Understanding the Relationship Between Mood Symptoms and Mobile App Engagement Among Patients With Breast Cancer Using Machine Learning: Case Study (Preprint)

Abstract

BACKGROUND Health interventions delivered via smart devices are increasingly being used to address mental health challenges associated with cancer treatment. Engagement with mobile interventions has been associated with treatment success, yet the relationship between mood and engagement among cancer patients remains poorly understood. One reason is the lack of a data-driven process for analyzing mood and app engagement data for cancer patients. OBJECTIVE The purpose of this study is to provide a step-by-step process for using app engagement metrics to predict continuously assessed mood outcomes in breast cancer patients. We describe the steps of data preprocessing, feature extraction, and data modeling and prediction. We then apply this process as a case study to data collected from breast cancer patients who engaged with a mobile mental health app intervention (IntelliCare) over 7-weeks. We compare engagement patterns over time (e.g., frequency, days of use) between high- and low-anxious and high- and low-depressed participants. We then use a Linear Mixed Model to identify significant effects and evaluate the performance of Random Forest and XGBoost classifiers in predicting weekly state mood from baseline affect and engagement features. METHODS We describe the steps of data preprocessing, feature extraction, and data modeling and prediction. We then apply this process as a case study to data collected from breast cancer patients who engaged with a mobile mental health app intervention (IntelliCare) over 7-weeks. We compare engagement patterns over time (e.g., frequency, days of use) between high- and low-anxious and high- and low-depressed participants. We then use a Linear Mixed Model to identify significant effects and evaluate the performance of Random Forest and XGBoost classifiers in predicting weekly state mood from baseline affect and engagement features. RESULTS We observed differences in engagement patterns between high- and low-anxious and depressed participants. Linear Mixed Model results varied by the featureset; these results revealed weak effects for several features of engagement, including duration-based metrics and frequency. Accuracy of predicting state mood varied according to classifier and featureset. The XGBoost classifier achieved the highest accuracy for state anxiety prediction when self-report scores and engagement features were used for only the most highly-used apps. The Random Forest classifier achieved the highest accuracy for state depression prediction when self-report scores and engagement features were used from all apps. CONCLUSIONS The results from the case study support the feasibility and potential of our analytic process for understanding the relationship between app engagement and mood outcomes in breast cancer patients. The ability to leverage both self-report and engagement features to predict state mood during an intervention could be used to enhance decision-making for researchers and clinicians, as well as assist in developing more personalized interventions for breast cancer patients.