GlucoTrace turns 24 hours of continuous glucose monitor readings into 128 numbers you can compare, search, and visualize. It's a small, readable Transformer that never fills in missing readings, and it's evaluated against tests written down before training.
A CSV with timestamp,glucose goes in, in mg/dL or with --unit mmol/L. You get back a fingerprint, a similarity score, or a single HTML report that works offline and loads no scripts.
$ pip install glucotrace $ glucotrace download-model $ glucotrace encode my-day.csv # 128-number fingerprint $ glucotrace compare monday.csv tuesday.csv $ glucotrace report my-day.csv --manifest corpus/manifest.json --output report.html
Most health-ML demos only show their best number. GlucoTrace was built so you can check its claims yourself.
Sensor gaps are tracked explicitly and never interpolated. A dropped reading can't pass for a real one.
Splits are made by participant and bound to checksummed manifests. The loader refuses splits that no longer match their data.
Thresholds, probes, and the number of candidates are fixed in writing first. Each test set is opened once.
A three-layer Transformer that trains on a laptop CPU in minutes, with 109 tests and no hidden services.
Measured on people the model never saw during training.
| Question | Answer |
|---|---|
| Stable when readings go missing? | YESMedian cosine 0.965 after removing 30% of readings, on new people (sealed test) |
| Links days from the same person? | OFTENA day's closest match is the same person about half the time. Treat fingerprints as personal data |
| Clinically validated? | NONo diagnostic, treatment, or safety claim is made |