Using natural language processing to assess proxy measures of therapeutic alliance across suicide risk tiers

Therapeutic alliance is a core predictor of psychotherapeutic outcome, in general, and suicide prevention, specifically. Although clinical literature suggests that maintaining positive alliance with high-suicide-risk patients can be particularly challenging, few studies evaluate this phenomenon, primarily due to alliance measurement difficulties. Our study derives proxy alliance metrics using natural language processing and models differences across classified suicide risk strata and suicide decedent status. We first identified all United States Department of Veterans Affairs (VA) psychotherapists who treated patients who died by suicide within discreet suicide-risk tiers (cases) and then identified suicide-risk matched patients who did not die (controls). We then extracted all psychotherapist electronic health records and used BERTopic, a topic modeling algorithm, to derive core topics within this corpus, and evaluated difference in topic slopes when controlling for suicide-risk-tier and case/control status. Using this method, we identified three clusters of psychotherapists with significantly different topic probabilities, each endorsing alternative proxy alliance frameworks. Our identified clusters centered on reduced psychosocial care as risk increases, increased psychosocial care as risk increases, and consistent care as risk increases. Rather than being solely a matter of clinical training or professional identity, these differences suggest clinicians' discrete emotional and behavioral reactions to patients' risk. Our findings suggest the possibility of scalable proxy assessments of alliance, offering potential utility in supporting clinicians working in high-risk care. Future work should examine whether these clusters predict patient engagement, dropout, or intervention impact, and explore whether targeted training can modulate clinicians' alliance-promoting behaviors in ways that improve outcomes.