Early Attrition Prediction for Web-Based Interpretation Bias Modification to Reduce Anxious Thinking: A Machine Learning Study.

Background: Digital mental health is a promising paradigm for individualized, patient-driven health care. For example, cognitive bias modification programs that target interpretation biases (cognitive bias modification for interpretation [CBM-I]) can provide practice thinking about ambiguous situations in less threatening ways on the web without requiring a therapist. However, digital mental health interventions, including CBM-I, are often plagued with lack of sustained engagement and high attrition rates. New attrition detection and mitigation strategies are needed to improve these interventions.

Objective: This paper aims to identify participants at a high risk of dropout during the early stages of 3 web-based trials of multisession CBM-I and to investigate which self-reported and passively detected feature sets computed from the participants interacting with the intervention and assessments were most informative in making this prediction.

Methods: The participants analyzed in this paper were community adults with traits such as anxiety or negative thinking about the future (Study 1: n=252, Study 2: n=326, Study 3: n=699) who had been assigned to CBM-I conditions in 3 efficacy-effectiveness trials on our team's public research website. To identify participants at a high risk of dropout, we created 4 unique feature sets: self-reported baseline user characteristics (eg, demographics), self-reported user context and reactions to the program (eg, state affect), self-reported user clinical functioning (eg, mental health symptoms), and passively detected user behavior on the website (eg, time spent on a web page of CBM-I training exercises, time of day during which the exercises were completed, latency of completing the assessments, and type of device used). Then, we investigated the feature sets as potential predictors of which participants were at high risk of not starting the second training session of a given program using well-known machine learning algorithms.

Results: The extreme gradient boosting algorithm performed the best and identified participants at high risk with macro-F₁-scores of .832 (Study 1 with 146 features), .770 (Study 2 with 87 features), and .917 (Study 3 with 127 features). Features involving passive detection of user behavior contributed the most to the prediction relative to other features. The mean Gini importance scores for the passive features were as follows: .033 (95% CI .019-.047) in Study 1; .029 (95% CI .023-.035) in Study 2; and .045 (95% CI .039-.051) in Study 3. However, using all features extracted from a given study led to the best predictive performance.

Conclusions: These results suggest that using passive indicators of user behavior, alongside self-reported measures, can improve the accuracy of prediction of participants at a high risk of dropout early during multisession CBM-I programs. Furthermore, our analyses highlight the challenge of generalizability in digital health intervention studies and the need for more personalized attrition prevention strategies.