Step-by-step deployment workflow

Supplement S3

Eight-step deployment workflow covering clinical context assessment through regulatory submission. Each step includes actions, deliverables, and links to relevant guide pages and templates.

Step 1 — Assess clinical context

Supplementary S3.1

Identify the clinical domain, patient population, and intended use
Determine regulatory jurisdiction (US FDA / EU MDR / both)
Classify device risk (Class I, IIa, IIb, III-PMA)
Document intended use statement per IEC 62304

Step 2 — ISO 14971 hazard analysis

Supplementary S3.1

Conduct hazard analysis per ISO 14971:2019
Identify hazardous situations related to model performance degradation
Assign hazard IDs to each identified risk
Derive all AEGIS thresholds from the hazard analysis (P_fail, R_G, τ, θ, P_PMS, δ)
Document clinical justification for each threshold value

Step 3 — Configure DARM

Supplementary S3.2

Define three dataset splits: Training D_T,k, Golden D_G, Drifting D_D,k
Lock and hash the golden dataset before first deployment
Verify patient/sample deduplication across all splits
Document the data accumulation rule: D_T,k+1 = D_T,k ∪ D_D,k
Establish dataset versioning and provenance metadata
Document label governance process

Step 4 — Configure MMM

Supplementary S3.3

Select primary performance metrics with clinical justification
Configure MLcps weights (0.5–3.0 per metric)
Select drift detection method and specify feature set
Define fairness metrics and relevant subgroups
Set up KPI historical database

Step 5 — Implement CDM

Supplementary S3.3

Implement decision hierarchy (P1–P4) from reference pseudocode
Implement alarm conditions (A1–A3)
Decide which intermediate categories to activate (CONDITIONAL APPROVAL, CLINICAL REVIEW)
Document polarity convention for all conditions
Implement iteration logging with all 11 required fields

Step 6 — Run iteration 0

Supplementary S3.4

Train initial model on D_T,0
Evaluate on golden dataset D_G
Record baseline metrics as gold standard
Verify CDM returns APPROVE for baseline performance
Lock gold standard with version tag and dataset hash
Document all metrics in audit trail

Step 7 — Enter iterative loop

Supplementary S3.5

For each iteration k ≥ 1:
Accumulate drifting data: D_T,k+1 = D_T,k ∪ D_D,k
Retrain model on D_T,k+1
Evaluate on D_G (performance) and D_D,k (drift)
Run CDM evaluation
Log all fields to audit trail
If REJECT or ALARM: initiate CAPA workflow
If APPROVE: release model with version tag

Step 8 — Document for submission

Supplementary S3.6–S3.7

Complete S2 configuration template with all deployment parameters
Map AEGIS configuration to PCCP components (if FDA pathway)
Map to EU MDR / AI Act requirements (if CE pathway)
Compile audit trail history
Prepare CAPA records for any REJECT/ALARM events
Submit documentation per regulatory pathway

All data stays in your browser. Nothing is sent to any server.