Prediction of Adolescent Suicide Attempt by Integrating Clinical, Neurocognitive and Geocoded Neighborhood Environment Data.

background and hypothesisSuicide attempt is a complex behavior influenced by a combination of factors including clinical, neurocognitive, and environmental. We aimed to leverage multimodal data collected during pre/early adolescence in research settings to predict self-report of suicide attempts by mid-late adolescence reported in pediatric settings. We hypothesized that different data types contribute to suicide attempt prediction and that clinical features would be most predictive of future suicide attempts. STUDY

designWe applied machine learning methods to clinical, neurocognitive, and geocoded neighborhood environmental data from the Philadelphia Neurodevelopmental Cohort study (Mean age [SD] = 11.1 [2.2], 53.3% female, 51.4% Black participants) to predict suicide attempt reported ~5 years later in two independent pediatric settings: primary care (n = 922, 5.3% suicide attempt) or emergency department (n = 497, 8.2% suicide attempt). We tested prediction performance using all data versus using subsets of features identified by three feature selection algorithms (Lasso, Relief, Random Forest). STUDY

resultsIn the primary care sample, suicide attempt prediction using subsets of selected features (predictors) was good, achieving AUC = 0.75, sensitivity/specificity 0.76/0.77. The use of highest-ranking features yielded similar prediction performance in external validation using the independent emergency department sample with AUC = 0.74, sensitivity/specificity 0.66/0.70. Different algorithms identified different high-ranking features, but overall multiple data domains were represented among the highest-ranking features. Besides suicidal ideation, the highest-ranking clinical predictive symptoms were from psychosis or mania spectrum.

conclusionsResults suggest that data collected at a single timepoint during preadolescence can inform suicide attempt prediction during mid-late adolescence, in different clinical settings. Findings encourage incorporation of multiple data types including neurocognitive and geocoded data, alongside clinical data, in machine learning suicide attempt prediction pipelines.