Results

Five independently trained models, 120 epochs each, evaluated on 14 clinical questions with subject-disjoint splits. Negative results included.

Headline

ROC-AUC is the probability that a randomly chosen positive case is ranked above a randomly chosen negative one. 0.5 is chance.

Chance

0.500

Raw glucose readings

0.607

Clinical metricstime-in-range, mean, variability

0.643

CGM-JEPAthe nearest published model, identical folds

0.643

OpenCGM-StateEvent

0.670

0.50

0.55

0.60

0.65

0.70

ROC-AUC. The axis starts at 0.50 — chance — not at zero.

Improvement over…	This model	CGM-JEPA	Reading
Clinical metrics	+0.0269[0.0222, 0.0316]	0.0000[-0.0049, 0.0050]	This interval excludes zero; the comparator’s does not.
Raw readings	+0.0628	+0.0359	Both improve on raw input.

Brackets are 95% confidence intervals over seeds. The macro advantage is real and its interval excludes zero, but it does not survive being broken into single questions: this model is ahead of the clinical baseline on 13 of 18 task-cohort rows, and on none of them does the difference reach significance after Holm correction — in any seed. Each row has 29 to 100 people in it. What the evidence supports is a small consistent lift across tasks, not a set of per-task wins.

Regenerate:just eval-headline && just head-to-headreports/eval/head_to_head_5seed.csv~80 min on an RTX 3090

Ablations

Each design choice was removed and the model retrained: ten conditions, three seeds each, 40 epochs. The seed-level standard deviation is 0.0084, so differences smaller than that are noise.

Remove this…	Score	Change	Verdict
Event stream only	0.6229	-0.0504	load-bearing
No decomposition (raw signal)	0.6500	-0.0233	load-bearing
No time-of-day signal	0.6580	-0.0152	load-bearing
State stream only	0.6592	-0.0141	load-bearing
Trained without Shanghai	0.6649	-0.0084	within noise
Trained without Stanford	0.6669	-0.0064	within noise
No temporal-dynamics loss	0.6710	-0.0023	within noise
Gaps filled in by interpolation	0.6738	+0.0005	within noise
No data augmentation	0.6757	+0.0024	within noise
Fixed (not learned) filter width	0.6765	+0.0032	within noise

The event stream alone loses 0.0504, six times the noise level. Removing the decomposition entirely loses 0.0233. The state/event split is where most of the benefit comes from.

Regenerate:just ablation-sweep && just ablation-aggregatereports/eval/tier1_ablations_3seed.csv~8 h on an RTX 5090

Two results went against expectation. Augmentation contributes nothing measurable. And interpolating the sensor gaps, which the design refuses to do, scores +0.0005 — level with the masked version at this evaluation depth. The never-interpolate rule is defensible on other grounds, but at 40 epochs it is not supported by a measured gain.

Artefacts

Everything the numbers above were computed from, and the files this site serves.

encoder.onnx	1.99 MB	sha256 b1349deffd15ab62…	CC-BY-NC-4.0
heads.json	18 probes	14 above the signal floor	CC-BY-NC-SA-4.0
encoder.meta.json	provenance	checkpoint, epoch, seed, architecture flags	Apache-2.0

findings/head_to_head.md — per-seed and per-task tables →findings/tier1_ablations.md — the full ablation matrix →REPRODUCE.md — measured wall-clocks, copy-paste commands →The CGM-JEPA port, so the comparison is checkable →Model card — intended use, limits, withheld heads →

Cite

@software{fourdrinier_opencgm_stateevent,
  author  = {Fourdrinier, Stephane},
  title   = {OpenCGM-StateEvent: an independent public-data reconstruction
             of the GlucoFM dual-stream CGM foundation model},
  year    = {2026},
  url     = {https://github.com/sfourdrinier/opencgm},
  note    = {Code Apache-2.0; encoder CC-BY-NC-4.0; probe heads CC-BY-NC-SA-4.0}
}