Using Machine Learning to Predict Uptake to an Online Self-Guided Intervention for Stress During the COVID-19 Pandemic.

Online self-guided interventions appear efficacious for alleviating some mental health concerns. However, among persons who are offered online interventions, only a fraction access them (i.e., achieve uptake). Machine learning methods may be useful to predict who will achieve uptake, which could inform improvements to interventions and their methods of delivery. We used secondary data from participants given access to a self-guided online stress intervention during the COVID-19 pandemic in a randomised trial (N = 301, among whom 158 achieved uptake). This study built and evaluated several models for predicting uptake. Putative predictors included demographic characteristics, mental health service utilization and interest, and mental health symptoms assessed before participants were provided access to the intervention. The best-performing model, a linear support vector machine model, had 70% accuracy and 0.70 area under the receiver operating characteristics curve in a held-out dataset, though these metrics were not significantly better than competitor models. Model inspection revealed that participants who reported interest in mental health treatment and lesbian, gay, bisexual, and other sexual minority participants were more likely to achieve uptake. Additionally, male participants were less likely to achieve uptake. The best-performing machine learning model achieved an acceptable level of performance in predicting uptake. Self-reported treatment interest was especially predictive of uptake. Future research should attempt to understand gender and sexual orientation differences in self-guided online mental health intervention uptake. Additionally, research should evaluate the utility of machine learning to inform targeted motivational enhancement of those less likely to achieve uptake.