Interpretable machine learning for identifying individual-specific cardiogram signatures

Abstract

This study investigates cardiogram-based biometric identification as a physiological authentication problem, focusing on the discriminative capacity of electrocardiogram (ECG) and impedance cardiogram (ICG) features and their robustness to emotional variability. A total of 29 features spanning four domains (temporal, amplitude, slope, and morphological) are evaluated using random forest (RF) models combined with multiple interpretability methods. Features derived from ECG and ICG rank consistently among the top 10 most important by Gini importance, permutation importance, and Shapley additive explanations (SHAP) values, with ECG-derived QRS descriptors occupying the highest positions. In parallel, BCX features derived from ICG provide complementary information, although their stability across methods is lower. Correlation analysis reveals substantial multicollinearity, where the RF distributes and diminishes importance across highly correlated pairs, confirming reduced independent contributions. Statistical analysis identifies 14 features that differ significantly between baseline and anger, without a clear pattern by domain. Feature selection with recursive feature elimination and genetic algorithms converges on a subset (12 features) that attains accuracy within 1% of the full set (99%), improving efficiency in storage and computation. Complementary analyses indicate that individual differences are primarily encoded in ECG features describing QRS amplitude, slope, morphology, and duration, with the first three remaining stable between baseline and anger, and duration being the only selected QRS feature that changes significantly. BCX amplitude features provide supportive but less stable discriminatory cues. More broadly, the proposed framework may help distinguish stable person-related characteristics, such as identity and personality-related traits, from transient affective changes.