ECG-derived heart rate variability achieves robust stress-state classification under leave-one-subject-out validation: a proof-of-concept for scalable alcohol use disorder relapse monitoring with cross-device heart rate agreement from consumer wearables

Abstract Heart rate variability (HRV) is a validated non-invasive marker of autonomic dysregulation associated with stress reactivity and relapse risk in alcohol use disorder (AUD). However, the generalizability of HRV-based stress classifiers across individuals has been limited by subject leakage in standard train-test splits. We evaluated HRV-based stress-state classification using leave-one-subject-out (LOSO) cross-validation to establish unbiased performance estimates. Using the WESAD public dataset (15 subjects, chest-worn ECG), we extracted time-domain (RMSSD, SDNN, pNN50), frequency-domain (LF, HF, LF/HF), and nonlinear HRV features. Under LOSO validation, Random Forest achieved 86.7% accuracy with ROC AUC of 0.946 (95% CI: 0.871-0.998) for binary stress classification — substantially higher than the 0.848 obtained from a conventional train-test split with subject leakage. Logistic Regression achieved AUC 0.873 and SVM achieved AUC 0.893. Deep sequence models (LSTM, CNN) did not improve over feature-based approaches, likely