Related Experiment Video
Updated: Jun 11, 2025
07:31
Implementation of a Real-Time Psychosis Risk Detection and Alerting System Based on Electronic Health Records using CogStack
Published on: May 15, 2020
7.0K
Using Natural Language Processing to develop risk-tier specific suicide prediction models for Veterans Affairs
Maxwell Levis1, Monica Dimambro2, Joshua Levy3
1White River Junction VA Medical Center, 215 North Main Street, White River Junction, VT, 05009, USA; Geisel School of Medicine at Dartmouth, 1 Rope Ferry Rd, Hanover, NH, 03755, USA.
Journal of Psychiatric Research
|October 1, 2024
Summary
New suicide prediction models for Department of Veterans Affairs (VA) patients improve accuracy for all risk levels. These models leverage electronic health records to better identify suicide risk in moderate- and low-risk veteran populations.
Area of Science:
- Public Health
- Mental Health
- Data Science
Background:
- Suicide is a significant cause of death, with elevated rates among Department of Veterans Affairs (VA) patients.
- Current VA suicide prevention efforts predominantly focus on high-risk individuals, who represent less than 10% of suicide deaths.
- Previous research identified moderate- and low-risk patient groupings within the VA system.
Purpose of the Study:
- To refine suicide prediction methods tailored to high-, moderate-, and low-risk veteran patient groups.
- To enhance the VA's leading suicide prediction model by incorporating risk-tier-specific algorithms.
- To improve the identification of suicide risk in underserved, non-high-risk veteran populations.
Main Methods:
- Utilized national VA data, including 4,584 veteran suicides from 2017-2018 and matched controls.
- Analyzed unstructured electronic health record (EHR) notes using natural language processing (NLP).
- Applied machine learning classification algorithms to develop and evaluate risk-tier-specific predictive models, assessing accuracy using Area Under the Curve (AUC).
Main Results:
- Developed risk-tier-specific models with significant predictive accuracy: high-risk (AUC=0.621), moderate-risk (AUC=0.669), and low-risk (AUC=0.673).
- These models demonstrated superior predictive performance compared to the VA's existing leading suicide prediction algorithm.
- Analysis of derived words revealed distinct patterns for each risk group: chronic conditions (high-risk), outpatient care (moderate-risk), and acute conditions (low-risk).
Conclusions:
- Leveraging unstructured EHR data through NLP and machine learning offers substantial benefits for suicide risk prediction.
- The study provides refined, risk-tier-specific prediction tools, expanding resources for moderate- and low-risk veteran populations.
- These advancements are crucial for addressing suicide prevention in historically underserved veteran groups within the VA system.
