Arrhythmia is a major cardiac abnormality associated with cardiovascular disease (CVD) and contributes substantially to morbidity and mortality worldwide. Automated analysis of electrocardiogram (ECG) signals is therefore essential for supporting the early detection of cardiac arrhythmia and clinical decision-making. This work demonstrates a comparative analysis of hybrid deep learning architectures, including CNN, CNN+LSTM, CNN+GRU, and CNN+Transformer for automated multi-class ECG arrhythmia classification using single-lead ECG signals. In the proposed architectures, one-dimensional convolutional neural networks (1D-CNNs) are employed to extract local morphological features, while LSTM, GRU, and transformer encoder networks are used to model temporal dependencies in ECG signals. Furthermore, gradient-weighted class activation mapping (Grad-CAM) is incorporated to enhance model interpretability by highlighting the important ECG regions that contribute most to the classification decisions. The developed models are evaluated on two publicly available benchmark datasets: CPSC2018, comprising nine arrhythmia classes, and the Georgia dataset, containing 67 SNOMED-CT diagnostic categories, of which 50 classes with sufficient heartbeat samples are considered for evaluation. On the CPSC2018 dataset, the CNN+GRU model achieved the best overall performance, with test accuracy of 90.31% and macro-averaged accuracy, AUC, precision, sensitivity, and F1-score of 97.85%, 98.89%, 90.26%, 89.92%, and 90.07%, respectively. On the Georgia dataset, the CNN+LSTM model achieved the best overall performance, with test accuracy of 85.27% and macro-averaged accuracy, AUC, sensitivity, and F1-score of 99.41%, 99.11%, 87.15%, and 86.98%, respectively. In addition, comprehensive ablation studies, computational complexity analysis, and comparisons with recent state-of-the-art methods demonstrate the effectiveness of the developed hybrid architectures. These results indicate that combining convolutional feature extraction with recurrent or transformer-based temporal modeling enables reliable multi-class ECG arrhythmia classification across diverse clinical datasets.
Adaptive recommender systems are updated as new user interactions occur, potentially changing both recommendations and their associated explanations. However, most explainable artificial intelligence methods evaluate explanations at a single point in time and do not measure how they evolve during continual updates. This study investigates explanation drift in dynamic graph-based recommender systems. We propose a temporal framework that integrates graph representation learning, streaming embedding updates, Shapley-based feature attribution, and drift-aware optimization. The framework employs three measures: Temporal Attribution Divergence (TAD), Explanation Stability Score (ESS), and Explanation Half-Life (EHL). We evaluated the proposed approach on MovieLens-1M and an Amazon Product Reviews subset under continual update settings. The framework achieved an F1@10 of 0.837 on MovieLens-1M, compared with 0.821 for the strongest reported baseline, and 0.810 on Amazon, compared with 0.789. Furthermore, it reduced TAD by approximately 42.2% relative to Static SHAP and increased ESS from 0.612 to 0.831. These results indicate that explanation drift can be reduced while maintaining recommendation performance under the evaluated conditions. However, the framework introduces additional computational overhead for embedding updates and explanation generation, which may affect response latency and its suitability for real-time deployment. The evaluation is also limited to offline benchmark datasets and feature-attribution stability; it does not assess users’ understanding or perceptions of the explanations. These findings support temporal explanation stability as a complementary evaluation criterion for adaptive recommender systems, while highlighting the need to consider computational efficiency in practical deployments.
