Abstract
Detection of cardiac abnormalities in 12-lead electrocardiograms (ECGs) requires models that are both accurate and interpretable. In this work, we investigate the diagnostic value of handcrafted inter-lead correlation features computed with Pearson, Spearman, and Kendall coefficients.Using the PTB-XL dataset, we evaluate statistical methods and classical machine learning methods on both binary (normal vs. abnormal) and multiclass (five diagnostic categories) tasks.Results show that simple statistical descriptors achieve strong performance in the binary setting and moderate performance in the multiclass setting.To interpret model decisions, we apply SHAP and permutation feature importance, which consistently highlight correlations between anatomically related leads as key predictors.These findings confirm that inter-lead dependencies carry physiologically meaningful information and can be effectively captured with simple, transparent models.While limited in their ability to represent nonlinear or patient-specific dynamics, correlation-based features provide a strong, interpretable baseline and motivate further exploration of adaptive, data-driven approaches.
DOI: 10.61416/ceai.v27i4.9849