Frequently Asked Questions

The study found significant differences in psychotherapy note themes between risk-matched Veterans with PTSD who died by suicide and those who did not, even though prior structured encounter-count analyses had found few service differences. In a matched sample of 722 cases and 8,664 controls, significant case-control differences were identified for the Risk topic at the moderate-risk tier, Treatment Planning at the moderate- and high-risk tiers, Evaluation at the high-risk tier, Psychosocial at the low- and moderate-risk tiers, and Medication at all risk tiers. The authors interpreted these findings as evidence that unstructured EHR notes can detect clinically meaningful differences in care content that structured utilization data may miss.

Suicide risk was stratified using the VA's Recovery Engagement and Coordination for Health-Veterans Enhanced Treatment (RV) algorithm, a machine learning-based prediction model built from structured EHR variables such as service use, psychotropic medication use, diagnoses, prior suicide attempts, and sociodemographic factors. The authors used 3 tiers previously derived from RV scores: high risk = scores 0-1, moderate risk = scores 2-24, and low risk = scores 25-100. Cases and controls were matched on identical RV scores at the time of the case's death.

The final matched sample included 722 cases and 8,664 controls. The high-risk tier included 171 cases and 2,052 controls, the moderate-risk tier included 402 cases and 4,824 controls, and the low-risk tier included 149 cases and 1,788 controls. Cases and controls were matched in a 1:12 ratio on identical RV suicide risk scores.

The topic model identified 5 psychotherapy note themes: Risk, Treatment Planning, Evaluation, Psychosocial, and Medication. Risk centered on emotionality and suicide risk disclosure; Treatment Planning on individual and group interventions; Evaluation on mental health concerns such as sleep, anxiety, alcohol, and depression; Psychosocial on group care and activity participation; and Medication on psychopharmaceutical care. These topic labels were assigned by a psychiatrist, social worker, and psychologist based on the model's word patterns.

The Medication topic was the most consistent differentiator, with controls showing significantly higher proportions of medication-related notes at every risk tier. The authors noted that this pattern was present in low-, moderate-, and high-risk subgroups. They suggested this may reflect less consistent medication use or greater wariness about prescribing certain medications in higher-risk patients, but did not test medication efficacy or causality.

As classified suicide risk increased, the probabilities of Risk, Treatment Planning, and Medication topics decreased, and the Evaluation topic also decreased from moderate- to high-risk patients, while the Psychosocial topic increased. The authors described these as 2 contrasting clusters: one involving Risk, Treatment Planning, Medication, and partly Evaluation, and the other involving Psychosocial content. They suggested that high-risk patients' notes may be more dominated by inpatient routines, group interventions, and care progress than by treatment planning or medication-related discussion.

Cases had significantly higher Risk topic probabilities than matched controls at the moderate-risk tier, significantly higher Evaluation topic probabilities at the high-risk tier, and significantly higher Psychosocial topic probabilities at the low- and moderate-risk tiers. For the Psychosocial topic, the difference between cases and controls narrowed as risk increased and was not significant in the high-risk tier. The authors therefore considered Psychosocial content less useful for distinguishing high-risk cases from controls.

No conclusion about trauma-focused or evidence-based psychotherapy use by risk tier can be drawn from this analysis because the study did not differentiate between psychotherapy types in the note-content models. The authors included all psychotherapy encounters identified by VA psychotherapy CPT codes and stated that their sample size and data sparsity precluded conclusions about intervention efficacy or the impact of evidence-based versus non-evidence-based treatment. They also noted that they could not control for specific intervention formats.

The researchers extracted text integration utility notes linked to psychotherapy encounters in the year before the index death, tokenized the text, removed stop-words and uninformative terms, and then applied Latent Dirichlet Allocation (LDA) to identify latent themes. They selected a 5-topic model based on perplexity curves, LDA visualizations, and reduced topic overlap and redundancy. For analysis, notes were classified as containing substantial presence of a topic when that topic proportion was at least 20%, because that cutoff produced the greatest number of significant case-control differences across risk tiers.

The findings should be interpreted as characterization of care patterns rather than proof of treatment effectiveness. Key limitations were that PTSD was identified by ICD codes rather than standardized assessment scores, the sample size and data sparsity prevented differentiation of psychotherapy types and conclusions about efficacy, the RV model itself includes service use and may bias risk-tier composition, and non-VA psychotherapy could not be captured. The authors also stated that the VA-based sample should not be generalized to civilian populations.

The study suggests that psychotherapy documentation patterns differ by suicide risk tier and by whether patients later died by suicide, which may help clinicians think about risk-tier-specific care needs. The authors found that higher-risk patients generally had less Risk, Treatment Planning, and Medication-informed care and more Psychosocial care than lower-risk patients, and they stated that these discrepancies may constrain intervention quality and efficacy. They suggested the results could be used to tailor psychotherapeutic interventions by risk level, monitor treatment quality, and enhance psychotherapeutic alliance, while avoiding causal claims about which treatments prevent suicide.