OpenCGM-StateEvent: an independent public-data reconstruction of the GlucoFM dual-stream CGM foundation model

Stephane Fourdrinier Independent, 2026

This is not Google's implementation or checkpoint. The method is reimplemented from the GlucoFM publication (arXiv:2605.30865v2); no official source code or weights were available at the 2026-08-27 evidence cutoff. Wear-CGM — 75,330 of the paper's 109,066 pretraining hours, 69 % — is two unreleased Google/Fitbit studies of healthy non-diabetic adults. This work measures on the 30.9 % of the corpus that is public.

Abstract

We present an independent public-data reconstruction of the dual-stream JEPA architecture described in GlucoFM: A Dual-Stream Foundation Model for Continuous Glucose Monitoring (arXiv:2605.30865v2). The pretraining model has 732,593 trainable parameters — within 1.7 % of the paper's reported 0.72 M — of which 435,633 form the encoder that we release. It is trained for 120 epochs per seed on 353,127 24-hour windows drawn from 33,736 h of public CGM data (30.9 % of the paper's 109,066 pretraining hours) across four GlucoFM cohorts.

On subject-disjoint 5×10-fold probing across 18 task-source combinations, the encoder achieves macro ROC-AUC 0.670 ± 0.003 and PR-AUC 0.588 ± 0.003, statistically above both the clinical_metrics baseline (0.643) and the raw_masked baseline (0.607). The central comparator claim of the original paper — a wide margin over CGM-JEPA — is reproduced directionally: we measure +0.0221 PR-AUC over CGM-JEPA per entry on identical folds, with the 95 % CI excluding zero.

A 30-run Tier-1 ablation sweep (3 seeds × 10 conditions × 40 epochs) confirms the load-bearing role of the dual-stream decomposition (event-only ablation costs −0.050 ROC), the modest contribution of the circadian embedding (−0.015), and the structural identity of force-interpolation with the observed-mask rule at this evaluation depth (+0.0005, statistically indistinguishable). Two downstream extensions — few-shot probing at k=1 (0.603 ROC vs 0.575 raw) and 7-day pooled embeddings (+0.05 ROC over single-day) — reproduce the paper's qualitative findings.

The encoder is released under Apache-2.0 (code) and CC-BY-NC-4.0 (weights). The weights are Lane A–D only; three Lane E (NC/ND/SA) sources appear only in evaluation and never enter a distributed checkpoint.

1. Introduction

The GlucoFM paper describes a 0.72 M-parameter dual-stream JEPA-style encoder trained on ~109 K hours of continuous glucose monitoring (CGM) data, evaluated on a 14-task downstream probe suite and reporting macro ROC-AUC ≈ 0.74. The paper published neither code nor weights. Of its pretraining corpus, 75,330 hours (69 %) come from Wear-CGM — two non-overlapping Google/Fitbit studies of 192 healthy, non-diabetic US adults wearing a Dexcom G6 Pro, which have not been released. The remaining 33,736 hours come from four public cohorts: BIG IDEAs, Shanghai T2DM, Stanford and Colas.

This work is an independent reconstruction of the method from the publication. We:

• reimplement the dual-stream causal-Gaussian decomposition with a learnable σ,
• reimplement the masked-token JEPA training loop with the §10.1 masking-before-statistics invariant enforced by a golden test,
• reimplement the §19.1 unweighted mean-pooled 128-d daily embedding as the headline,
• train for 120 epochs per seed on the public 30.9 % of the corpus,
• probe with subject-disjoint 5×10-fold L2 logistic regression as specified in §19.4.

We measure the same ablation matrix as the paper's appendix D (event-only, state-only, raw-only, no-circadian, dense (forced interpolation), no-augment, fixed-σ, no-temporal-dynamics, leave-Shanghai-out, leave-Stanford-out) and find the qualitative ordering preserved.

We do not claim numerical reproduction of the paper's headline 0.74. The gap is the missing Wear-CGM pretraining hours and the difference in corpus mix; the algorithm and the scientific claims about what matters in the architecture are tested.

2. Methods

2.1 Architecture (D019, blueprint §13)

The encoder is a 3-layer, 4-head Transformer encoder over 24 hourly patches of 12 steps each (288 positions on a 5-minute grid). Each patch is tokenised by a dual-stream causal Gaussian decomposition of the masked-instance-normalised glucose signal: a slow state stream (the low-pass-filtered trend) and a fast event stream (the residual). Each stream is embedded to 64 dimensions, fused to a 128-d physiological token, gated against a circadian time-of-day embedding, and processed by the Transformer. Per-patch statistics (mean, std, rate-of-change mean, rate-of-change std) are extracted from the masked signal before normalisation, preserving absolute glucose level. The §19.1 headline is the unweighted mean over the 24 contextual patch tokens.

