Category: Suicide and Self-Injury
Keywords: Suicide | Longitudinal | Research Methods
Presentation Type: Symposium
Traditional approaches to predicting suicide attempts may be inaccurate (Franklin et al., 2017). In this study, we used a machine learning approach to predict future suicide attempts among a large sample of hospital patients.
This longitudinal cohort study was built from a database of clinical electronic health records (EHRs) reflecting two decades of care at a single academic medical center. To accurately identify suicide attempters within this database, two suicide experts coded records from all individuals who received an E950x ICD code (i.e., self-injury code). A total of 3,250 suicide attempters were identified (cases); these were compared to 2,260 non-suicidal self-injurers (NSSI controls), 8,003 patients with depression but no suicidal behavior (depressed controls), and 12,695 individuals randomly selected from the hospital database (hospital controls). Predictive modeling via a machine learning (ML) framework incorporated algorithms including L1-regularized regression, random forests, and traditional logistic multiple regression.
Results revealed that ML approaches, particularly random forests, produced accurate prediction of future suicide attempts (AUCs = 0.83 - 0.91). Three general patterns emerged within these findings. First, accuracy improved significantly as the comparison group became less severe (NSSI controls [AUC = 0.83], depressed controls [AUC = 0.84], and hospital controls [AUC = 0.92], p p<.001). Third, accuracy was consistently poor for traditional multiple logistic regression models (AUCs = 0.62 - 0.64), despite the fact that they included the same predictors as ML models. These findings suggest that machine learning applied to EHRs may provide an accurate and scalable method of suicide attempt risk detection.
Vanderbilt University Medical Center
Friday, November 17
3:30 PM – 4:30 PM
The asset you are trying to access is locked. Please enter your access key to unlock.