LibraryFamily 3 · Predictive Risk Stratification
Sepsis early warning system — bedside deployment
Under reviewPilotEmerging evidence
Undergoing in-depth clinical, technical and governance review.
Continuously estimates the risk of sepsis from vital signs and laboratory results, prompting earlier clinical review of deteriorating patients.
Plain-language summary
AIChoose an audience and generate a tailored summary on demand. The AI uses only what is on this page.
Clinical context
- Clinical problem
- Predict a future risk
- Point of care
- Monitoring
- Nature of AI output
- A risk score
- Clinical specialty
- Intensive care
- Care setting
- Hospital — inpatient
- Patient population
- Adult inpatients on general medical and surgical wards.
- Intended use
- Continuously estimates the risk of sepsis from vital signs and laboratory results, prompting earlier clinical review of deteriorating patients.
Technology
- AI technique
- Classical machine learning
- Input data
- Vital signs, Laboratory results
- Output type
- Risk score
- Autonomy level
- Informs a human (advisory)
- Model provenance
- Built in-house
- Model version
- sepsis-ews v1.4
- Built on a general-purpose model
- No
Deployment
- Status
- Pilot
- Country
- Netherlands
- Deployment date
- 15 January 2025
- Sites
- 1
Regulatory & governance
- EU AI Act risk tier
- High-risk
- High-risk basis
- Annex III use case
- Medical device
- No
- EU MDR class
- Not a device
- CE marking
- Not required
- FDA status
- Not applicable
- ISO 14971 risk class
- Medium
- GDPR processing basis
- Public interest
- GDPR DPIA
- Completed
- Data identifiability
- Pseudonymised
- Explainability method
- Intrinsic
- Human oversight model
- Generates a ward review prompt; the clinical team decides on escalation and the sepsis bundle.
NICE evidence standards
- ESF tier
- Tier C — treat / diagnose / calculate risk
- Evidence category
- Category 2
Performance summary
- Headline metric
- AUC / AUROC
- Value
- 0.84
- Subgroup performance assessed
- Yes
- Known bias signals
- Early thresholds produced more alerts for older patients; recalibrated during the pilot to balance alert burden.
Similar deployments
AILooking for close matches in the Library…
Safety signals
Post-deployment concerns flagged by clinicians, patients or monitoring (proposal §3.7). High or critical signals re-enter a validated use case into maturation automatically.
Higher-than-expected alert volume on night shifts raised by ward staff; threshold recalibration under review.
Raised 17 May 2026, 17:58
MediumInvestigating
Evidence records
Studies and evaluations attached to this use case.
Retrospective validation
AUC / AUROC: 0.81
Population: General ward inpatients
Subgroup: age >= 75
Retrospective validation
AUC / AUROC: 0.84
Population: General ward inpatients, 40k admissions
Contributors
- Deploying organisation
- [demo] Amsterdam UMC · Hospital / health system · Netherlands
- AI vendor
- —
- Product name
- —