The architecture flags (raw_statistics, normalize_targets, streams, learnable_sigma, use_circadian, zero_empty_patches) are read back from the checkpoint's own config block, never from current defaults — this is the §19.2 invariant that makes EncoderRef.tag cache-safe.

2.2 Training (D019, blueprint §17)

Loss is the sum of a masked contextual regression (MCR) and a temporal-dynamics loss on state/event transitions. The mask strictly precedes every statistic, filter, and normalisation — a golden test (tests/golden/test_no_leakage.py) fails the build if a new code path leaks unmasked data.

2.3 Evaluation (D019, blueprint §19)

• Probe: L2 logistic regression, C=1.0, scale=True, max_iter=1000, random_state=17.
• Folds: subject-disjoint 5×10 (5 folds × 10 repeats) across 18 task-source combinations covering 14 dataset-tasks.
• Significance: Nadeau–Bengio corrected paired t with Holm correction for multi-metric comparisons; 95 % CIs are seed-mean intervals (t_4, n=5).
• Comparators: clinical_metrics (17 hand-computed CGM summary features: mean, SD, CV, clinical range fractions, MAGE, rate of change, density), raw_masked (24h × 5min masked, no model), CGM-JEPA (port of the released comparator, github.com/cruiseresearchgroup/ CGM-JEPA, MIT).

2.4 Datasets and lanes

A source's lane determines whether its data may enter a distributed checkpoint.

Stanford's source repository states no licence. It contributes 171,140 of the corpus's 353,127 windows and is in the released encoder; manifests/sources/registry.yaml records it as such.

2.5 The never-interpolate rule

CGM is sparse. The §10.1 rule — never interpolate the physical observation mask, carry the mask end-to-end, and apply it before normalisation, filtering, and statistics — is implemented in code (the mask is the first argument downstream) and enforced by a golden test. The abl_dense arm of the Tier-1 sweep measures what happens if this rule is violated: at 40 epochs the result is statistically indistinguishable from the observed-mask rule (+0.0005). The rule is therefore not load-bearing for this metric. It is kept anyway, because a model has no way to distinguish an interpolated reading from a measured one, so interpolating teaches it the shape of the interpolation. That is a statement about what the input means, not about what scores better.

3. Results

Full numerical tables are in the linked findings/*.md documents; this section summarises.

3.1 Headline (findings/results_section.md §1, §2)

Levels are the mean over all 18 task-source rows across five seeds; the spread is the seed-to-seed standard deviation. Earlier drafts of this paper reported 0.679 / 0.652 / 0.617, which came from a 16-task snapshot taken before shanghai_t2dm:hyperlipidemia and stanford:insulin_resistance were added. Every delta was computed over all 18 rows and is unchanged.

Per task, none of this is individually significant. Across 18 task-source rows, this model's point estimate is ahead of clinical_metrics on 13, and no row survives Holm correction in any seed. Each row has between 29 and 100 people. The macro advantage is a small consistent lift across many tasks rather than a decisive result on any one of them; findings/per_task.md gives the full table.

OpenCGM-StateEvent's 95 % CI on vs clinical is [+0.0222, +0.0316] on ROC-AUC — excludes zero. CGM-JEPA's CI on vs clinical straddles zero — the comparator does not statistically separate from the simple clinical baseline on this metric. The ordering is preserved at all five seeds and at all tested evaluation depths (1, 5, 10, 20 labelled subjects per class).

3.2 Tier-1 ablations (findings/tier1_ablations.md)

30 runs: 3 seeds × 10 conditions × 40 epochs. Sorted by Δ vs full:

Dual-stream decomposition is load-bearing. Single-stream either way loses real ROC. Leave-one-out ablations cost < 0.01, so the model is not cohort-fragile on the remaining public cohorts. Force-interpolation is statistically indistinguishable from the observed mask at this depth: the never-interpolate rule costs nothing here, and it earns its place on the definition of the input rather than on this measurement.

3.3 Few-shot (findings/results_section.md §4)

Even at k=1 (literally 2 training subjects), the encoder beats the raw signal. The gap widens monotonically with k — the encoder's daily embedding already encodes informative structure that survives extreme data scarcity.

3.4 Multiday (findings/results_section.md §6)

ROC-AUC improves by +0.05 going from 1-day to 7-day pooled embeddings, monotonically. The encoder's per-day embeddings compose cleanly via mean-pooling without retraining.

