Achieving and maintaining optimal glycemic control remains fundamental to preventing microvascular and macrovascular complications in diabetes mellitus. Early identification of patients likely to have poor long-term control could guide more proactive, individualized management. This study aimed to develop a clinically interpretable machine learning model to predict glycemic control status three years after diagnosis and to identify major determinants influencing glycemic outcomes in real-world settings.

Data were retrieved from e-Nabız, the national electronic health record (EHR) system in Turkey, covering over 94% of Istanbul’s population. Patients newly diagnosed with diabetes mellitus in 2017 were included if they had initial HbA1c, lipid profile, serum creatinine, and at least four HbA1c measurements during follow-up. Glycemic control was defined as the last two HbA1c values below 7% (under control) versus ≥7% (poor control). A total of 105 baseline and first-year variables were analyzed, including demographics, HbA1c levels, lipid parameters, comorbidities (44 ICD-10 categories), and medication exposure (antidiabetic and other drugs classified by ATC codes). After data cleaning and exclusion of incomplete records, a Random Forest algorithm was trained and validated using five-fold cross-validation. Model performance and feature importance were assessed to ensure clinical interpretability.

The model achieved a mean accuracy of 0.884 (±0.02) in predicting glycemic control at three years. The most influential predictors were baseline HbA1c, early HbA1c change within the first year, age, LDL cholesterol, serum creatinine, and cumulative doses of insulin and oral hypoglycemic agents (particularly metformin and glimepiride). Comorbidities such as obesity, hypertension, nephropathy, and depression were also associated with poor glycemic control, underscoring their clinical relevance.

Our findings demonstrate that routinely collected clinical and laboratory data can effectively predict long-term glycemic control in patients with diabetes mellitus. The identified predictors align with established metabolic and behavioral risk factors, supporting the model’s clinical credibility. Integrating such predictive tools into endocrinology practice may facilitate early risk stratification, individualized treatment adjustment, and improved chronic disease management.