AACE 2026: How machine learning models predict hemoglobin A1c response 

Machine learning models using routine clinical data outperformed logistic regression in predicting glycemic response over 6 to 18 months in people with type 2 diabetes, according to research presented at the 2026 AACE Annual Meeting in Las Vegas.

In a real-world cohort from a tertiary diabetes center in India, researchers evaluated whether machine learning (ML) models could identify patients likely to achieve meaningful hemoglobin A1c (HbA1c) improvement or target control. The primary outcome was a composite “responder” outcome, defined as a reduction in HbA1c of at least 1% or as achieving an HbA1c level below 7% between 6 and 18 months. 

The analysis used de-identified electronic medical records to construct patient-level datasets incorporating demographics, anthropometric measures, glycemic and renal indices, and medication classes, including metformin, insulin, DPP-4 inhibitors, SGLT2 inhibitors, and sulfonylureas. A subset analysis included approximately 3 months of HbA1c or fasting glucose values. 

Models evaluated included logistic regression and two tree-based approaches—random forest and extreme gradient boosting (XGBoost)—trained on a stratified 75/25 split. 

Logistic regression showed modest discrimination, with an area under the curve (AUC) of 0.53. Tree-based models performed better, with AUCs of 0.69 for random forest and 0.67 for XGBoost. Calibration was acceptable for both, with Brier scores from 0.20 to 0.21. 

Feature importance analyses identified baseline HbA1c and 3 month glycemic measures as the strongest predictors of response, followed by body mass index or weight, renal markers, and selected medication indicators. 

In secondary analyses, regression models predicting continuous HbA1c change showed modest performance, with mean absolute error of about 2.6 percentage points and R² values ranging from 0.24 to 0.26, indicating substantial unexplained interindividual variability in glycemic trajectories. 

“Baseline glycemic burden and early metabolic trajectory were the strongest determinants of success,” the researchers wrote. They added that although the models are not yet suitable as stand-alone clinical tools, they may support decision-making by identifying patients at high risk for inadequate control and prompting earlier treatment intensification. 

Funding sources and conflicts of interest were not reported. 

(Editor's Note: These findings are from a conference presentation on an abstract and should be considered preliminary.)