Project Overview
An NHS ICU Network covering 8 hospital intensive care units with 280 ICU beds and 18,000 annual admissions wan...
Technology Stack
Compliance & Standards
The Challenge
An NHS ICU Network covering 8 hospital intensive care units with 280 ICU beds and 18,000 annual admissions wanted to deploy an AI-powered sepsis early warning system — analysing electronic observations (vital signs, blood results, NEWS2 score) to detect sepsis risk before clinical deterioration. MHRA Class IIa SaMD (clinical decision support), IEC 62304, DCB0129 (ICU sepsis mortality is the critical safety risk), DTAC all 5 domains, NHS FHIR R4, UK GDPR Article 9, and WCAG 2.1 AA were mandatory. Budget: £100,000.
Our Approach
Training dataset
42,000 ICU admissions from NHS Digital Clinical Practice Research Datalink (CPRD — HRA ethics approval).
Features
vital signs (HR, BP, RR, SpO2, temperature), blood results (WBC, lactate, creatinine, bilirubin), urine output, and NEWS2 score.
Model
XGBoost classifier (NEWS2-augmented — 24 clinical features).
Output
- sepsis risk score (0–100), risk category (low/moderate/high), and predicted time to deterioration (0–6 hours).
- Sepsis-3 definition: SOFA score >
- = 2 with suspected infection — model trained to predict Sepsis-3 criterion before clinical recognition.
- Real-
NHS Electronic Observations
vital signs entered by nurses in EPR (Epic, SystmOne, or Cerner) → HL7 v2 ORU (Observation Result Unsolicited) message → ClickMasters FHIR R4 Observation (vital sign) → model input.
Blood results
HL7 v2 ORU from laboratory information system (LIS) → FHIR R4 Observation (pathology) → model input.
Observation latency
- vital sign recorded in EPR → sepsis risk score updated within 60 seconds.
- Real-time dashboard: ICU coordinator sees all patients with high sepsis risk, time since last observation, and intervention recommendation.
Highest clinical safety hazard
AI false negative (high-risk patient shown as low-risk → delayed sepsis treatment → death).
Mitigation
- 1model sensitivity calibrated at 94% (higher false positive rate — better to alert than miss),
- 2AI risk score is an alert tool — does NOT replace clinical NEWS2 assessment,
- 3manual override: ICU nurse can escalate any patient regardless of AI score,
- 4audit: all AI alerts and clinical responses reviewed by clinical governance weekly.
MHRA Class IIa justification
AI assists clinician decision, clinician makes treatment decision — AI does not act autonomously.
Sepsis high risk alert
- AI score >
- = 70 → FHIR R4 Communication resource (alert) → GOV.UK Notify SMS to ICU nurse and duty ICU consultant → alert displayed on ICU dashboard with amber urgency.
Sepsis critical alert
- AI score >
- = 90 → immediate pager alert (ICU consultant and consultant intensivist) + escalation to ICU charge nurse.
Alert acknowledgement
nurse acknowledges alert within 15 minutes or second alert triggered.
Clinical response recording
nurse records clinical assessment and intervention against each alert (evidence for clinical governance audit).
The Results
MHRA Class IIa registration obtained.
DTAC approved all 5 domains.
Platform live at 18 weeks, £92,000 — under budget.
Sepsis detection sensitivity (model vs clinical chart review): 94.2% (target: 90%).
Time from observation entry to alert: 52 seconds average.
Sepsis treatment initiation time (from retrospective data vs matched cohort): 42 minutes → 28 minutes (33% improvement — faster treatment drives outcome improvement).
Mortality in high-alert sepsis patients: 18.4% vs 24.6% matched historical cohort (6.2 percentage point improvement).
“Sepsis mortality 18.4% versus 24.6% matched historical cohort — 6.2 percentage points. In ICU, that is lives saved. Treatment initiation 42 minutes to 28 minutes. 94.2% detection sensitivity — above our 90% target. MHRA Class IIa first submission. DCB0129 accepted without amendment. DTAC first submission. The clinical governance team described this as the most impactful clinical technology deployment in the network's history." — Network Clinical Director, NHS ICU Network (name withheld)”
Project Details
Related Case Studies
View AllMedTech Wearable Data Analytics Platform — UK NHS Trust Research
An NHS Foundation Trust conducting clinical research in cardiac rehabilitation wanted to build a wearable data...
MedTech AI Diagnostic Imaging Platform — UK Radiology Group
A UK radiology group with 18 sites and 340,000 annual scans wanted to build an AI-assisted diagnostic imaging ...
MedTech NHS Antimicrobial Stewardship Platform — UK Acute Trust
An NHS Acute Trust with 520 beds managing 14,000 antibiotic prescriptions per month needed an antimicrobial st...
Ready to Transform Your Business?
Let's discuss how our technical expertise can help you achieve remarkable results.