AEGIS: An Operational Infrastructure for Post-Market Governance of Adaptive Medical AI Under US and EU Regulations

The gist: AEGIS presents an AI/ML Evaluation and Governance Infrastructure for Safety that operationalizes FDA PCCP requirements and EU AI Act Article 43(4) provisions through three modules: DARM, MMM, and CDM.

How it works

AEGIS comprises three core modules to operationalize regulatory change control concepts; these are the Dataset Assimilation and Retraining Module (DARM), the Model Monitoring Module (MMM), and the Conditional Decision Module (CDM) AEGIS comprises three core modules to operationalize regulatory change control concepts; these are the Dataset Assimilation and Retraining Module (DARM), the Model Monitoring Module (MMM), and the Conditional Decision Module (CDM)

The DARM manages controlled ingestion and structured integration of new observations, assigning definitive status to every data point before they can be assigned to dataset splits The DARM addresses EU AI Act data governance requirements, MDR evidence needs, and PCCP control over retraining data This module maintains three datasets at each iteration: the Training dataset (DT,k) for parameter updates, the Golden dataset (DG) as a static benchmark, and the Drifting dataset (DD,k) for drift monitoring

The MMM calculates metrics on dataset splits and maintains historical records of Key Performance Indicators (KPIs), including performance, fairness, and distribution categories The MMM provides unbiased evidence for continuous monitoring by calculating performance metrics on both static and dynamic sets from the DARM It tracks parameters such as Accuracy, sensitivity, specificity, AUC, F1-score for model performance metrics and KL divergence or JS distance for distribution shift metrics

Key Governance Mechanisms

AEGIS implements a four-category deployment decision taxonomy (APPROVE, CONDITIONAL APPROVAL, CLINICAL REVIEW, REJECT) combined with an independent PMS ALARM signal This composite output enables detection of the critical governance state in which no deployable new model is available and the currently released model is simultaneously assessed as at risk The decision logic is formalized using a ternary operator D = f(C1, C2,..., CN)? D1: D2, where f is an aggregated Boolean function

The deployment decision taxonomy maps directly to regulatory structures; for instance, APPROVE aligns with PCCP scope and EU Article 43(4) exemptions REJECT triggers escalation under both jurisdictions, while the ALARM signal addresses PMS obligations under EU Medical Devices Regulation (MDR) Article 83 and FDA PMS expectations The composite output captures the critical REJECT + ALARM state, where no deployable new model is available and the released model is simultaneously failing, representing the critical safety escalation state

Performance Comparison and Thresholds

AEGIS parallels FDA’s concept of SE used in 510(k) submissions by implementing a “fixed performance reference comparison mechanism” New iterations must perform within defined bounds of the fixed performance reference metrics or demonstrate improvement This mechanism serves two functions: it acts as a trigger for the deployment decision at P2 (CLINICAL REVIEW) and provides the reference against which the PMS ALARM trigger A2 is evaluated The fixed performance reference comparison coupling ensures that the candidate model performance on Golden dataset at iteration k (PG,k) does not silently degrade below the released performance reference (P rel G) or the fixed performance reference (P ref)

The system supports two approaches for evaluating acceptable performance: deterministic threshold comparison, where point estimates are compared directly against thresholds, and CI-based equivalence testing, which is recommended for production deployments with smaller post-market samples The fixed performance reference (P ref) is derived from data available at the time of the device’s initial regulatory-authorized release or from externally established benchmarks

Demonstrated Generalizability and Outcomes

AEGIS was applied to two heterogeneous domains to demonstrate generalizability, accommodating different data modalities through configuration alone The sepsis prediction experiment across 11 iterations yielded 8 APPROVE decisions, 1 CONDITIONAL APPROVAL from minor cross-site drift, 1 CLINICAL REVIEW from regression from fixed performance reference, and 1 REJECT when sensitivity fell below the threshold

