Introduction and Objective: HbA1c outcomes in diabetes vary widely due to differences in clinical characteristics and real-world adherence. This study aimed to develop and validate machine-learning models to predict HbA1c at 120 days using demographic, clinical, and behavioral adherence data from a digital diabetes platform. Methods: Data from 507 participants enrolled in the Lillia Care digital diabetes management platform were analyzed. The cohort had a mean age of 47 years, mean BMI of 26.7 kg/m², and mean diabetes duration of 6.2 years. Mean HbA1c improved from 8.74% at baseline to 7.31% at follow-up (mean reduction: 1.43%). Predictors included demographic and clinical variables, comorbidities, lifestyle factors, and adherence scores for nutrition, exercise, medication use, and platform engagement. Nutrition and exercise adherence varied widely, while medication adherence was consistently high (mean 96.3%). Multiple regression models were evaluated using cross-validation and held-out testing, with performance assessed using mean absolute error (MAE) and root mean squared error (RMSE). Results: Among all evaluated models, a Gradient Boosting regressor incorporating adherence metrics demonstrated the strongest predictive performance of HbA1c, achieving a cross-validation MAE of 0.62, test MAE of 0.63, and RMSE of 0.98. Tree-based ensemble models consistently outperformed linear, kernel-based, and neural-network approaches, highlighting nonlinear relationships between adherence behaviours and glycemic outcomes. Conclusion: Machine-learning models leveraging real-world adherence data can accurately forecast short-term HbA1c outcomes and support personalized diabetes care. Beyond risk stratification, the proposed model can serve as a motivational and educational tool by visually demonstrating adherence-driven glycemic outcomes, reinforcing the impact of improved nutrition, physical activity, and medication adherence on HbA1c trajectories and encouraging sustained behavior change. Disclosure P. Nerkar: None. A.K. Joshi: None. A. Essat: None. M. Mundra: None. S. Chakrabarty: None. P. Samant: None. G. Talbot-Lachance: None. R. Barai: None.
NERKAR et al. (Fri,) studied this question.