Unsupervised phenotype discovery in 12-lead ECGs using deep multimodal clustering techniques

Document Type

Article

Publication Date

8-1-2026

Abstract

Cardiovascular diseases represent a significant worldwide health issue necessitating scalable and interpretable diagnostic approaches. The 12-lead electrocardiogram (ECG) serves as a fundamental tool for detecting cardiac issues, commonly employed in supervised learning; nevertheless, supervised learning algorithms are constrained by costly annotations and face challenges in identifying novel abnormalities. We define ECG phenotypes as observable patterns in waveform morphology, rhythm, and intervals reflecting underlying cardiac states. This study presents an unsupervised deep clustering approach for 12-lead ECG analysis, aimed at identifying clinically significant cardiac abnormalities independent of labeled data. The approach extracts multimodal characteristics, such as latent representations from autoencoders, statistical descriptors, and medical-specific properties, and combines them using an attention-based technique to get complementary signal embeddings. A deep clustering module uses a multi-objective loss to find the best balance between keeping things together inside a cluster and separating them from other clusters. Theoretical contributions include a novel anti-collapse loss and feature-level attention fusion for ECG. The proposed framework shows strong clustering performance on the PTB-XL dataset, with Silhouette scores over 70%, Adjusted Rand Index (ARI) over 60%, and cluster purities over 87%. Ablation studies validate the importance of attention fusion and weighted loss terms, while validation trials exhibit reproducibility with a coefficient of variation (CV) below 3%, resilience to noise (ARI > 85%), and statistical significance (p< 0.001). Cross-dataset validation on Chapman-Shaoxing confirms generalizability. The model recognizes established diagnostic categories and mixed aberrant patterns, demonstrating its capability for scalable, interpretable, personalized ECG-based diagnostics.

Keywords

12-lead ECG, Attention mechanism, Cardiovascular diseases (CVD), Deep clustering, Feature fusion, Multimodal AI

Publication Title

Biomedical Signal Processing and Control

ISSN

1746-8094

DOI

10.1016/j.bspc.2026.110336

Volume

121

Share

COinS