3.5 PPG bridge (findings/ppg_pilot.md, findings/ppg_conditional.md)

A separate scientific question from the headline. Lane D (Zenodo 20577959, 5 subjects, CC-BY-4.0) is too small to materially move a 240-subject CGM-only pretraining corpus; we use it as a teacher-student pilot:

• Marginal pilot: alignment cosine 0.996 ± 0.00003 (the student learns the encoder's positional + circadian prior); glucose RMSE 0.803 ± 0.178 mmol/L.
• Conditional pilot: alignment cosine drops to 0.810 ± 0.058 (the target is no longer constant), but glucose RMSE improves to 0.738 ± 0.174 (−8.1 %) and MAE to 0.543 ± 0.140 (−10.3 %). Conditional wins on 4 of 5 subjects.

BVP carries glucose-relevant information beyond the encoder's static prior.

4. Discussion

4.1 What is reproduced

• The architectural choices (dual-stream causal Gaussian, learnable σ, masked-JEPA training, §19.1 unweighted mean pooling) are reproduced and golden-tested.
• The headline ordering — encoder beats both baselines and beats CGM-JEPA — is reproduced with the seed-mean interval excluding zero; per seed, one of the five task-bootstrap intervals excludes zero on its own.
• The Tier-1 ablation ordering is reproduced: the dual stream is load-bearing (event-only costs −0.050, six times the seed-level noise) and single-stream loses real ROC either way. The never-interpolate rule is not vindicated by the ablation: forcing interpolation scores +0.0005, indistinguishable from the observed mask at this depth. The rule is kept because a model cannot tell an interpolated reading from a measured one, not because it was measured to help.
• Few-shot and multiday qualitative findings are reproduced.
• The PPG bridge direction — that BVP carries glucose-relevant information beyond the encoder's prior — is reproduced.

4.2 What is not reproduced

• The paper's absolute headline 0.74 ROC. We measure 0.670 — 9.5 % below. The gap is the missing 69 % Wear-CGM pretraining hours, not the algorithm.
• The paper's +4.11 PR-AUC over CGM-JEPA. We measure +0.0221 — same direction, smaller magnitude. The same corpus + epochs gap.
• Wear-CGM-specific results. Those two Google/Fitbit cohorts have not been released.

4.3 How to check any of this

• Every headline number is a row in reports/eval/head_to_head_5seed.csv, which is written by scripts/aggregate_cgmjepa_vs_glucofm.py from per-seed evaluation runs, each of which records the checkpoint SHA it read. Nothing in this paper is typed in by hand.
• Comparisons are paired at the level of the fold: the same subject-disjoint splits, the same labels and the same probe are used for every method, so a per-row difference is structurally paired rather than a comparison of two independently-tuned pipelines.
• Multi-seed results carry a seed-level standard deviation. The multiday, cross-dataset and few-shot extensions are single-seed and are reported without error bars; they should be read as directional until the multi-seed sweep lands.
• The macro comparison is reported with a task-bootstrap interval and not with a Nadeau-Bengio t: each task seeds its own folds, so pairing cells across tasks is arbitrary and that interval comes out too narrow.
• No Lane E data entered pretraining, so the released encoder is Lane A only. This is not the same as saying Lane E never leaves the building: eight of the eighteen probe heads are fitted on CGMacros and are distributed, under the share-alike licence that CGMacros imposes. See §4.4 and D025.

4.4 What is open-weight, and what is not

• Code (this repository): Apache-2.0.
• Encoder weights (glucofm_encoder.onnx): CC-BY-NC-4.0. Pretrained on Lane A only, so no share-alike or no-derivatives source contributed to them.
• Probe heads (glucofm_heads.json): CC-BY-NC-SA-4.0. Eight of the eighteen heads are fitted on CGMacros, whose terms are share-alike, and CC BY-NC-SA §3(b) requires adapted material to be offered under the same licence. Labelling the bundle accordingly discharges that obligation; keeping it a separate artefact stops the term reaching the encoder, which never saw CGMacros. Recorded as D025.
• Never distributed: anything fitted on UCHTT1DM. Its no-derivatives term admits no labelling remedy — a derived classifier cannot be redistributed under any licence.
• Unresolved: the Stanford licence, discussed in §2.4. It bears on the encoder itself, not on any single head.

5. Reproducibility

A cold checkout, the released tarball of glucofm_encoder.onnx and glucofm_heads.json, and access to the public-data corpus are sufficient to run any headline number from this paper.

