AEGIS Implementation Guide
What is AEGIS?
AEGIS (An Operational Infrastructure for Post-Market Governance of Adaptive Medical AI) is a modular, rule-based framework for governing medical AI systems that undergo iterative learning after initial market authorisation. It provides a structured pipeline for continuous monitoring, performance evaluation, and regulatory-compliant decision-making throughout the post-market lifecycle of adaptive AI/ML-enabled medical devices.
AEGIS is not a machine learning library, a training framework, or a model architecture. It is an operational governance layer that sits on top of any ML pipeline and provides:
- Structured dataset management across training, golden, and drifting splits
- Composite performance scoring with clinically weighted metrics
- Covariate and concept drift detection with statistical rigour
- A deterministic, auditable decision engine with priority-ordered logic
- Full traceability to ISO 14971:2019, FDA PCCP, and EU MDR/AI Act requirements
**Sepsis example:** In the sepsis prediction case study, AEGIS governed 11 iterations (k = 0 through k = 10) of a Random Forest classifier trained on 20,336 patients from Hospital A of the PhysioNet Sepsis Challenge, with 4,067 reserved as a fixed performance reference (Golden evaluation set). The framework autonomously issued 8 APPROVE decisions (iterations 0--5 and 8--9), 1 CONDITIONAL APPROVAL (iteration 6), 1 CLINICAL REVIEW (iteration 7), and 1 REJECT (iteration 10) — exercising all four deployment decision categories across the experiment. ALARM was co-issued at iterations 8 and 10. At iteration 10, sensitivity dropped to 0.428 (below the 0.65 threshold), triggering simultaneous REJECT + ALARM and demonstrating the framework's safety boundaries under extreme conditions.
When to use AEGIS
AEGIS is designed for medical AI systems that meet all three of the following criteria:
| Criterion | Description |
|---|---|
| Adaptive | The model is retrained or fine-tuned on new data after deployment |
| Regulated | The device falls under FDA 21 CFR Part 11, EU MDR, or the EU AI Act |
| Continuous | Post-market surveillance requires ongoing performance monitoring |
If your system is a locked (non-learning) model, AEGIS still provides value for post-market surveillance (PMS) via the alarm channel, but the full iterative governance loop will not apply.
Three-module pipeline
AEGIS operates through three sequential modules executed at each iteration:
Iteration k
┌──────────┐ ┌──────────┐ ┌──────────┐
│ DARM │ ──> │ MMM │ ──> │ CDM │
│ Dataset │ │ Model │ │ Clinical │
│ Assimilat.│ │ Monitor │ │ Decision │
│ & Retrain│ │ Metrics │ │ Module │
└──────────┘ └──────────┘ └──────────┘
│ │
│ ┌────────────────────────┐ │
└────│ Feedback: D_T,k+1 │<───┘
│ = D_T,k ∪ D_D,k │
└────────────────────────┘
Reading order
This guide is structured as a sequential implementation walkthrough. Each page builds on the previous one. We recommend reading in order:
| # | Page | What you will learn |
|---|---|---|
| 1 | Core concepts | DARM, MMM, CDM definitions; fixed performance reference; notation |
| 2 | Configuration | How to instantiate AEGIS for your domain |
| 3 | Thresholds | ISO 14971 threshold derivation and calibration |
| 4 | CDM logic | Decision hierarchy, alarm conditions, pseudocode |
| 5 | MLcps | Composite performance scoring mechanics |
| 6 | Drift detection | Covariate and concept drift methodology |
| 7 | Audit trail | Logging requirements and JSON format |
| 8 | PCCP mapping | FDA PCCP regulatory alignment |
| 9 | EU regulatory mapping | EU MDR and AI Act compliance |
Relationship between guide and supplements
Each guide page references one or more sections of the supplementary material. The supplement field in each page's metadata traces back to the exact supplement section where the formal specification resides.
| Guide page | Supplement section(s) | Downloadable form |
|---|---|---|
| Core concepts | S1 | S1 Checklist |
| Configuration | S2.1--S2.2 | S2 Configuration |
| Thresholds | S2.3 | S2 Threshold Table |
| CDM logic | S3.3 | S3 CDM Pseudocode |
| MLcps | S2.4 | -- |
| Drift detection | S2.5 | -- |
| Audit trail | S1.4 | -- |
| PCCP mapping | S2.8, S3.5 | -- |
| EU regulatory mapping | S2.8 | -- |
Quick start
If you are short on time, focus on these three pages first:
- Core concepts -- understand the vocabulary
- Configuration -- fill in the S2 configuration form
- Thresholds -- derive your safety-critical thresholds
Then return to the remaining pages as you move toward deployment.
Source
Manuscript Sections I--II. The overall AEGIS infrastructure is introduced in Section I (Introduction) and formally defined in Section II (Methods). The three-module pipeline (DARM, MMM, CDM) is described in Section II-B. The DARM architecture is depicted in Figure 1 and the overall AEGIS infrastructure in Figure 4 of the manuscript.