Personalized ECG Arrhythmia Classification Using Single Normal Beat Guidance: A Minimal Supervision Approach for Atypical Morphology Cases

Electrocardiogram (ECG) exhibits diverse morphologies, allowing for individual distinction but often leading to varying interpretations of similar patterns. This complexity, especially with atypical morphologies, poses challenges for deep learning (DL) algorithms, potentially causing hidden stratification issues. While patient-specific studies show promise, their performance on unseen data remains uncertain, requiring labor-intensive corrections.The proposed model uses the morphology of the target beat and its intervals with surrounding beats. A single normal beat selected by an expert serves as a guide to influence the classification of all beats. Morphological similarity and RR intervals are utilized to classify beats into normal (N), supraventricular ectopic (S), and ventricular ectopic (V) types.ECGs were collected from 1,366 patients using patch-type devices, categorized into typical (91%; QRS duration (QRSd) \((<)\) 120 ms for N, S, and \((\geq)\) 120 ms for V) and atypical (9%; QRSd \((\geq)\) 120 ms for N, S, or \((<)\) 120 ms for V) groups. Inter-patient 5-fold cross-validation showed improved F1-score for the unseen atypical group, averaging 87.49 (\((+)\)7.66) with a standard deviation of 14.57 (\((-)\)5.91).The proposed method enhances ECG analysis for atypical cases by leveraging minimal patient-specific guide data. It demonstrates potential for improving productivity in clinical ECG analysis.