(PS9-88) From Therapy to Daily Life: Understanding Youths’ Challenges in Applying Therapy Skills with Large Language Models

Sunday, November 23, 2025

10:00 AM - 11:00 AM CST

Location: Elite Hall A, Level 1

Author(s)

ST

Sean Toh, B.S.

Clinical Psychology PhD Student
Harvard University
Cambridge, Massachusetts, United States

Co-Author(s)

KA

Kate Adams, B.A.

Research Fellow
Harvard University
Cambridge, Massachusetts, United States
DL

Daniel Low, Ph.D.

Pos Doc
Department of Psychology
Somerville, Massachusetts, United States
JW

John R. Weisz, ABPP, Ph.D.

Research Professor
Harvard University
Cambridge, MA, United States

Abstract Text: Introduction

Despite the growing number and efficacy of youth mental health treatments, recent research suggests that youths rarely apply therapeutic skills learned in psychotherapy to real-life situations. Given that consistent use of therapy skills in real-life is strongly associated with significant improvements in treatment outcomes, it is crucial to identify barriers that hinder their application.

Methods

This study analyzed 2,196 recorded psychotherapy sessions ( >2,500 hours) from 198 children and adolescents who participated in the Modular Approach to Therapy for Children with Anxiety, Depression, Trauma, or Conduct Problems (MATCH-ADTC) between 2009 and 2018. MATCH-ADTC is an evidence-based modular treatment designed to address common childhood mental health problems, with a strong emphasis on helping youths apply therapy skills outside of sessions. All psychotherapy recordings were de-identified, transcribed, and speaker-labeled using state-of-the-art speaker separation deep learning models. Large Language Models (LLMs; GPT-4o via a Python API) were then applied to these transcripts to identify barriers to real-life therapy skill application and whether youths had used these skills prior to the session, as reported by their parents in-session. Each factor was coded on a binary scale (0 = construct not present or unclear; 1 = clearly present).

Results

Of the 2,196 psychotherapy sessions, 617 (28.1%) included mentions of youths successfully applying learned therapy skills in real life, and 1,120 (51.0%) reflected difficulties in skill application. Our NLP analysis identified 534 distinct barriers related to therapy skill use in real-life, which were grouped into nine categories along with their prevalence rates: “Time Constraints” (21.4%), “Symptoms Affecting Skill Use” (20.7%), “Lack of Interest in Applying Skills” (16.9%), “Lack of Parental and Family Support” (15.8%), “Perceived Ineffectiveness of Skills” (10.0%), “Social Challenges” (6.3%), “Stigma and Social Perception” (3.6%), “Lack of Confidence in Skill Execution” (2.7%), “Lack of Resources” (2.1%). Time-lagged mixed-effects generalized linear models showed that "Lack of Parental and Family Support" (β = -0.482, OR: 0.618, p < .05) and "Symptoms Affecting Skill Use" (β = -0.385, OR: 0.682, p < .05) were significantly associated with a lower likelihood of having practiced therapy skills in real-life prior to the next session. These findings highlight the importance of addressing immediate symptoms, while also recognizing that successful real-life application of skills among youths heavily depends on the involvement and support of their families.

Conclusion

This study highlights the potential of AI-driven approaches (automated speech recognition, speaker separation, LLMs) in analyzing large-scaled psychotherapy data to uncover valuable insights into using learned therapy skills in real-life. As real-life skill application remains a largely underexplored area, future research could leverage similar approaches to examine additional mechanisms influencing therapy skill use.