SHAP-Driven Explainable EDA Biometrics: Feature Importance, Selection, and Recognition

Physiological signals have attracted growing interest in biometric recognition because they contain individual-specific characteristics that can support person identification. However, Electrodermal Activity (EDA) remains less investigated than other physiological modalities. This study evaluates EDA-based biometric recognition across different recording locations and physiological conditions, focusing on hand and foot signals. The recordings were divided into overlapping temporal windows, from which 48 features were extracted covering statistical, temporal, area-based, dynamic, tonic, phasic, frequency, and complexity characteristics. Six machine-learning classifiers were evaluated using purged within-subject cross-validation. Hand and foot signals were analyzed separately and subsequently combined through feature-level fusion, resulting in a 96-feature representation. SHAP (SHapley Additive exPlanations) was applied exclusively to the training data to rank feature importance and prevent information leakage. The selected features were then used to retrain the classifiers and assess the effect of feature reduction. In the hand–foot fusion experiment, MLP achieved a mean accuracy of 68.35% with 40 SHAP-selected features, compared with 68.31% using all 96 features, corresponding to a 58% reduction in feature dimensionality with no statistically significant loss in recognition accuracy (p = 0.275). KNN showed the largest improvement from feature reduction, increasing from 55.90% with 96 features to 63.43% with 40 features, corresponding to an absolute increase of 7.53 percentage points, which was statistically significant according to the paired Wilcoxon signed-rank test (p = 0.00195). SHAP analysis identified statistical, tonic, and area-based features among those contributing most to subject recognition. The results show that feature reduction has a classifier-dependent effect, while SHAP can identify an appropriate feature subset for each model. Overall, EDA contains subject-specific information across recording locations and physiological conditions, supporting its potential as a wearable biometric modality and its future integration with complementary physiological signals.