Comprehensive Summary
Bunnel et al. evaluated deep learning models for detecting and predicting intentional self-harm (ISH) from electronic health record (EHR) notes. Using data from 4,500+ adult patients across two institutions, convolutional neural networks (CNNs) were compared with traditional bag-of-words approaches. CNNs achieved superior performance, with AUC = 0.99 and F1 = 0.94 for detecting concurrent ISH, and AUC = 0.81–0.82 (F1 = 0.61–0.64) for predicting future ISH events. These findings demonstrate that deep learning applied to EHR text can outperform traditional models, offering a promising tool for clinical suicide risk detection.
Outcomes and Implications
Suicide risk assessment is often inefficient, time-consuming, and limited in predictive power. Machine learning and deep learning models can provide more accurate, timely detection of intentional self-harm compared with traditional methods such as interviews and self-reports. Integration into electronic health records could deliver real-time alerts and track risk trajectories, giving clinicians better tools to intervene early and adapt care strategies over time.