Using natural language processing to evaluate temporal patterns in suicide risk variation among high-risk Veterans.

Area of Science:

• Computational linguistics
• Mental health informatics
• Suicidology

Background:

• Assessing dynamic suicide risk in high-risk populations is challenging.
• Electronic health records (EHRs) contain valuable longitudinal data.
• Natural Language Processing (NLP) offers potential for analyzing unstructured EHR data.

Purpose of the Study:

• To apply Dynamic Topic Modeling (DTM) to unstructured EHRs of high-suicide risk patients.
• To identify temporal patterns in clinical topics related to suicide risk fluctuation.
• To develop a method for time-sensitive suicide risk measurement.

Main Methods:

• Utilized a Python-based DTM algorithm to analyze EHRs of Veterans Affairs patients.
• Sample included matched cases (died by suicide) and controls (did not die by suicide) from 2017-2018.
• Analyzed clinical notes from diagnosis until the relevant end date.

Main Results:

• Identified five key topics: Medication, Intervention, Treatment Goals, Suicide, and Treatment Focus.
• Observed divergent temporal changes: pathology-focused care increased in cases, supportive care in controls.
• Case topics exhibited greater fluctuation (lability) compared to controls.

Conclusions:

• DTM can effectively monitor suicide risk fluctuation in high-risk patients.
• Increased topic lability in cases suggests a potential marker for heightened risk.
• Findings support the development of dynamic, time-sensitive suicide risk assessment tools.