AACE 2026: How machine learning models predict hemoglobin A1c response 

Machine Learning Models Predict HbA1c Response in Type 2 Diabetes

Overview

Machine learning models using routine clinical data outperformed logistic regression in predicting meaningful hemoglobin A1c (HbA1c) improvement or target control over 6 to 18 months in patients with type 2 diabetes. Tree-based models such as random forest and XGBoost showed better discrimination with AUCs around 0.67-0.69 compared to logistic regression's AUC of 0.53.

Background

Predicting glycemic response in type 2 diabetes is critical for optimizing treatment strategies. Traditional logistic regression models have shown limited accuracy in forecasting HbA1c outcomes. Machine learning approaches may leverage complex clinical data to improve prediction of patients likely to achieve significant HbA1c reduction or target control. This study evaluated ML models using real-world electronic medical records from a tertiary diabetes center in India.

Data Highlights

Key Findings

• Machine learning models (random forest and XGBoost) outperformed logistic regression in predicting HbA1c response with AUCs of 0.69 and 0.67 versus 0.53.
• Calibration of tree-based models was acceptable with Brier scores between 0.20 and 0.21.
• Baseline HbA1c and 3-month glycemic measures were the strongest predictors of response.
• Other important predictors included body mass index or weight, renal markers, and medication classes.
• Regression models predicting continuous HbA1c change had modest performance with mean absolute error ~2.6% and R² of 0.24-0.26, indicating substantial unexplained variability.
• Models may help identify patients at high risk for inadequate glycemic control to prompt earlier treatment intensification.

Clinical Implications

Machine learning models can enhance prediction of glycemic response beyond traditional logistic regression, potentially aiding clinicians in identifying patients unlikely to achieve HbA1c targets. Early identification of suboptimal responders may facilitate timely treatment adjustments. However, these models are not yet ready for standalone clinical use and should complement clinical judgment.

Conclusion

Machine learning approaches using routine clinical data show promise in predicting HbA1c response in type 2 diabetes, with baseline glycemic burden and early metabolic trajectory as key determinants. Further validation is needed before clinical implementation.

References

1. AACE 2026 Annual Meeting -- Machine learning models predict hemoglobin A1c response