Research map / Probabilistic and scientific modeling / Molecules and proteins

GlucoBench: Curated List of Continuous Glucose Monitoring Datasets with Prediction Benchmarks

ICLR 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Spatiotemporal forecasting and imputation · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Presents GlucoBench, a curated set of five public CGM datasets with a standardized preprocessing pipeline, two prediction tasks (accuracy and uncertainty), and baseline benchmarks of shallow, deep, and hybrid models with analysis of performance factors.

Problem

CGM-based glucose trajectory prediction is evaluated on small private datasets with no standardized tasks or comparisons, which hinders reproducibility and clinical adoption.

Approach

Consolidates public CGM datasets with a uniform preprocessing pipeline, defines two standardized tasks (predictive accuracy via RMSE/MAE and uncertainty quantification via likelihood and calibration error), and benchmarks a common set of shallow, deep, and hybrid models under a fixed out-of-the-box tuning protocol.

Key results

• On predictive accuracy, ARIMA or linear regression achieve the lowest error on most datasets, with Latent ODE best on Hall and Transformer best on Weinstock.
• Gluformer achieves the best likelihood on all five datasets, and the best calibration on all but two (TFT best on Colas and Weinstock per the tables).
• Deep models generalize better to out-of-distribution patients than simple baselines (ARIMA, linear regression), though performance generally degrades OD vs ID.

Measured effects reported in the paper

Surprising findings

• Simple ARIMA and linear regression beat deep models on most datasets for accuracy.
• Adding covariates hurt performance on several datasets, likely due to overfitting on small data.

Limitations

• Out-of-the-box baselines without model-specific tuning, pre-training, or augmentation, so results may not reflect best achievable performance.
• Datasets are small and mostly Type 1; few Type 2 datasets and only one has time-varying covariates.
• Deep models underperform on several datasets due to limited observations per patient.

Experimental setup

Datasets: Broll Colas Dubosson Hall Weinstock

Benchmarks: Broll Colas Dubosson Hall Weinstock

Metrics: RMSE MAE log-likelihood calibration error

Baselines: ARIMA Linear regression XGBoost Transformer Temporal Fusion Transformer NHiTS Latent ODE Gluformer

Builds on: Darts NHiTS Temporal Fusion Transformer Latent ODE Gluformer

Closest papers by meaning

• GlucoFM-Bench: Benchmarking Time-Series Foundation Models for Blood Glucose Forecasting (NeurIPS 2026 · similarity 0.89)
• Hybrid$^2$ Neural ODE Causal Modeling and an Application to Glycemic Response (ICML 2024 · similarity 0.80)
• NutriBench: A Dataset for Evaluating Large Language Models in Nutrition Estimation from Meal Descriptions (ICLR 2025 · similarity 0.79)
• Adaptive Shrinkage Estimation for Personalized Deep Kernel Regression in Modeling Brain Trajectories (ICLR 2025 · similarity 0.79)
• Uncertainty-Calibrated Prediction of Randomly-Timed Biomarker Trajectories with Conformal Bands (NeurIPS 2025 · similarity 0.78)
• Modeling All-Atom Glycan Structures via Hierarchical Message Passing and Multi-Scale Pre-training (ICML 2025 · similarity 0.78)