Predicting psychiatric symptom deterioration using heart rate variability: the role of temporal feature engineering

Background Psychiatric care critically lacks objective biomarkers for monitoring symptom trajectories, often delaying the detection of clinical deterioration. While heart rate variability (HRV) offers promise as a continuous, non-invasive marker of autonomic function, prior prediction models have achieved only modest performance (AUC 0.55–0.65). This limitation likely stems from a reliance on static features that fail to capture the complex temporal dynamics and substantial inter-individual variability inherent in longitudinal physiological data. Aims & Objectives To develop and validate a comprehensive temporal feature engineering framework for HRV-based prediction of psychiatric symptom deterioration, and to identify statistically validated predictors with clinical interpretability. Method We analyzed longitudinal data from 193 psychiatric inpatients (633 assessments) across four symptom domains: general psychopathology (BPRS), depression (MADRS), mania (YMRS), and anxiety (HAMA). Beyond 164 standard HRV features, we engineered temporal features including delta (visit-to-visit change) and cumulative (expanding window) aggregations to capture multi-scale temporal dynamics, expanding the feature space to 600–800 features. Models were validated using TRUE nested cross-validation with rigorous four-level data leakage prevention (patient-level splitting, temporal leakage prevention, independent feature engineering, and nested feature selection). Generalized estimating equations (GEE) were employed to statistically confirm feature associations while accounting for within-patient correlation. All GEE models included adjustment for age, sex, psychiatric diagnosis, psychotropic medication class and duration, and body mass index to address potential confounding. Results The comprehensive temporal model achieved an average test AUC of 0.70 (95% CI: 0.654–0.746), improving substantially over a baseline model using 164 standard HRV features only (AUC 0.464; +0.236 AUC, +51%). In ablation analyses, both delta features (+0.040 AUC vs. NO Delta) and cumulative features (+0.025 AUC vs. Markov) contributed independently to predictive accuracy. GEE analysis with Bonferroni correction (adjusted α = 0.0125 for 4 symptom scales) identified cumulative HRV instability—particularly Mean RR cumulative standard deviation—as the strongest predictor of deterioration (OR = 1.98 for depression, OR = 1.61 for mania; both p < 0.01). Higher parasympathetic activity (HF power) showed a protective effect against manic deterioration (OR = 0.67, p < 0.001). The number needed to harm (NNH) for 1-SD increase in Mean RR cumulative standard deviation for depressive deterioration was 3.1 (95% CI: 2.3–4.8), indicating clinically meaningful magnitude. Cross-validation estimates closely matched test performance (gap = +0.034), supporting robust generalization without major data leakage. Discussion & Conclusions Temporal feature engineering substantially improves HRV-based prediction of psychiatric symptom deterioration by effectively capturing within-subject dynamics. Cumulative HRV instability emerges as a novel transdiagnostic biomarker, offering a pathway for integrating objective, wearable-based monitoring into proactive clinical decision support systems.