Research map / Time series / Foundation models for time series

GlucoFM-Bench: Benchmarking Time-Series Foundation Models for Blood Glucose Forecasting

NeurIPS 2026 · Time series › Foundation models for time series · approach: Time series foundation models · 0 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.

Benchmarks four pre-trained time-series foundation models, two LLM-based forecasters, and two supervised DL models for CGM glucose forecasting across 15 datasets under zero-, few-, and full-shot protocols.

Problem

No prior benchmark evaluates general-purpose time-series foundation models (TSFMs) for blood glucose forecasting across diverse diabetes populations and data regimes.

Approach

A standardized benchmark that pairs zero-, few-, and full-shot protocols with glycemic-range and diabetes-cohort stratified analyses, using RMSE plus Surveillance Error Grid no-risk ratios, across 12 open-access and 3 controlled-access CGM datasets.

Key results

• Zero-shot TimesFM 2.5 achieves overall RMSE 18.75 mg/dL vs 22.25 for persistence, within about 4.3% of the best full-shot LSTM (17.97 mg/dL).
• Few-shot Chronos-2 achieves the best overall RMSE of 18.95 mg/dL and SEG no-risk ratio 88.30%.
• Full-shot LSTM achieves the lowest overall RMSE of 17.97 mg/dL, ahead of Chronos-2 at 18.79 mg/dL.

Surprising findings

• Moirai 2.0 and TimesFM 2.5 do not improve consistently with full-shot fine-tuning, and in some cases degrade relative to few-shot.
• Larger TSFM model size does not necessarily yield better full-shot glucose forecasting performance.

Limitations

• Univariate CGM only; no covariates such as carbohydrates or insulin, and few TSFMs support multivariate inputs.
• Does not exhaustively cover newer TSFM architectures or per-model tuning strategies.
• Retrospective benchmark, not validated for clinical deployment.

Experimental setup

Datasets: Hall2018 D1NAMO Colas2019 BIG IDEAS Lab ShanghaiT1DM ShanghaiT2DM UCHTT1DM HUPA-UCM CGMacros T1DM-UOM Bris-T1D AZT1D

Benchmarks: GlucoFM-Bench

Metrics: RMSE MAE SEG no-risk ratio CEG zone A ratio PI80 coverage

Baselines: Persistence LSTM GPFormer TimeLLM CALF

Builds on: Time-LLM CALF Chronos TimesFM Moirai

Closest papers by meaning

• GlucoBench: Curated List of Continuous Glucose Monitoring Datasets with Prediction Benchmarks (ICLR 2024 · similarity 0.89)
• Hybrid$^2$ Neural ODE Causal Modeling and an Application to Glycemic Response (ICML 2024 · similarity 0.84)
• FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs (NeurIPS 2026 · similarity 0.83)
• Adaptive Shrinkage Estimation for Personalized Deep Kernel Regression in Modeling Brain Trajectories (ICLR 2025 · similarity 0.82)
• Specialized Foundation Models Struggle to Beat Supervised Baselines (ICLR 2025 · similarity 0.82)
• NutriBench: A Dataset for Evaluating Large Language Models in Nutrition Estimation from Meal Descriptions (ICLR 2025 · similarity 0.82)