Language signals of behavioral activation predict real-world outcomes for Ash users

An SDMH 2026 flash talk: LLM-detected behavioral-activation language in 279 Ash users tracked self-reported activation and predicted lower later depression.

Behavioral-activation language that an LLM detected in Ash conversations tracked users' self-reported activation and predicted lower later depression — a proof of concept for monitoring therapeutic change without adding user burden.

Study type

Observational study (NLP of session transcripts)

Published in

Society for Digital Mental Health

Partner

New York University

Citation

Stamatis, C. A., Malgaroli, M., & Hull, T. D. (2026). Natural Language Processing of Behavioral Activation Predicts Real-World Clinical Outcomes with AI Mental Health Support [Conference flash talk]. Society for Digital Mental Health Annual Meeting.

A flash talk at the Society for Digital Mental Health 2026 meeting presented a study of 279 Ash users over six weeks. A clinician-refined prompt had Claude 3.5 Haiku rate 5,062 session transcripts for nine behavioral-activation (BA) language markers. BA language correlated with the gold-standard self-report (r=.36) and predicted lower subsequent depression at both the between- and within-person levels, suggesting language-based markers can index treatment progress in real time.

Between-person effects also linked BA language to lower anxiety, though within-person anxiety effects were not significant; self-reported behavioral activation showed parallel but stronger effects. As a proof-of-concept observational analysis, it indicates automated, language-based metrics could enable continuous, low-burden monitoring of treatment trajectory and adaptive intervention in digital mental health.