Mining Association Rules From a Multimodal Dataset of a Digital Therapeutics Application for Sleep Improvement Through a Healthy Lifestyle : Quantitative Study

Abstract

Background: The demand for sleep interventions is high and steadily growing. Digital therapeutics (DTx) can help individuals improve their sleep remotely, over an extended period, and with less effort from medical professionals. Obstructive sleep apnea (OSA), one of the most prevalent and consequential sleep disorders, can be treated with health-supporting behavior changes, such as physical exercise and weight loss, and, therefore, acts as a promising application for DTx. Objective: The study aimed to analyze a digital intervention from both medical and technological perspectives by moving beyond clinical markers and exploring more deeply how the DTx application was used. This study aimed to propose a novel way in which association rules can function as an exploratory tool to analyze the sleep, behavior, and engagement of participants with the DTx application on a day-to-day level. Methods: A lifestyle intervention study (N=192) targeted at adults with mild-to-moderate OSA aimed to reduce their OSA severity using a DTx application and an exercise program over a study period of 12 weeks. The participants’ OSA severity was assessed through polysomnography at the beginning and at the end of the study period, and the participants tracked their sleep with a digital sleep diary and a smartwatch over the course of the entire study. The DTx application provided data on when and how the participants pursued the proposed lifestyle interventions. These heterogeneous data sources were combined into one multimodal dataset, which was explored through descriptive statistics. Ultimately, the data were turned into a transaction-based format, and association rules were derived using the Apriori algorithm. Results: Analyzing the participants’ interaction with the application revealed the lifestyle interventions they pursued and how their behavior and sleep patterns changed over time. The Apriori algorithm generated a set of association rules with lift and confidence scores that were significantly higher than those for the co-occurrence of items through random chance. The rules show co-occurrence of missions and items from the sleep diary, as well as items derived from the watch measurements. Conclusions: The study showed the richness of the various data sources provided by a digital intervention using wearables and how they can be used to get an in-depth understanding of the study. The generated association rules showed the presence of significant co-occurrences across the different data modalities and highlighted their effectiveness as an exploratory tool for multimodal health data.