Abstract / Summary
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). While electrocardiography (ECG) remains the gold standard for HRV assessment, its scalability in outpatient rehabilitation is limited. Consumer-grade wearables using photoplethysmography (PPG) offer a practical alternative, but systematic calibration of PPG-derived HRV against ECG standards has been lacking. We evaluated the feasibility of using calibrated PPG-derived HRV for stress-state classification as a proxy for relapse-relevant autonomic signatures. Using the WESAD public dataset (15 subjects, ECG + PPG), we extracted time-domain (RMSSD, SDNN, pNN50), frequency-domain (LF, HF, LF/HF), and nonlinear (sample entropy, DFA) HRV features. Classical machine learning models achieved 80% accuracy with ROC AUC of 0.85 for binary stress classification using ECG-derived features. Linear calibration mapping PPG features to ECG space improved PPG-based ROC AUC from 0.79 to 0.85, matching ECG performance. Deep sequence models (LSTM, CNN) did not improve over feature-based approaches in this dataset, likely due to limited sample size. We additionally collected seven-day concurrent PPG recordings from Empatica EmbracePlus (FDA-cleared) and Fitbit (consumer-grade) wearables worn by a study participant to characterize cross-device agreement in ambulatory settings. These results establish that calibrated PPG-derived HRV from consumer wearables can approximate ECG-level discriminative performance for stress-state classification, supporting the feasibility of scalable HRV-based relapse monitoring in AUD rehabilitation.