Assessing Suicide Risk Reduction in the Interval Before Death by Suicide Using Natural Language Processing

Objective: High-suicide-risk patients who quickly respond to psychotherapy, such that they cease being classified as high-risk, frequently die by suicide soon after risk declassification. It is accordingly difficult to distinguish symptom improvement and changes that may be fleeting or associated with evading treatment. This exploratory study used natural language processing to examine care patterns preceding suicide among Veterans Affairs (VA) patients whose classified-risk status remained elevated and those whose classified risk diminished in the interval before death by suicide.

Methods: We identified (n=89) patients classified as high-suicide-risk whose high-risk status was later removed and died by suicide within 6 months, between 2015-2018 (Reduced Clinical Risk Classification [RCRC]). We then selected comparison (n=89) patients who retained high-risk status and died within the same interval (Sustained Clinical Risk Classification [SCRC]). We analyzed electronic health records using statistical models and topic modeling, a natural language processing method that gauges lexical patterns, and monitored risk fluctuations.

Results: RCRC patients were younger, with fewer psychiatric diagnoses. Topics included "Psychotherapy," "Medication," "Outpatient," "Follow-up," and "Suicide-Assessment." After risk declassification, RCRC patients demonstrated sharp declines across all topics, whereas SCRC patients maintained or increased engagement.

Conclusion: Our preliminary results suggest that RCRC patient records featured more psychotherapy-related terminology, while SCRC patient records featured more structured and medically oriented terminology. This dichotomy may infer distinct care trajectories. Perceived symptom improvement could prompt reduced surveillance, creating unrecognized vulnerability. Our findings are VA specific and should be interpreted cautiously. Future work will compare patients who died and did not die after risk declassification.