Unsupervised classification of ECG signals using DFA and wavelet entropy

This study proposes a novel approach for the unsupervised classification of electrocardiogram (ECG) signals by leveraging two nonlinear features: detrended fluctuation analysis (DFA) and wavelet entropy (WE). Unlike supervised techniques that require extensive labeled datasets, this method applies K-Means clustering to group ECG recordings into normal and arrhythmic categories without prior annotations. DFA captures long-range correlations and fractal dynamics of the heart signal, while WE quantifies the signal’s frequency-domain complexity. The combination of these two physiologically meaningful features offers an interpretable and compact representation of ECG dynamics. ECG signals were sourced from two benchmark PhysioNet databases NSRDB (normal sinus rhythm) and MIT-BIH (arrhythmias) and preprocessed to extract DFA exponents and WE values. The resulting feature vectors were then clustered using K-Means, assuming a natural grouping into two physiological states. Results demonstrate that the algorithm effectively separates the signals into distinct clusters that align with known clinical classifications. The proposed method provides a computationally efficient and interpretable framework suitable for real-time ECG monitoring in wearable devices or low-resource healthcare settings. Furthermore, it aligns with the growing demand for explainable AI in medical diagnostics and can serve as a foundation for scalable and annotation-free arrhythmia screening tools.