Predicting Suicidality Indicators from Depression Symptom Checklists Using Interpretable Machine Learning

Abstract

Early identification of individuals at elevated risk of suicidality is a critical component of mental health prevention and intervention. In many applied settings, assessment relies on self-reported depression symptom checklists rather than comprehensive clinical evaluations. This study investigates whether suicidality indicators can be predicted using only non-suicidality depression symptoms, while emphasizing interpretability and screening-oriented evaluation. Using a publicly available depression checklist dataset comprising 2,523 unique responses to 25 Likert-scale items, suicidality outcomes were constructed from three suicidality-related items and excluded from the predictor set to prevent information leakage. Interpretable machine learning models, including logistic regression and shallow decision trees, were compared with a gradient boosting model under stratified five-fold cross-validation. Model performance was evaluated using discrimination metrics, threshold-dependent screening measures, and probability calibration. The best- performing model achieved an AUROC of approximately 0.78, while interpretable baselines performed competitively. High-recall operating thresholds substantially reduced false negatives, highlighting the suitability of the approach for screening applications. Calibration analysis showed that isotonic calibration improved the reliability of predicted probabilities. Symptoms reflecting negative self-concept, hopelessness, and anhedonia emerged as the strongest predictors of suicidality indicators. Although outcomes were proxy measures derived from self-report data rather than clinical diagnoses, the results demonstrate that interpretable machine learning can extract clinically plausible screening signals from standard depression checklists. This work provides a transparent and reproducible framework for ethically grounded suicidality risk screening research.