Automated detection and prediction of suicidal behavior from clinical notes using deep learning.

Area of Science:

• Artificial Intelligence
• Clinical Informatics
• Computational Psychiatry

Background:

• Deep learning (DL) models offer potential for enhanced suicide prediction.
• Existing research faces challenges in performance consistency and site-to-site replicability.
• Validation of DL for intentional self-harm (ISH) detection and prediction is needed across diverse academic medical centers.

Purpose of the Study:

• To validate a DL approach for detecting and predicting ISH using clinical notes.
• To evaluate the generalizability of DL models for ISH prediction across multiple institutions.

Main Methods:

• Extracted clinical notes from EHRs for 1,538 ISH patients and 3,012 controls.
• Evaluated traditional bag-of-words models (Naïve Bayes, Random Forest) and CNN models (CNNr, CNNw).
• Assessed model performance for detecting ISH within 24 hours and predicting ISH 1-6 months prior.

Main Results:

• CNN models significantly outperformed bag-of-words models in detecting concurrent ISH (AUCs .99, F1 0.94).
• CNN models showed superior performance over Naïve Bayes in predicting future ISH (AUCs 0.81-0.82, F1 0.61-0.64).

Conclusions:

• Leveraging EHRs and DL models for ISH detection/prediction using clinical notes is feasible and replicable.
• The study demonstrates the potential for multi-institutional validation of DL approaches for suicide risk assessment.
• Future research will explore broader multi-site validation using diverse EHR data.