Sentinel: An honestly-evaluated, calibrated, and deployed decision-support system for 30-day hospital readmission risk on a public benchmark

Published July 8, 2026 | Version v1.0.0

Preprint Open

Thirty-day hospital readmissions are frequent, costly, and partly preventable, and predicting

them is a crowded problem on which most models are only modestly discriminative. We argue

that on such a saturated task the useful contribution is not a higher discrimination score but

the disciplined integration and honest evaluation that the literature repeatedly calls for and

rarely delivers together. We present Sentinel, a decision-support system for 30-day readmission

risk built end-to-end on the public UCI Diabetes-130 benchmark and deployed as a publicly

reachable service. Sentinel couples a leakage-disciplined, patient-grouped evaluation protocol

with probability calibration, decision-curve utility analysis, a pre-registered subgroup fairness

audit, exact per-patient SHAP explanations, and a pre-registered distribution-shift robustness

study. Its discrimination is deliberately honest—a sealed holdout AUROC of 0.677, at the

field’s ceiling for this data—while isotonic calibration reduces expected calibration error (ECE)

from 0.33 to 0.026 (exact: 0.3305 to 0.0264). The robustness study shows that discrimination

transports across simulated shift but the probability scale is prevalence-specific, so a site with

a different base rate must recalibrate. We report what the system does not establish—most

importantly, it has not been validated on an external cohort—and release all code, a live demo,

and the trained artifacts for reproduction

Files

Name	Size	Download all
sentinel_preprint.tex md5:da1a0905f0732d45eae910e026d26a19	24.4 kB	Download
SentinelPreprint.pdf md5:6d60a8224ea8c4eb3f972deb7734dc1a	405.2 kB	Preview Download