GLM-DM: language model boosted neural networks for HbA1c trend prediction in diabetes mellitus

Predict Hemoglobin A1c (HbA1c) trends, a key metric in diabetes mellitus (DM) management, using readily available patient variables and language models (LMs).

We propose GLM (Language Model Boosted Neural Network) -DM, which leverages data augmentation and language model-driven feature encoding to predict HbA1c trends using easily accessible patient-level variables. Our model captures complex relationships among patient characteristics and enhances predictive performance through Generative Adversarial Networks (GANs) for synthetic data augmentation and LMs for feature embedding. By transforming patient profiles into rich latent representations, our approach enables a more comprehensive analysis of how patient-level variables correlate with HbA1c trends over time.

Using clinical data from 257 DM patients, GLM-DM achieves 70.2% accuracy of HbA1c trend prediction, outperforming classic classifiers and transformer-based models. Ablation studies confirm the effectiveness of GAN-based augmentation and LM-driven embedding. Our model achieves 68.2% prediction accuracy for Type 1 DM and 72.7% for Type 2 DM.

Proposed approach learns the underlying complex function of HbA1c using clinical variables easily available at the patient visit and leveraging the power of LMs to accurately predict the trend of HbA1c in a period. The model can enhance patient advisories for daily diabetes management without the need for continuous glucose monitoring.