Incorporating a deep‐learning client outcome prediction tool as feedback in supported internet‐delivered cognitive behavioural therapy for depression and anxiety: A randomised controlled trial within routine clinical practice

Abstract Introduction Machine learning techniques have been leveraged to predict client psychological treatment outcomes. Few studies, however, have tested whether providing such model predictions as feedback to therapists improves client outcomes. This randomised controlled trial examined (1) the effects of implementing therapist feedback via a deep‐learning model (DLM) tool that predicts client treatment response (i.e., reliable improvement on the Patient Health Questionnaire‐9 [PHQ‐9] or Generalized Anxiety Disorder‐7 [GAD‐7]) to internet‐delivered cognitive behavioural therapy (iCBT) in routine clinical care and (2) therapist acceptability of this prediction tool. Methods Fifty‐one therapists were randomly assigned to access the DLM tool (vs. treatment as usual [TAU]) and oversaw the care of 2394 clients who completed repeated PHQ‐9 and GAD‐7 assessments. Results Multilevel growth curve models revealed no overall differences between the DLM tool vs. TAU conditions in client clinical outcomes. However, clients of therapists with the DLM tool used more tools, completed more activities and visited more platform pages. In subgroup analyses, clients predicted to be ‘not‐on‐track’ were statistically significantly more likely to have reliable improvement on the PHQ‐9 in the DLM vs. TAU group. Therapists with access to the DLM tool reported that it was acceptable for use, they had positive attitudes towards it, and reported it prompted greater examination and discussion of clients, particularly those predicted not to improve. Conclusion Altogether, the DLM tool was acceptable for therapists, and clients engaged more with the platform, with clinical benefits specific to reliable improvement on the PHQ‐9 for not‐on‐track clients. Future applications and considerations for implementing machine learning predictions as feedback tools within iCBT are discussed.