CRUISEResearchGroup/CGM-JEPA-Pretraining

CGM-JEPA Pretraining Corpus

Continuous glucose monitor (CGM) time-series corpus used for self-supervised pretraining of CGM-JEPA, X-CGM-JEPA, GluFormer, and TS2Vec encoders in the paper CGM-JEPA: Learning Consistent Continuous Glucose Monitor Representations via Predictive Self-Supervised Pretraining.

Pretraining-only corpus. For the labeled downstream-evaluation cohorts (insulin resistance and β-cell dysfunction classification), see CRUISEResearchGroup/CGM-JEPA-Downstream. For pretrained model weights, see CRUISEResearchGroup/CGM-JEPA.

Quick start

Option 1 — datasets library

from datasets import load_dataset
ds = load_dataset("CRUISEResearchGroup/CGM-JEPA-Pretraining")
# DatasetDict({
#   train: Dataset({features: ['timestamp', 'glucose_value', 'subject'],
#                   num_rows: 389365})
# })

The natural unit for SSL pretraining is a 288-step (24-hour) window per (subject, day); the code repo's data_loaders/JEPALoader performs the windowing on top of this raw flat CSV.

Option 2 — file download for the code repo's pretraining scripts

huggingface-cli download CRUISEResearchGroup/CGM-JEPA-Pretraining \
  --repo-type dataset --local-dir Dataset_Open

Then from the code repository:

# (Optional, only for X-CGM-JEPA) Precompute the Glucodensity cache once.
python -m utils.precompute_glucodensity \
  --data_path Dataset_Open/cgm_initial_cohort.csv \
  --output_path Dataset_Open/gluco_cache_stride288.pkl \
  --stride 288

# Pretrain
python -m pretrain.pretrain_x_cgm_jepa     # X-CGM-JEPA (cross-view)
python -m pretrain.pretrain_cgm_jepa       # CGM-JEPA (temporal only)
python -m pretrain.pretrain_gluformer      # GluFormer baseline
python -m pretrain.pretrain_ts2vec         # TS2Vec baseline

Files

File	Size	Purpose
cgm_initial_cohort.csv	~12 MB	Raw CGM readings (5-min sampling) — primary pretraining input

X-CGM-JEPA additionally needs a Glucodensity cache (~54 MB) which is not shipped here; it's derived deterministically from the CSV by utils/precompute_glucodensity.py in the code repository (one-time, ~few minutes on CPU).

Schema

cgm_initial_cohort.csv

Column	Type	Description
timestamp	datetime	Reading time at 5-min resolution
glucose_value	float	Glucose value in mg/dL
subject	str	Subject identifier (S01..S22 for Stanford; colas_<subject>_<day> for Colas subject-days)

• 413 unique subject identifiers (22 Stanford + 391 Colas subject-days = 206 unique Colas subjects across multiple days).
• 389,365 total readings.
• Median 288 readings per subject (= one 24-hour window at 5-min sampling); max 18,527 readings for the longest Stanford continuous-CGM record.

Cohort composition

Source	Subjects	Rows	Share
Stanford CGM Study	22	276,757	71.1 %
Colas et al. (2019)	206 (391 subject-days)	112,608	28.9 %
Total	228 unique subjects	389,365	100 %

How this corpus was built

The corpus was assembled by scripts/preprocess_dataset.py in the code repository, from two upstream sources:

1. Stanford CGM Study (Metwally et al. 2025, Nature Biomedical Engineering) — data distributed through the Metabolic_Subphenotype_Predictor repository under MIT license (files filtered_cgm_03222026.csv, filtered_ogtt_…csv, filtered_metabolic_tests.csv).
2. Colas et al. (2019) — Open-access CGM time series from Detrended Fluctuation Analysis in the prediction of type 2 diabetes mellitus in patients at risk, PLOS ONE 14(12):e0225817, distributed in the paper's Supporting Information files under CC BY 4.0.

Stanford rows were smoothed onto a 5-min grid via cubic smoothing splines (scipy.interpolate.make_smoothing_spline(lam=0.35)). Sensor "Low"/"High" markers were replaced with the empirical numeric min/max. Colas daily windows were concatenated with prefixed subject IDs (colas_<subject>_<day>) to disambiguate them from Stanford subjects.

Intended use

• Self-supervised pretraining of CGM encoders (masked-patch prediction, contrastive, autoencoding objectives).
• Comparison of CGM representation learning methods.

License & attribution

This corpus is released under CC BY 4.0 to match the licenses of its two upstream sources:

• Stanford CGM Study data — released through Metabolic_Subphenotype_Predictor under the MIT license, accompanying Metwally et al. 2025 (Nature Biomedical Engineering).
• Colas et al. (2019) — CGM time series from the PLOS ONE Supporting Information files, licensed under CC BY 4.0.

When using this dataset, please cite all three sources: the Stanford CGM Study (via Metabolic_Subphenotype_Predictor), Colas et al. 2019, and our CGM-JEPA paper.

Citation

Citation block to be filled once the CGM-JEPA paper has a stable venue / arXiv link.

Code repository

https://github.com/cruiseresearchgroup/CGM-JEPA