Assessing the Utility of Social Determinants of Health Data in Suicide Prediction Models

This study seeks to explore the utility of social determinants of health (SDoH) variables in suicide prediction models. We aim to assess the impact of individual- and geographic-level SDoH factors on improving the performance of suicide prediction models and the identification of individuals at high risk for suicide.

A retrospective sample of 1214 deaths by suicide and 815,544 living patients was identified in the Maryland Suicide Data Warehouse (MSDW) and linked to census tract data through geo-coding. Three machine learning algorithms were trained and validated with cross-validation to assess model performance across different data category combinations, including demographics, clinical features, and individual and geographic level SDoH.

Models incorporating clinical information demonstrated substantial improvements in predictive performance compared with demographic-only baselines across all algorithms. In the absence of clinical data, the addition of individual level SDoH significantly improved performance compared to baseline models, where 73% of individual level SDoH factors were identified as significant. Conversely, geographic level SDoH provided limited improvement in model performance, with only 7% of variables showing significance marginally.

Within SDoH data, most individual level SDoH are among the most important risk factors associated with death by suicide, with improvement in model performance over just clinical and demographic information. Geographic level SDoH appeared to have limited added value.