Responsible data selection method for algorithmic personalization of health apps: a case study on promoting mental health

Abstract

Digital technologies are on the rise to promote health. To improve the engagement and effectiveness of these technologies, there is a growing interest in algorithmic personalization. However, the user input data for these algorithms (e.g., data from wearables or self-reported data) can come with ethical and regulatory implications. Despite a growing amount of theoretical work, there is no practical precedent on how to consider these implications in the development of personalization algorithms. Therefore, our work aims to tackle this challenge by proposing a stepwise method for Responsible Data Selection (ReDS) for algorithmic personalization of mHealth. The ReDs method acts from a duty of care and promotes an active search for ethically less risky data. We demonstrate the six-step method through a real-world use case on an mHealth app promoting adolescents' mental well-being, using a dataset of 1181 adolescents (5199 interactions) who received coping strategy challenges based on cognitive behavioral therapy. First, we identified the personalization objective in the case study (step 1). The objective was to personalize the type of challenge to promote adherence while diversifying the coping strategy types within the completed challenges. Next, we identified the emotional state of the adolescent and prior completion rates as promising input data (step 2). However, personal emotion data can be considered sensitive, personal, and private, implying ethical implications (step 3). As a potential alternative, tiredness data can be perceived as less sensitive to share and collect (step 4). Subsequently, we analyzed the utility of all data features (step 5) using evaluative simulations with reinforcement learning models. This revealed that solely using the completion rates of the previous day could already benefit the personalization objective and that adding emotion data or tiredness data could similarly further increase the performance of the personalization algorithm. When determining the utility-risk trade-off (step 6), we conclude that tiredness data can be used as an alternative for emotion data if risk mitigation strategies are deployed. Through this case study, we demonstrate the practical utility of the ReDS method. We hope that our work will inspire future developers of personalization algorithms to explicitly incorporate ethical considerations in the algorithm development process.