Biomedical Signal Processing and Control Обработка сигналов

Which representation clusters? A chance-corrected benchmark of ECG representations for unsupervised arrhythmia structure

Dongho Chun, Axel Faes
Elsevier · · DOI 10.2139/ssrn.7094872

Аннотация

Unsupervised analysis of the electrocardiogram (ECG) promises label-freediscovery of cardiac structure, yet whether clustering succeeds depends less onthe algorithm than on how the signal is represented. We benchmark fourrepresentation families for unsupervised separation of sinus rhythm (SR) fromatrial fibrillation (AF) on PTB-XL: whole-recording raw flattening, a convolutionalreconstruction autoencoder, compact hand-crafted rhythm and morphology features,and two frozen foundation-model embeddings (HuBERT-ECG and ECGFounder).Evaluation is chance-corrected (adjusted Rand index, ARI), patient-grouped,repeated over ten seeds, and bounded by a supervised ceiling. ECGFounder clustersbest (ARI up to 0.885), but 16 interpretable hand-crafted features are competitive(0.711) and far ahead of the raw and autoencoder pipelines (ARI $\approx 0$), whichmatch a majority-class assignment despite supervised ceilings well above chance.The same ordering holds on a MIT-BIH beat-morphology task. Two cautions emerge: thesilhouette score rewards the degenerate one-cluster solution (silhouette 0.81 atARI 0) and cannot validate ECG clustering alone, and the feature scaler materiallychanges the outcome. The ranking is largely expected, since SR and AF differ bydefinition in rhythm regularity; the contribution is the chance-corrected controlsthat locate where naive pipelines fail and the finding that the cheap interpretablerepresentation is the more deployable one. Under heavy class imbalance andfederation across ten heterogeneous sites, the hand-crafted features are the mostrobust, overtaking the foundation embedding under imbalance while the embedding'scentralized advantage erodes under federation. The representation, not thealgorithm, governs label-free ECG structure.

Читать у издателя →