Personalized Forecasting of Glycemic Control in Type 1 and 2 Diabetes Using Foundational AI and Machine Learning Models

TL;DR

This study evaluates modern machine learning models for predicting week-ahead glucose metrics in diabetes, showing that while predictions are feasible with reasonable accuracy, low-prevalence hypoglycemia remains challenging, and advanced models require significant computational resources.

Contribution

It compares the performance of several recent tabular learning models, including foundation models, for diabetes glycemic forecasting, highlighting their relative accuracy and computational costs.

Findings

AutoGluon and tabPFN outperform XGBoost in some targets.

Models achieve 8.5-16.5% MARD for T1DM metrics.

Low-prevalence hypoglycemia prediction remains difficult.

Abstract

Background: Accurate week-ahead forecasts of continuous glucose monitoring (CGM) derived metrics could enable proactive diabetes management, but relative performance of modern tabular learning approaches is incompletely defined. Methods: We trained and internally validated four regression models (CatBoost, XGBoost, AutoGluon, tabPFN) to predict six weekahead CGM metrics (TIR, TITR, TAR, TBR, CV, MAGE, and related quantiles) using 4,622 case-weeks from two cohorts (T1DM n=3,389; T2DM n=1,233). Performance was assessed with mean absolute error (MAE) and mean absolute relative difference (MARD); quantile classification was summarized via confusion-matrix heatmaps. Results: Across T1DM and T2DM, all models produced broadly comparable performance for most targets. For T1DM, MARD for TIR, TITR, TAR and MAGE ranged 8.5 to 16.5% while TBR showed large MARD (mean ~48%) despite low MAE. AutoGluon and tabPFN showed lower MAE than XGBoost for several targets (e.g., TITR: p<0.01; TAR/TBR: p<0.05 to 0.01). For T2DM MARD ranged 7.8 to 23.9% and TBR relative error was ~78%; tabPFN outperformed other models for TIR (p<0.01), and AutoGluon/ tabPFN outperformed CatBoost/XGBoost on TAR (p<0.05). Inference time per 1,000 cases varied markedly (PFN 699 s; AG 2.7 s; CatBoost 0.04 s, XGBoost 0.04 s). Conclusions: Week-ahead CGM metrics are predictable with reasonable accuracy using modern tabular models, but low-prevalence hypoglycemia remains difficult to predict in relative terms. Advanced AutoML and foundation models yield modest accuracy gains at substantially higher computational cost.