Lightweight ECG Signal Classification via Linear Law-Based Feature Extraction

Abstract

This paper introduces LLT-ECG, a novel semi-supervised method for electrocardiogram (ECG) signal classification that leverages principles from theoretical physics to generate features without relying on backpropagation or hyperparameter tuning. The method identifies linear laws that capture shared patterns within a reference class, enabling compact and verifiable representations of time series data. We evaluate the method on two PhysioNet datasets, TwoLeadECG and VPNet. On TwoLeadECG, a minimal configuration - using only the linear law-based transformation (LLT) and a linear decision rule - reaches 73.1% accuracy using just two features. On VPNet, LLT-ECG combined with classifiers like k-nearest neighbors (KNN) and support vector machines (SVM) achieves up to 96.4% accuracy, comparable to deep learning models. These results highlight LLT-ECG’s promise for lightweight, interpretable, and high-performing ECG classification.