1587-P: Beyond the Mean: Machine Learning–Driven Discovery of Behavioral Drivers and Barriers to Glycemic Control

Introduction and Objective: Despite the indisputable benefits of tight glycemic control and well defined clinical targets, real-world goal attainment remains hard to achieve. We leveraged the mySugr app database (over 5 million downloads) to measure the impact of digital engagement on glycemic outcomes. Our goal was to apply machine learning to this real-world data to automatically detect complex behavioural interactions driving success, avoiding the need to explicitly hypothesize and model these terms as required in standard linear regression. Methods: We analyzed glycemic and non-glycemic outcomes after 90 days of app use. Unlike traditional linear modeling, which assumes additive effects unless interactions are manually specified, we employed non-linear machine learning models paired with SHAP (SHapley Additive exPlanations). This methodology allowed us to automatically discover and quantify non-linear synergies between user characteristics and behaviors that drive goal achievement variability. Results: Baseline glycemic status was the strongest predictor of outcome. SHAP interaction plots revealed that the relationship between engagement and outcome was non-linear: increased digital engagement (specifically session counts, tagging, and logging) correlated with improved glycemic control, but this benefit was significantly amplified in users with a baseline blood glucose >200 mg/dL. Subgroup analysis further identified distinct behavioural drivers: for Type 1 DM, irregular meal timing (specifically dinner) hindered improvement, while for Type 2 DM, overall BG irregularity was the primary predictor of poorer outcomes. Conclusion: In summary, this work highlights the pivotal role of mobile health apps in diabetes management. Furthermore, it demonstrates the utility of explainable AI in decoding intricate, high-order interactions between behaviour and glycemic control that would likely remain hidden under the constraints of linear assumptions. Disclosure E. Cinar: Employee; Current; mySugr.