Personalized Impact of Lifestyle on Type 1 Diabetes Patients: A Comprehensive Regression Analysis

While lifestyle behaviors play important roles in Type 1 Diabetes (T1D), the individualized effects of lifestyle factors on T1D patients have not been studied. In this paper, we present a regression analysis-based approach to understand the personalized impact of lifestyle factors on blood glucose (BG) in T1D patients using the OhioT1DM dataset. Our method addresses the following challenges: (1) effectively structuring the multi-modal lifestyle, insulin, and BG data and (2) modeling the data to derive personalized insights. To solve the first challenge, a patient's data are segmented into time windows based on BG-affecting events. Consequently, we can better reveal each lifestyle factor's effect because windows capture BG trends at higher granularity and contain fewer concurrent BG-affecting factors. To model the data and derive personalized insights, we first utilized variance inflation factor (VIF) and correlation analysis to eliminate and combine lifestyle features to avoid multicollinearity issues. We then trained multiple machine learning models and found multivariate linear regression (MLR) yielded the best BG prediction. We derived personalized insights about the relationship between lifestyle factors and BG using the MLR model's statistical properties, including β coefficients, P-values, and R2 values. Our results show that T1D patients differ significantly in how lifestyle factors affect their BG, indicating that personalized lifestyle interventions are necessary for T1D management.