Category: Suicide and Self-Injury
Keywords: Suicide | Longitudinal | Research Methods
Presentation Type: Symposium
Science’s ability to predict future suicide has been on par with random guessing for the past 50 years (Franklin et al., 2017). We conducted the present project in an effort to improve this state of affairs. The aim of the present project was three-fold. First, we applied machine learning to a large dataset of electronic health records (EHRs) to develop risk algorithms to predict suicide death. Second, we used machine learning to predict suicide risk over time. Third, to provide a direct comparison to the traditional approach, we compared prediction using machine learning to prediction derived from conventional logistic regression.
The sample was drawn from a curated data repository of EHRs that included data from over two million patients. We identified 502 suicide decedents; cause of death was verified through the National Death Index. Suicide decedents were compared to 11,774 general patient controls and 26,516 depressed controls. Multiple imputation of chained equations was applied to handle missing data. Random forest was applied as our main ML approach. Bootstrapping was used to assess for variance and correct for optimism (i.e., guard against overfitting). For comparison, logistic regression analyses were conducted. Kolmogorov-Smirnov testing was used to assess normality and inform direct comparisons of performance.
ML algorithms accurately predicted suicide death, with improvement in predictive accuracy as suicide death became more imminent. Prediction was strongest one week before suicide death, comparing against general patient (AUC = .88 [.85, .91]) and depressed controls (AUC = .85 [.82, .88]). Machine learning algorithms also outperformed traditional logistic regression, which produced accuracy estimates only marginally better than chance (logistic regression: .58 [.48, .67]; p<.001).
Taken together, these findings suggest that ML approaches have the potential to make substantive advances toward accurate and scalable suicide risk detection. As these results also speak to the temporal variance of suicide risk, they may help to inform the timing of risk detection and prevention efforts.
Florida State University
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.