A silent pilot in one adult ICU.
For one to three months the model scores your unit's patients every hour and logs what it would have flagged, and at the end your team reviews the results with us.
- Scope
- One adult ICU
- Duration
- One to three months
- Shown to clinicians
- Nothing
- Thresholds
- Fixed before the first hour is logged
- At the end
- Chart review with your clinicians, and a report on your own patients
A week of flags, 16 beds.
559 flags
5.0 per patient-day · 21% of 2,688 patient-hours
Each line is a bed; each tick is a flag. The count comes from our test results; where flags land is illustrative. Conservative mode caught no kidney or liver failure in retrospective scoring.
A report on your own patients.
The metrics are agreed before the first prediction is logged.
Calibration, top decile
Predicted Observed
What it covers
- Discrimination, flag rate and lead timePer endpoint
- Sensitivity, specificity, PPV and NPV at the frozen thresholdsPer endpoint
- Calibration and drift over the pilotPer month
- Chart review of flags and missed eventsWith your clinicians
- Data availability, scoring failures and uptimePer hour
For IT and information governance.
Governance and data access
- The protocol requires prediction logs to stay inside the hospital's approved environment.
- Your governance team decides who can see patient-level logs and how long prediction, monitoring and outcome logs are kept.
- Silent mode means no bedside alert, escalation, paging or order suggestion.
Data feeds
The pilot will connect to your existing FHIR or HL7 feeds and score every hour. Before scoring starts, each input will be mapped to its source field, unit, expected refresh cadence and acceptable range.
Identifiers
Patient, encounter and ICU stay identifiers will be logged as salted hashes. No raw MRN will appear in analytic exports. Your site confirms which identifiers may be logged and who owns and stores the hash salt.
Stale and missing data
Scoring will be suppressed when input data are stale or missing. A row will be written for each ICU stay at each prediction time, recording its status: scored, suppressed for stale data, suppressed for missing data, or error. Logs will be stored row by row so the analysis can be replayed exactly.
Frozen version
The model version, threshold policy and input schema will be frozen before the pilot starts. Each row will record the model version, a hash of the model artefact or manifest, and the threshold policy version, along with the code revision used for scoring.
Rollback and downtime
A model rollback procedure, a downtime communication path and an incident review process for unexpected logging or scoring behaviour will all be defined before the pilot starts.
Monitoring
A monitoring row will be written for each unit every hour, tracking scored eligible stays, data missingness and freshness, scoring failures and latency, drift in input distributions, and calibration and alert burden. Each prediction row will record model latency and the latency from ingestion to log.
Named owners
Your site names a clinical owner, a technical owner and a safety monitoring owner, and confirms the chart-review process and who is responsible for adjudication.
Before any alerts
Only once all of these are true:
- Calibration is stable enough for the intended threshold policy.
- Alert burden is acceptable to clinicians.
- Median lead time is clinically useful.
- No patient subgroup shows a major unsafe loss of performance.
- Clinicians approve an action pathway for each endpoint.
- Regulatory and institutional review requirements are satisfied.
Request a silent pilot.
Tell us about your unit. We reply from hello@zentus.health within two working days.