A two-stage diagnostic classification framework integrating heart rate variability and clinical features to identify...

Accurate identification of recent suicide-related risk remains challenging in psychiatric settings. Although clinical assessment is essential, objective physiological indicators such as heart rate variability (HRV) may provide complementary information. This study developed a two-stage machine-learning framework integrating clinical features and HRV to identify recent suicide-related status among psychiatric inpatients.

This retrospective study included 619 psychiatric inpatients who underwent standardized clinical assessment and 5-min resting HRV recording. Patients were classified as no suicidal ideation or behavior (no SI/SB; n = 416), suicidal ideation without suicidal behavior (SI; n = 137), or suicidal behavior (SB; n = 66) based on suicide-related information within the month before admission. Stage 1 classified patients with any recent SI/SB versus no SI/SB, and Stage 2 further classified recent SB versus SI among patients with suicide-related risk. Demographic-clinical, demographic-HRV, and combined feature sets were evaluated using multiple machine learning algorithms with repeated nested cross-validation.

The combined feature set achieved the best overall performance. In Stage 1, the best combined model showed good discrimination (AUC = 0.894; AUPRC = 0.831). In Stage 2, the best combined model showed moderate performance (AUC = 0.778; AUPRC = 0.686). Depressive symptom severity, self-harm, past suicide attempt, younger age, diagnosis, and reduced vagally mediated HRV contributed to identifying any SI/SB, whereas past suicide attempt was the leading predictor of SB.

Integrating clinical and HRV features may support the identification and stratification of recent suicide-related status in psychiatric inpatients. HRV provided modest complementary information, particularly for broad SI/SB identification, but external prospective validation is required.