Dynamic prediction of inpatient self-harm and suicide attempts using machine learning: Toward data-driven clinical monitoring

Objective: Self-harm and suicide attempts are notoriously difficult to predict and to prevent, particularly as they are low base-rate events. This study aimed to improve prediction by combining static information—unchanging patient characteristics collected at admission—with dynamic self-reported mental health data collected daily during inpatient care. Method: Seventeen thousand five hundred eight psychiatric inpatients self-reported their mental health daily throughout their hospital stay (M_age = 40.72, female = 71%). Machine-learning models predicted self-harm over the next 7-days, which made new risk projections daily based on available self-report data. Prediction was compared between models that used only static information and those that included dailyupdated information. Results: The dynamic daily-updated prediction model exhibited stable prediction performance over time, while the static model reliant on information collected at admission decreased over time. Although all models performed well in detecting events, the rate of false positives increased substantially over time in static models (excluding daily data), and model accuracy tended to decrease. Conclusions: The improved performance of prediction models leveraging dynamic (daily) self-reported data could support just-intime alerts for clinical staff. Further research is needed to identify salient markers for when the risk of self-harm and suicide attempts may be heightened to further enhance prediction accuracy.