Related Experiment Video
Updated: Oct 28, 2026
Closed-Loop Neurostimulation for Biomarker-Driven, Personalized Treatment of Major Depressive Disorder
Published on: July 7, 2023
An explainable machine learning framework integrating heart rate variability and insomnia for severe BDI-II-defined
Sanghoon Choi1, Hyo-Chang Seo2, Hong Jin Jeon3
1Digital Therapeutics Research Center, Smart Healthcare Research Institute, Samsung Medical Center, Seoul, 06351, South Korea; Department of Medical Device Management and Research, SAIHST, Sungkyunkwan University, Seoul, 06355, South Korea.
Background:
Depression is a common mental health condition requiring repeated symptom severity assessment. While short-term electrocardiography (ECG)-derived heart rate variability (HRV) provides objective autonomic markers, it may not fully capture the clinical burden of depressive symptoms. We developed an interpretable machine learning framework combining short-term ECG-derived HRV and evaluated the incremental value of insomnia symptoms for classifying severe depressive symptom burden (Beck Depression Inventory-II ≥ 29).
Methods:
We analyzed 2663 psychiatric outpatients. Seven classifiers, a penalized logistic regression baseline, and three voting ensembles were evaluated using stratified 5-fold cross-validation with validation-selected decision thresholds. Feature ablation, decision curve analysis, and explainable artificial intelligence (SHAP and LIME) examined the contributions of HRV and Hamilton Depression Rating Scale (HAMD)-derived insomnia score.
Results:
The weighted soft-voting ensemble achieved the highest performance, with an area under the receiver operating characteristic curve (AUROC) of 0.837 (95%CI, 0.821-0.852), accuracy 0.756, sensitivity 0.721, specificity 0.776, and F1 score 0.686. Adding the insomnia score improved ensemble AUROC from 0.701 to 0.837 (Δ = 0.136, P < 0.001), while adding HRV features to age, gender, and insomnia score increased AUROC from 0.737 to 0.837 (Δ = 0.100, P < 0.001). Excluding nine duration-sensitive HRV indices yielded an AUROC of 0.787. SHAP and LIME identified age, insomnia score, and mean heart rate as primary contributors, confirming patient-level interpretability.
Conclusions:
Adding insomnia score to ECG-derived HRV appears to improve the classification of severe depressive symptom burden over HRV alone. This interpretable, low-burden framework shows potential clinical utility, suggesting that it may serve as an adjunctive screening tool for depression care.
