Time–Frequency Deep Learning for High-Accuracy Arrhythmia Classification: A Wavelet-Scalogram Squeeze-and-Excitation Residual Network

Abstract

Continuous cardiac monitoring on wearable and embedded devices demands heartbeat classifiers that are simultaneously accurate and lightweight. We present WSRNet, a wavelet-scalogram squeeze-and-excitation residual network that classifies single-lead electrocardiogram (ECG) beats from a two-dimensional time–frequency image. Each beat is transformed by the continuous wavelet transform, using a Morlet mother wavelet, into a scalogram that makes cardiac morphology and spectral content jointly explicit; a compact two-dimensional residual convolutional network with squeeze-and-excitation channel attention then performs the classification. Evaluated under the intra-patient protocol on the MIT-BIH Arrhythmia Database over the four principal AAMI classes {N, S, V, F}, WSRNet attains 99.13 % overall accuracy and a macro-F1 of 0.932, detecting the rare supraventricular and fusion classes at F1 scores of 0.926 and 0.829 respectively despite a fifty-to-one class imbalance. An ablation study shows the model is robust to component choices, every variant exceeds 98 %, and that, in this intra-patient regime, a plain cross-entropy loss suffices (class re-weighting, essential inter-patient, is here counterproductive) and a size-matched raw-1-D network is competitive, so the scalogram is an effective though not exclusive route to high accuracy. The model uses only 716 K parameters (about 700 KB after 8-bit quantisation) and classifies a beat in a few milliseconds on a single CPU thread, comfortably within real-time budgets. It thus occupies the same lightweight regime as recent edge-oriented systems while matching the accuracy of far larger transformer- and wavelet-based models, making it a practical candidate for on-device arrhythmia screening.