A minimalist 1-lead ECG beat classification without handcrafted morphological feature extraction for inter-patient, edge-oriented diagnosis

The conventional machine learning models are relying on handcrafted feature extraction. These methods require domain expertise and increase computational cost, which limits their implementation on resource-constrained wearable edge device settings. To investigate whether ECG beat classification can be done using machine learning approach on raw ECG signal without handcrafted morphological or spectral feature extraction while maintaining clinical-level accuracy in a patient-independent evaluation setting. The experiments have been conducted on MIT-BIH Arrhythmia dataset using Lead-II ECG signals, which have been split into five AAMI beat classes and evaluated according to inter-patient (DS1/DS2) protocol when training and testing beats belong to different patients. Three heart-beat segment variants are cross-validated on DS1 to select the most suitable one, which is progressively compressed to analyse the effect of compression ratio on classification accuracy. The Support Vector Machine (SVM) classifier with radial basis kernel was trained on the compressed raw ECG beat segment augmented with strictly causal RR interval values calculated from R-peaks detection without additional computation. With the operating point of 64 samples/beat, the proposed 1-lead approach reaches overall accuracy of 90.8% with per class F1-scores of 0.95 for normal beats and 0.73 for ventricular ectopic beats at the model size of 680 KB and inference latency of 0.297 ms per beat on a general-purpose CPU. Supraventricular ectopic beats remain the most challenging category, which is consistent with the difficulty reported in other inter-patient studies. These findings demonstrate that a small number of raw ECG samples in the vicinity of R-peaks along with some beat-timing information is sufficient for reliable ventricular ectopic beats detection in an inter-patient setting. Therefore, the proposed method can be used for low-latency and resource constrained wearable ECG monitoring.