A Comprehensive Benchmark of Time-Series Foundation Models for Blood Glucose Forecasting

This repository contains the official source code for the paper "A comprehensive benchmark of time-series foundation models for blood glucose forecasting". We provide a framework to evaluate ten Time-Series Foundation Models (TSFMs) across zero-shot, fine-tuned, and high glycemic variability scenarios. This codebase facilitates research on applying TSFMs to address the cold-start problem in personalized medical time-series forecasting.

Evaluation Overview

All models are evaluated under three scenarios: zero-shot forecasting, full fine-tuning, and generalization evaluation on a high glycemic variability (GV) dataset. The table below summarizes model participation:

✅ = applicable; NA = not applicable due to training, API, or protocol constraints.

Repository Structure

tsfm-bgf/
├── TSFMs/          # Time-series foundation models (10 models)
├── Tsline/         # Deep learning & statistical baseline models (8 models)
└── README.md

TSFMs/

Contains evaluation and fine-tuning scripts for 10 TSFMs. Each model is organized into zero-shot/ and fine-tuned/ (if applicable) subdirectories. Models retain their original inference pipelines — no unified wrapper is imposed. See TSFMs/README.md for upstream repo versions and per-model usage.

Tsline/

Contains 8 baseline models (6 deep learning + 2 statistical) built on a unified wrapper architecture. All models share a consistent BaseModelWrapper interface for training and evaluation. Model hyperparameters can be found in the BASE_CONFIG dictionary within each wrapper script. See Tsline/README.md for details.

Getting Started

Data Preparation

This benchmark uses publicly available continuous glucose monitoring (CGM) datasets. To reproduce the experiments:

1. Obtain the raw CGM data from the public dataset sources referenced in the paper.
2. Segment each subject's recordings into continuous CSV files named {subjectID}_{fileID}.csv. Each CSV must contain a GlucoseValue column.
3. Organize the files into the following directory structure:

glucose_data/
├── op_split2/          # OP dataset
│   ├── Train/
│   ├── Val/
│   └── Test/
├── re_split2/          # RE dataset
│   ├── Train/
│   ├── Val/
│   └── Test/
└── lg_split2/          # LG (high GV) dataset
    ├── Train/
    ├── Val/
    └── Test/

Running Evaluations

Navigate to the target model's directory and run the corresponding evaluation script directly:

# Example: TSFMs zero-shot evaluation
python TSFMs/Sundial/zero-shot/sundial_eval.py

# Example: Tsline baseline training & evaluation
python Tsline/PatchTST/patchtst_wrapper.py

Refer to each subdirectory's README.md for model-specific instructions, CLI arguments, and fine-tuning workflows.