Multimodal machine-learning discrimination of recent preadmission suicidal behavior in adolescents hospitalized with depressive episodes.

BACKGROUND

Objective sleep and autonomic measures may add information beyond clinical characteristics when distinguishing adolescents with recent suicidal behavior. We evaluated multimodal machine-learning models for concurrent discrimination of recent preadmission suicidal behavior in adolescents with depressive episodes.

METHODS

This retrospective cross-sectional study included 262 inpatients aged 11-17 years: 59 with actual and 27 with interrupted attempts within 3 months before admission. Stage 1 characterized 45 candidates using seven selectors; Stage 2 compared fixed predictor sets across six algorithms using five-fold nested cross-validation repeated 10 times. Additional analyses evaluated end-to-end feature and algorithm selection within outer training folds and an actual-attempt-only outcome. Mean repeat ROC-AUC was the primary measure.

RESULTS

The nine-predictor XGBoost model achieved ROC-AUC 0.767 (SD, 0.012), versus 0.612 for clinical logistic regression, with PR-AUC 0.662 and Brier score 0.173. End-to-end ROC-AUC was 0.695 (SD, 0.024), versus 0.601 for matched clinical logistic regression. Actual-attempt-only ROC-AUCs were 0.751 for fixed-nine XGBoost and 0.678 for the end-to-end process. Adding inflammatory and metabolic variables provided only a small, uncertain gain in the original analysis.

LIMITATIONS: This single-center, retrospective cross-sectional study used same-cohort data-informed feature-set development; additional end-to-end internal validation showed lower discrimination, and external validation remains unavailable.

CONCLUSIONS

Clinical, sleep, and nocturnal autonomic features provided complementary information for concurrent discrimination of recent preadmission suicidal behavior. External and prospective validation is required before clinical use.