Using Artificial Intelligence to Identify Effective Components of Computer-Assisted Cognitive Behavioural Therapy

Abstract

Although clinician-supported computer-assisted cognitive-behaviour therapy (CCBT) is well established as an effective treatment for depression and anxiety, less is known about the specific interventions used during coaching sessions that contribute to outcomes. The current study used artificial intelligence (AI) to identify specific components of clinician-supported CCBT and correlated those scores with therapy outcomes. Data from a randomized clinical trial comparing clinician-supported CCBT with treatment as usual in a primary care setting were utilized. Participants (n = 95) engaged in CCBT with coaching sessions. The primary outcome was the Patient Health Questionnaire (PHQ-9), with Generalized Anxiety Disorder (GAD-7), Satisfaction with Life Scale (SWLS) and Automatic Thoughts Questionnaire (ATQ) ratings as secondary outcomes, which were assessed at 12 weeks (post), 3- and 6-month follow-up. The Lyssn system utilized AI technology to code CBT techniques and common general psychotherapeutic techniques. After controlling for initial ratings, 13 Lyssn-variables were observed to be significantly associated with reducing anxiety on the GAD-7 after 12 weeks of treatment. Among the most effective CBT interventions for anxiety included the use of guided discovery, understanding, interpersonal effectiveness and agenda setting. The most beneficial intervention was the proportion of open questions across all variables. Lyssn did not identify any CBT-specific interventions significantly associated with PHQ-9, SWLS or ATQ. Therapist use of CBT-specific techniques was significantly associated with reduction of anxiety symptoms after 12 weeks, but such gains were not observed at follow up. Therapist use of open questions was observed to be the most impactful technique contributing to treatment outcomes.