A machine learning model that analyzes language from therapy sessions accurately predicted which patients with treatment-resistant depression would respond to psilocybin treatment. Transcripts of psychological support sessions held one day after COMP360 (a synthetic psilocybin formulation) administration were analyzed using a zero-shot classifier based on the BART large language model to measure sentiment (valence and arousal) for both participant and therapist. These scores, combined with the Emotional Breakthrough Index and treatment arm, were used to predict treatment outcome measured by MADRS scores. Two multinomial logistic regression models predicted responder status at week 3 and through week 12 with 85% and 88% accuracy, and AUC values of 88% and 85%, respectively. The approach enables rapid prognostication of personalized response to psilocybin treatment and insights into therapeutic model optimization.
In a 12-week clinical trial of 25 mg COMP360 psilocybin for 22 participants with post-traumatic stress disorder, audio recordings showed that during drug-administration sessions speech by either party was rare: silence filled 78% of the time on average, compared to 25% to 30% in non-administration sessions. Thematic analysis of post-dosing interviews revealed that support was minimally enacted but experientially salient, autonomy was promoted through the introspective state and non-directive support, and primary modes of support during altered states included reassurance and validation. The minimal verbal interaction distinguishes this monitoring and support from conventional psychotherapies and MDMA-assisted therapy.