ARDIA PRECISION HEALTHGoverned AI for healthcare revenue & precision care
Free TrialBOOK A DEMO
● Measured, not claimed

How we test our models

Ardia is pre-revenue — so instead of made-up numbers, here is what we can actually measure today, reproducibly, from public data and our own test suite.

The honest line: One model (Cadence) is genuinely trained and measured on a public benchmark; Meridian's math is unit-tested; the persona models (Lumen, Aria, MolecuIQ, ToxIQ, PulmoIQ, TARA) run on a frontier model under guardrails and have no published accuracy yet — their benchmarks will be built the same way Cadence's was, once measured on real data under a BAA. Nothing here is invented, and none of it is a diagnosis.
95.45%
Cadence held-out accuracy
0.954
Macro-F1 (6 activities)
34/34
Automated tests passing
100%
Reproducible from public data

Cadence — the trained model

● 6 / 6 tests passing

Activity-recognition model trained from scratch on the public UCI HAR benchmark (Anguita et al., 2013, CC BY 4.0), evaluated subject-independently — the 9 test people never appear in training, so this measures generalisation to a new person. Model chosen by 5-fold cross-validation on the training split only, then scored once on the held-out subjects.

Accuracy vs. baselines

held-out (subject-independent), 2,947 windows
Balanced accuracy 95.33% · macro-F1 0.9545. Baselines shown so the score is judged against what "always guess the biggest class" (18.2%) or random (16.7%) would get.

Model selection (train-only CV)

5-fold on the training subjects, before touching the test set
Logistic regression (98.46%) narrowly beat random forest (98.04%) in cross-validation, so it was the one refit and scored once on held-out subjects. No test-set peeking.

Per-activity F1

how well each of the six activities is recognised
Sitting vs. standing is the hardest pair — both are still, upright postures — and that's the only real confusion (see the matrix). It does not affect the Active/Sedentary/Resting rollup, which is 99.63%.

Confusion matrix

rows = true · columns = predicted · held-out test
The near-diagonal is the model getting it right. The one warm off-diagonal cell (58) is sitting predicted as standing — the expected posture confusion, not a safety-relevant error.

Full metrics — precision · recall · F1

per activity, held-out test (2,947 windows)

Dataset, artifact & reproduction

everything needed to re-run this exact result
DatasetUCI HAR · CC BY 4.0
Windows (train / test)7,352 / 2,947
Subjects (train / test)21 / 9 · disjoint
Features561 engineered
ModelLogistic regression (scikit-learn 1.6.1)
Worst / mean / best subject85.71% / 95.23% / 99.48%
Artifactcadence_model.joblib (41 KB)
$ python models/ml-wellbeing/train_activity_model.py
$ python -m pytest tests/test_cadence.py -q

⚠ The honest limitation — stated, not buried

UCI HAR's volunteers are aged 19–48. Older adults move differently, so this 95.45% proves the method and pipeline — it does not prove Cadence is 95.45% accurate on a frail 82-year-old, and we won't imply it does. Before elder deployment, Cadence will be re-trained and re-validated on older-adult movement data under a BAA, on this same subject-independent protocol. It is non-diagnostic and not a fall detector.

Meridian — deterministic & unit-tested

● 7 / 7 tests passing

Meridian isn't an LLM — it's a deterministic model of the PAMA / CLFS rate cliff. The math lives in models/meridian/clfs.py and its tests pin the exact figures the site shows, so the numbers can never drift from the code.

What the tests lock down

tests/test_meridian.py — 7 assertions
Baseline CLFS revenue (seed mix)$895,551
Cuts compound 2027→2028→2029×0.85 each yr
Cumulative revenue at risk$728,419
2029 run-rate vs baseline61%
Reporting-penalty ceiling$10,000 / day
Rejects cut > 15% statutory cap✓ raises error
$ python -m pytest tests/test_meridian.py -q  # 7 passed

Why it's trustworthy

deterministic, sourced, no PHI

Same inputs always give the same answer. It applies the PAMA statute (up to 15%/yr for 2027–2029) to 2025 CLFS national rates (public CMS data) and the test mix you enter. It runs entirely in your browser — no upload, no server, no PHI. Try it live on the Meridian page →

The test suite

Everything above is enforced by an automated test suite in the repo — run it yourself.

tests/test_cadence.py6 passed  subject-independence · dataset shape · accuracy gate · worst-subject floor · saved-model reproduction
tests/test_meridian.py7 passed  baseline · compounding · revenue-at-risk · run-rate · penalty ceiling · statutory-cap guard
tests/test_sentinel.py10 passed  18 HIPAA Safe-Harbor categories · SSN/phone/email/MRN/date/ZIP redaction · CPT codes preserved · deterministic
tests/test_crucible.py11 passed  6 guardrail gates reported by the engine: de-identification · non-diagnostic · scope-of-practice · safety-escalation · honesty · human-in-the-loop — a failed gate withholds the answer entirely. Cite-or-abstain and policy-override are enforced in the retrieval and answer-binding layers, not counted as Crucible gates.
Total34 / 34 passing
$ python -m pytest tests/ -q  # 34 passed