PREDICTING SUICIDE ATTEMPT IN THOSE WITH IDEATION USING CLINICAL, GENETIC, AND SOCIAL FACTORS

Understanding and identifying which patients with suicidal ideation (SI) will attempt suicide (SA) is a significant clinical challenge. According to the ideation-to-action framework, SI and SA share many common risk factors but are distinct behaviors. Only a small subset of individuals with SI progress to SA, and this progression depends on an individual’s social context and capability for suicide. Current statistical suicide risk models reliant on clinical factors may be inadequate here. We investigated whether incorporating genetic data and social and behavioral factors (SBFs) into existing suicide risk models might improve recognition of individual risk and improve their predictive performance. We examined integrating individual-level genetic data and SBFs with clinical data for patients presenting to the emergency department with SI (ED-SI visit) using de-identified electronic health record (EHR) data from Vanderbilt University Medical Center (VUMC). To simulate a real-time clinical scenario, the ED-SI visit was the prediction point. Regularized Cox regression and random survival forests (RSF) were used to predict SA within 90 days. We assessed performance first using a validated risk score alone. Then, six SBFs (adverse childhood experiences (ACE), chronic stress, financial insecurity, homelessness, social isolation, loneliness) were extracted from clinical notes using a natural language processing (NLP)-based phenotype retrieval algorithm, appended to the risk score, and performance was reassessed. Finally, 16 polygenic risk scores including anxiety, bipolar disorder, schizophrenia, and risk-taking behavior were added. Model performance was assessed with area under the precision-recall curve (AUPRC) and positive predictive value (PPV). Feature importance was analyzed using Shapley Additive Explanations (SHAP). We identified 8,841 ED-SI visits; 328 (3.7%) resulted in SA within 90 days. Regularized Cox regression and RSF showed integrating SBFs achieved nearly a three-fold increase in AUPRCs and two-fold increase in PPVs compared to clinical risk scores alone. The Cox model with SBFs achieved an AUPRC of 0.138 and a PPV of 0.101, while the RSF model with SBFs achieved an AUPRC of 0.113 and a PPV of 0.091. The clinical risk score-only models had significantly lower performance metrics (Cox: AUPRC = 0.052, PPV = 0.053; RSF: AUPRC = 0.046, PPV = 0.049). Genetic data inclusion did not improve predictive performance. SHAP analysis of the regularized Cox regression model further highlighted the impact of SBFs, identifying homelessness, chronic stress, and ACE as the strongest predictors of SA, while the clinical risk score was identified as the weakest predictor. Social isolation, loneliness, and financial insecurity exhibited variability in risk impact, leading to lower overall feature importance. The study demonstrated that integrating NLP-extracted individual-level SBFs into suicide risk prediction models significantly improves predictive performance for patients presenting to the ED with SI. Homelessness, chronic stress, and ACE were identified as the strongest predictors of SA, while social isolation, loneliness, and financial insecurity had variable effects on suicide risk. These findings emphasize the need for further investigation into the context-specific impact of these factors. Future work includes refining these insights and expanding to incorporate biologic factors building based on genetic data analyzed here.