Machine Learning Model to Predict Assignment of Therapy Homework in Behavioral Treatments: Algorithm Development and Validation

• Published in: JMIR formative research Vol. 7; p. e45156
• Main Authors: Peretz, Gal, Taylor, C Barr, Ruzek, Josef I, Jefroykin, Samuel, Sadeh-Sharvit, Shiri
• Format: Journal Article
• Language: English
• Published: Canada JMIR Publications 15.05.2023
• Subjects: Algorithms; Artificial intelligence; Clinical outcomes; Datasets; Decision making; Homework; Intervention; Machine learning; Mental health; Natural language processing; Original Paper; Review boards; Subject specialists; Supervision; Therapists; Therapy; Voice recognition; deep learning; natural language processing; homework; mental health; therapy; mHealth; empirically-based practice; behavioral treatment; treatment fidelity; intervention; machine learning; artificial intelligence
• DOI: 10.2196/45156
• ISSN: 2561-326X

Therapeutic homework is a core element of cognitive and behavioral interventions, and greater homework compliance predicts improved treatment outcomes. To date, research in this area has relied mostly on therapists' and clients' self-reports or studies carried out in academic settings, and there is little knowledge on how homework is used as a treatment intervention in routine clinical care. This study tested whether a machine learning (ML) model using natural language processing could identify homework assignments in behavioral health sessions. By leveraging this technology, we sought to develop a more objective and accurate method for detecting the presence of homework in therapy sessions. We analyzed 34,497 audio-recorded treatment sessions provided in 8 behavioral health care programs via an artificial intelligence (AI) platform designed for therapy provided by Eleos Health. Therapist and client utterances were captured and analyzed via the AI platform. Experts reviewed the homework assigned in 100 sessions to create classifications. Next, we sampled 4000 sessions and labeled therapist-client microdialogues that suggested homework to train an unsupervised sentence embedding model. This model was trained on 2.83 million therapist-client microdialogues. An analysis of 100 random sessions found that homework was assigned in 61% (n=61) of sessions, and in 34% (n=21) of these cases, more than one homework assignment was provided. Homework addressed practicing skills (n=34, 37%), taking action (n=26, 28.5%), journaling (n=17, 19%), and learning new skills (n=14, 15%). Our classifier reached a 72% F -score, outperforming state-of-the-art ML models. The therapists reviewing the microdialogues agreed in 90% (n=90) of cases on whether or not homework was assigned. The findings of this study demonstrate the potential of ML and natural language processing to improve the detection of therapeutic homework assignments in behavioral health sessions. Our findings highlight the importance of accurately capturing homework in real-world settings and the potential for AI to support therapists in providing evidence-based care and increasing fidelity with science-backed interventions. By identifying areas where AI can facilitate homework assignments and tracking, such as reminding therapists to prescribe homework and reducing the charting associated with homework, we can ultimately improve the overall quality of behavioral health care. Additionally, our approach can be extended to investigate the impact of homework assignments on therapeutic outcomes, providing insights into the effectiveness of specific types of homework.