Recent advances in deep learning based arrhythmia classification using ECG and PPG signals: a systematic review

This systematic review analyzes recent advances in deep learning–based arrhythmia classification using electrocardiography (ECG) and photoplethysmography (PPG) signals from 2023 to 2025. The review examines model architectures, preprocessing pipelines, datasets, optimization strategies, and evaluation metrics reported in recent studies. Particular emphasis is placed on hybrid models, transformer-based approaches, and lightweight architectures designed for wearable monitoring systems. Despite reported accuracies exceeding 97% on benchmark datasets such as MIT-BIH, several challenges remain. These include dataset imbalance, poor cross-dataset generalization, motion artifacts in PPG signals, limited large-scale datasets, and lack of standardized evaluation frameworks. The review also highlights emerging research directions including self-supervised learning, multimodal fusion, domain adaptation, and uncertainty-aware models. The findings emphasize the need for standardized benchmarking, physiologically grounded synthetic data generation for rare arrhythmias, and efficient interpretable models capable of deployment in real-world wearable monitoring systems.