Scored hourly on 12,431 held-out patients, and without retraining on 27,484 patients in three external databases, the results support silent validation in a hospital, the step before any alert reaches a clinician.
Zentus scores every ICU patient-hour for five deterioration events: circulatory failure, hyperglycaemia and sepsis 8 hours ahead, and kidney and liver failure 48 hours ahead. All results are on held-out or external patients.
Internal test set
On 12,431 held-out patients from MIMIC-III, MIMIC-IV and eICU, scored hourly, the five endpoints reached a macro AUROC of 0.932 and a macro AUPRC of 0.725. Liver failure was strongest at 0.982; sepsis was weakest at 0.845.
Macro AUROC across five endpoints on the internal rolling test set: 12,431 patients, 17,411 ICU stays.
Three external databases
Without any retraining, macro AUROC was 0.935 on NWICU, 0.854 on Zigong and 0.798 on the paediatric PICDB. AUPRC is much lower externally, 0.286, 0.324 and 0.104, because events are rarer there, 0.1% to 6.4% of patient-hours.
Each endpoint, in each outside database.
NWICU 17,349 patients
Zigong 2,255 patients
PICDB, children 7,880 patients
Each field is 100 pairs of patient-hours, one before deterioration and one stable. Bright dots: pairs where Zentus ranked the deteriorating hour higher (AUROC × 100). No retraining.
What it supports
The packet concludes that the evidence is enough to request prospective silent validation. It lists what is not yet done: no hospital's data feeds have been mapped or audited, thresholds have not been frozen with clinicians, silent logging has not run in a hospital, calibration drift over time is unknown, and the clinician adjudication workflow is not yet complete.
