Using Natural Language Processing to Evaluate Differences in Psychotherapeutic Services for Posttraumatic Stress...

Objective: Posttraumatic stress disorder (PTSD) is a common psychiatric condition, especially among US Veterans. Individuals diagnosed with PTSD have higher likelihoods of experiencing suicidal thoughts, attempting suicide, and dying by suicide compared to those without PTSD diagnoses. Although the US Department of Veterans Affairs (VA) emphasizes psychotherapy as a leading treatment for PTSD and associated suicide risk, there is limited understanding about how these treatments are used by VA patients who did and did not die by suicide. To better assess these patients' psychotherapy usage differences, we used Latent Dirichlet Allocation, a natural language processing topic modeling methodology, to study patients' unstructured electronic health record (EHR) note corpus.

Methods: We evaluated VA suicide-risk-stratified patients (high-, moderate-, and low-suicide-risk) who died by suicide in 2017-2018 (cases) and suicide-risk-matched patients with similar demographics, diagnoses, and care who did not die by suicide (controls). After collecting all psychotherapy EHR notes within 1 year of case death and completing corpus preprocessing, we derived topics and used a binomial logistic regression model to evaluate topic differences, calculate odds ratios and P values, and examine topic clinical relevance.

Results: We identified 5 topics: Risk, Treatment Planning, Evaluation, Psychosocial, and Medication. Cases and controls had several significantly different topic patterns, including Risk differences for moderate-risk patients, Treatment Planning differences for moderate- and high-risk patients, Evaluation differences for high-risk patients, Psychosocial differences for low- and moderate-risk patients, and Medication differences for all patient subgroups.

Conclusion: Topic differences help distinguish closely matched cases and controls, aiding understanding of psychotherapy utilization and risk monitoring. Our findings suggest divergent care priorities, such that evaluation and risk monitoring are more central for high-risk cases while collaborative treatment planning and medication management are more central for high-risk controls.