive alignment, and covariance-based feature alignment to improve domain robustness. A deep ensemble of 5 models is then used, along with temperature scaling, to quantify prediction uncertainty and improve probability calibration. To facilitate interpretation, Integrated Gradients and Transformer attention rollout methods explain the ECG regions the model relies on for its predictions, and an attentive prediction mechanism is designed to pass uncertain cases to an expert for evaluation. Experimental results show that the proposed framework achieves 98.7% accuracy on the test set. Beyond classification performance, the framework is assessed using class-wise predictive metrics, cross-dataset generalisation, ablation analysis, calibration, uncertainty estimation, and explanation analysis. The resulting framework offers a unified solution that integrates accurate ECG classification, robustness, calibrated confidence, interpretability, and selective decision-making, providing a more reliable basis for intelligently detecting cardiac abnormalities using an ECG.