A deep learning framework for arrhythmia classification: mitigating data leakage and class imbalance in ECG analysis

Accurate and timely identification of cardiac abnormalities remains essential for reducing mortality associated with life-threatening cardiac arrhythmias. However, automated diagnostic systems often struggle with imbalanced clinical data and limited generalizability. In this work, we propose an efficient deep learning framework to classify the various ECG arrhythmias based on time-frequency features generated through Continuous Wavelet Transform (CWT) and efficient VGG16 transfer learning. By avoiding data leakage through a stratified split-then-augment partitioning method, the method achieved a test accuracy of 96.05% and stable performance across five-fold cross-validation (96.0% ± 0.4%). The model achieved a macro-average F1 score of 0.933 despite class imbalance with a sensitivity that supports real-world detection needs, including ventricular fusion beat recognition. CWT improved the multi-resolution representation of non-stationary ECG signals, while architectural changes, including batch normalization, global average pooling and dropout, improved model performance without increasing computational load. The system has only 166K trainable parameters with an inference latency of ∼25-30 ms, suggesting the possible application in emergency triage, bedside monitoring, and telemedicine-assisted decision support. These findings show the feasibility of the proposed method for early cardiac event detection and providing clinical patient assistance in high-risk environments.