Sufficient Biometric Signals for Wearable Exercise Classification

An Exhaustive Subset Analysis of Empatica E4 Data

Abstract

Wrist-worn wearable devices record multiple biometric signals, but not all contribute meaningfully to exercise classification. This study tested which signals could be omitted by evaluating all 31 non-empty subsets of five features — heart rate (HR), blood volume pulse standard deviation (BVP std), inter-beat interval (IBI), electrodermal activity (EDA), and skin temperature — extracted from 94 Empatica E4 sessions across 35 participants. Classification used Random Forest and Logistic Regression with participant-stratified GroupKFold cross-validation. The full model achieved 65.09 ± 10.53% accuracy; the pre-registered hypothesis pair HR + EDA achieved 65.96 ± 12.07% (permutation test p = 0.620), confirming performance within the 5-percentage-point sufficiency threshold. The best two-feature subset was HR + IBI at 68.01 ± 5.82% (p = 0.215 vs. full model). Per-class analysis showed strong resting-session discrimination (AUC = 0.96) but weaker aerobic–anaerobic separation (AUC = 0.78 and 0.76), indicating the classifier primarily distinguishes rest from exercise. Because HR and IBI both derive from a single PPG sensor, a device designer could potentially omit dedicated EDA, temperature, and BVP sensors without statistically significant accuracy loss.