Model v1.0.0Research baseline · release blocked
Governance-first ML · Care operations

Prioritize follow-up.
Prove every decision.

Source-grounded 30-day readmission risk with calibrated probability, transparent drivers, subgroup auditing, and a release gate that refuses unsupported clinical claims.

◇Source-groundedUCI dataset with DOI and lineage.
◉Calibrated probabilityReliability measured on held-out patients.
⚖Release-gatedClinical promotion stays blocked.
Reference case

30-day readmission risk

Research only
…moderate model score
0%Model flag threshold —45%+
Top local driverslog-odds contribution
✓

Model provenance verifiedUCI 296 · model v1.0.0 · deterministic API

Decision support research only — not a diagnosis or clinical recommendation.

▤

Public dataset101,766 encounters130 US hospitals · 1999–2008

⌁

DiscriminationROC-AUC 0.628AP 1.67× prevalence baseline

◎

CalibrationBrier 0.093ECE 0.8%

⚑

Operating pointRecall 78.1%Precision 13.0%

Performance truth

Useful probability, not a magic answer.

The positive class appears in only 10.7% of the held-out set. CarePath therefore reports ranking, calibration, recall, precision, and subgroup behavior together—not accuracy alone.

ROC-AUC0.628
Balanced accuracy0.577
Recall0.781
Precision0.130
Audit surface

Every score carries its context.

Data lineageUCI 296 · CC BY 4.0verified
Patient leakageGroup-aware splitpassed
Sensitive inputsExcluded from modelpassed
External validationNot completedrequired