Predicting suicidal and self-injurious events in a correctional setting using AI algorithms on unstructured medical...

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Public Health

Background:

• Suicidal and self-injurious incidents in correctional settings strain resources and staff.
• Traditional statistical analyses are limited by structured data collection and high implementation costs.

Purpose of the Study:

• To extract actionable insights from medical and mental health progress notes using AI.
• To predict suicidal and self-injurious events for improved healthcare triage and prevention in Orange County Jails.

Main Methods:

• Utilized AI algorithms, specifically a Transformer Encoder model, to analyze unstructured progress note data.
• Compared AI model performance using notes data alone versus a hybrid approach combining notes data with structured Electronic Health Record (EHR) data.
• Employed under-sampling techniques to address class imbalance in the dataset.

Main Results:

• Progress notes contained richer information on suicidal/injurious behaviors than structured EHR data.
• The Transformer Encoder model using notes data achieved an AUC-ROC of 0.862, Sensitivity of 0.816, and Specificity of 0.738.
• Integrating notes data into traditional Machine Learning models improved Sensitivity to 0.89 (AUC-ROC: 0.77, Specificity: 0.65).

Conclusions:

• AI analysis of clinical notes offers a more effective method for predicting inmate self-harm than structured data alone.
• Hybrid models incorporating AI-extracted features show promise for enhancing predictive accuracy.
• Under-sampling is a viable strategy for managing imbalanced data in this context.