How UnitMesh Works
Every unit follows the same governed pathway — shadow observation, formal audit, active deployment, and continuous monitoring. No recommendation reaches a clinician without passing all four gates.
Shadow Mode — 90-Day Observation
The AI runs silently alongside clinical workflows. Recommendations are generated and logged, but never surfaced to clinicians.
What runs
The model generates recommendations on live patient data in real time — but outputs go only to the audit log, never to the EHR. Zero clinical impact.
What's tracked
Recall, FPR, and override rate logged per 100 decisions. Clinician workflows are unchanged — they see nothing different during shadow mode.
What triggers exit
90-day minimum and ≥ 500 scored decisions — both required before a formal audit can be scheduled.
Formal Audit — Dual Approval
Shadow data is reviewed against fixed thresholds. No approval chain shortcuts — every reviewer must sign before activation proceeds.
Reviewers
CDO + CMO required for all tiers. Red-tier units (highest patient risk) additionally require Legal sign-off — triple approval.
Pass thresholds
Recall ≥ 0.80 · FPR ≤ 0.15, measured on the shadow period dataset. Both thresholds must be met before activation can proceed.
If audit fails
The unit returns to extended shadow mode. A second failed audit triggers a model review by the clinical AI team before any further activation attempt.
Sign-off record
CDO, CMO (and Legal for Red-tier) each provide a dated digital signature — immutable, stored in the audit log, retrievable for FDA/CE inspection.
Active Deployment — EHR Integration
Recommendations reach clinicians via FHIR CDS Hooks — no EHR replacement, no workflow disruption. Every decision is logged with full provenance.
Delivery
Recommendations surface as FHIR CDS Hooks cards inside the clinical workflow. The EHR calls UnitMesh; UnitMesh returns a structured card with summary, rationale, and suggested action.
What clinicians see
A card with the recommendation, a confidence indicator, and a one-line rationale. Clinicians can accept, modify, or override — all responses are captured.
Decision provenance
Every scored decision is logged with model version, feature snapshot, timestamp, clinician ID, and response. Full audit trail for every patient interaction.
Ongoing Operations — Continuous Monitoring
Activation is not the end. Every unit is monitored continuously — drift triggers retraining, adverse events trigger suspension, and deactivation is always one step away.
Fitness score
Platform-wide 0–100 score computed from recall, FPR, override rate, and adverse event count across all active units. Displayed on the Governance Portal dashboard.
Drift detection
Monthly post-market surveillance (PMS) cycle reviews rolling recall and FPR. If either metric degrades past threshold, the unit is flagged for review.
Retraining
Flagged units enter a federated retraining cycle — each hospital's data contributes under differential privacy constraints. The retrained model re-enters shadow mode before re-activation.
Adverse events
Any reported adverse event triggers immediate unit suspension, root-cause analysis, and a mandatory CAPA (Corrective and Preventive Action) before reinstatement.
Deactivation path
A unit can be deactivated back to shadow at any time by CDO or CMO — no approval chain required to reduce risk.
Ready to start a shadow deployment?
Shadow mode has zero clinical impact. Your team observes. The platform proves itself.
Book a Demo →