• End-to-end recipe: REPRODUCE.md
• Status from disk: just status — single source of truth
• Numerical parity for the released ONNX: tests/export/test_onnx_parity.py
• Lane E leakage gate: tests/unit/test_source_rights.py
• Mask ordering golden test: tests/golden/test_no_leakage.py
• Frozen spec: bundle/glucofm_public_reproduction_blueprint.md (the paper we reimplemented) + bundle/BLUEPRINT_AMENDMENTS.md (the A1–A7 deltas)

Wall-clock from a cold checkout to the headline number: ~10 h on an RTX 5090; ~30 h on an RTX 3090. The full Tier-1 ablation sweep adds ~8 h on a 5090.

6. Licensing and intended use

This project is research software. It is not a medical device, and nothing it produces should inform a decision about anyone's health.

The source code is Apache-2.0. The encoder weights are CC-BY-NC-4.0. The probe heads are CC-BY-NC-SA-4.0, one step stricter, because eight of them are fitted on share-alike material; the reasoning is in §4.4 and D025. Non-commercial use is the intended use and needs no permission from anyone. Commercial use of the weights requires a licence from the maintainer, and in the Stanford case (§2.4) would require resolving the upstream question first.

7. Acknowledgements

This is an independent project. It is not affiliated with, sponsored by, or endorsed by Google, the GlucoFM authors, or any of the dataset providers. The method was reimplemented from the publication without access to the authors' source code or weights. The PPG bridge data (Zenodo 20577959) is CC-BY-4.0. The four public CGM cohorts are credited in manifests/sources/registry.yaml under their respective licenses.

References

1. GlucoFM: A Dual-Stream Foundation Model for Continuous Glucose Monitoring. arXiv:2605.30865v2. The method reimplemented in this work.
2. The blueprint: bundle/glucofm_public_reproduction_blueprint.md — frozen spec from the publication, with corrections and additions documented in bundle/BLUEPRINT_AMENDMENTS.md.
3. CGM-JEPA. github.com/cruiseresearchgroup/CGM-JEPA, MIT. The comparator implementation is a faithful port of the authors' released code (master @ 2026-05-11).

Appendix A — Architecture parameters

model:        OpenCGMStateEvent
raw_statistics:        True        # D019: derive patch stats from raw, not normalised
normalize_targets:    False        # D019: target layer_norm disabled (no ablation activation)
streams:              "both"       # D019: state + event fused, not zeroed
learnable_sigma:      True         # paper §3.3 learnable causal Gaussian σ
use_circadian:        True         # paper §3.3 time-of-day embedding
zero_empty_patches:   False        # D020: as trained; the current repo default is True

Trainable parameter count: 732,593 — encoder 435,633 + predictor 132,480 + transition heads 2 × 82,240. Within 1.7 % of the paper's reported 0.72 M. The EMA target is a frozen copy of the encoder and contributes no trainable parameters. The released ONNX artifact is the encoder alone: 435,633 parameters.

For the avoidance of doubt, 521,584 is the CGM-JEPA comparator's parameter count (src/opencgm_stateevent/baselines/cgm_jepa.py), not ours.

Appendix B — Per-task reliability (the heads)

See findings/results_section.md §3.4 for the cross-dataset transfer table and §7 for the manifest of source licenses. Per-head roc_auc, coverage_p05/p95/median, has_signal for all 18 fitted task-source combinations is in artifacts/glucofm_heads.json (the consumer JSON shipped to HF and the web demo).

Appendix C — The 5 findings documents

• findings/results_section.md — the headline numbers, ablation table, few-shot, cross- dataset, multiday, PPG bridge pilots, architecture, and the gap analysis.
• findings/head_to_head.md — paired 5-seed comparison vs CGM-JEPA with CIs.
• findings/tier1_ablations.md — the 30-run Tier-1 ablation matrix.
• findings/ppg_pilot.md — the marginal PPG teacher-student pilot (D023, A7).
• findings/ppg_conditional.md — the input-conditioned PPG teacher-student pilot (D023 + A7 extension).

Appendix D — Standing rules

Reproduced here so this paper is self-contained.

• Never interpolate CGM. The physical observation mask is authoritative.
• Mask patches before normalisation, filtering, statistics. (§10.1)
• Lane E sources (NC/ND/SA) never enter a distributed checkpoint.
• Improvements live in separate checkpoints and separate tables.
• Claim "independent public-data reconstruction", never "reproduced Google's model".
• Every consequential choice is tagged PAPER_EXACT | SOURCE_VERIFIED | INFERRED_RECONSTRUCTION | PROPOSED_EXTENSION. A guess never silently becomes "the GlucoFM recipe".