Prediction of adolescent suicide attempt by integrating clinical, neurocognitive and geocoded neighborhood environment data

Abstract

Objective- The authors used multimodal data collected during pre/early-adolescence in research settings to predict self-report of past suicide attempt (SA) by mid-late adolescence in pediatric care settings. The study further aimed to determine types of data that contribute to the prediction; and to test generalizability of the prediction in an independent sample. Methods- The authors applied machine learning methods on clinical, neurocognitive and geocoded neighborhood environmental data from the Philadelphia Neurodevelopmental Cohort study (PNC, Mean age 11.1, SD=2.2, 52.3% female and 51.4% Black participants) to predict SA reported ~5 years later in a community pediatric clinic (n=922, 5.3% SA) or emergency department (n=497, 8.2% SA). The authors compared prediction performance when using all data versus individual data types, then used feature selection algorithms (Lasso, Relief and Random Forest) to identify important predictors and compared performance of models relying on feature subsets. Results- In the training dataset, SA prediction was good, reaching AUC=0.75, sensitivity/specificity 0.76/0.77 when relying on feature subsets identified using feature selection pipelines. Use of highest-ranking feature subsets from the training dataset yielded similar prediction in the testing dataset with AUC=0.74, sensitivity/specificity 0.66/0.70. Different algorithms showed different high-ranking features, but overall multiple data domains were represented among the highest-ranking predictors selected by each algorithm. Conclusions- These findings provide evidence for feasibility of predicting youth SA using data collected at a single timepoint early in life in a diverse cohort. Results encourage incorporation of multiple data types including neurocognitive and geocoded environmental measures in machine learning SA prediction pipelines.