The brain tumor segmentation instantiation demonstrated that the same governance logic naturally produces APPROVE, ALARM, and REJECT decisions from real distributional shifts across imaging cohorts The two case studies together confirm both that all categories are reachable through appropriate configuration and that selective activation is legitimate design practice AEGIS bridges both regulatory paradigms by mapping the four-category deployment taxonomy to both FDA PCCP and EU Article 43(4) exemptions

Trustworthy AI Dimensions

The AEGIS architecture accommodates the full spectrum of Trustworthy AI requirements through concurrent monitoring of explainability metrics derived from SHAP analysis, fairness metrics assessing performance parity across demographic subgroups, and calibration quality via Brier scores and the relationship between PPV/NPV and clinical prevalence For EU AI Act transparency requirements, manufacturers could formalize these dimensions as additional PCCP boundaries

The system’s capacity to distinguish between data distribution changes (triggering the PMS ALARM signal for investigation) and actual safety violations (triggering REJECT for the new candidate model) is a key capability The MLcps metric provides a principled approach to composite performance assessment by projecting multiple evaluation metrics onto a two-dimensional polar coordinate system

Limitations and Future Directions

The current implementation relies on aggregate performance metrics that may not capture shifts in model uncertainty distribution, and it does not explicitly address adversarial threats like data poisoning attacks The authors propose extending the AEGIS’s monitoring layer to incorporate disagreement-based signals adapted from ensemble learning approaches to detect data integrity compromise Future work should incorporate uncertainty quantification methods to strengthen confidence estimation for release decisions by providing per-prediction reliability indicators alongside aggregate metrics

The fixed performance reference comparison mechanism serves two distinct functions in the composite CDM output, ensuring that the candidate model performance on Golden dataset at iteration k (PG,k) does not silently degrade below the released performance reference (P rel G) or the fixed performance reference (P ref) The two instantiations establish that AEGIS produces the correct decision under controlled conditions and under real distributional change, but neither constitutes epidemiological evidence of any drift types clinical prevalence

The study demonstrates that AEGIS provides what is termed operational implementation, translating high-level policies into executable procedures The authors acknowledge that the sepsis prediction and brain tumor segmentation applications serve exclusively as proof-of-concept demonstrations of governance infrastructure functionality, not clinical evaluation or validation The authors emphasize that all thresholds are illustrative and production values require ISO 14971:2019 hazard analysis with documented traceability

The proposed extension to the temporal ensemble approach would involve maintaining previously validated model versions in a “shadow” or “sentinel” mode, continuously comparing their predictions against the current deployed model The final composite output is reported as (deployment decision) or (deployment decision + ALARM) when any of A1–A3 fires This composite state, which requires regulatory escalation beyond standard rollback, is unrepresentable in prior single-hierarchy frameworks

The authors thank the Stockholm Medical Artificial Intelligence and Learning Environments (SMAILE) core facility at Karolinska Institutet and by EU grant 240037 from EIT Health for support All code is available through the GitHub repository https://github.com/ki-smile/aegis The authors declare no competing interests The study used fully anonymized and public datasets, and as no human participants or patient data were involved, institutional review board approval was not required All scientific interpretations remain the sole responsibility of the authors

The paper is structured to align with the FDA’s PCCP components as specified in the August 2025 final guidance, providing manufacturers a structured approach to implementing PCCPs as defined by FDA guidance The four-category deployment taxonomy combined with the parallel PMS ALARM signal is designed to support documentation of predetermined governance pathways in 510(k) submissions Manufacturers should engage early with regulators to align parameters with expectations for specific device classifications This work presents AEGIS, a governance infrastructure for continuously learning medical AI systems The modular architecture (DARM, MMM, CDM) provides manufacturers a structured approach to implementing PCCPs as defined by FDA guidance Demonstration across the PhysioNet Sepsis Challenge dataset established generalizability and practical feasibility The experimental results demonstrate five key AEGIS capabilities, including the REJECT + ALARM composite state

The authors conclude that AEGIS provides what is termed operational implementation, translating high-level policies into executable procedures The two case studies together confirm both that all categories are reachable through appropriate configuration and that selective activation is legitimate design practice The final composite output is reported as (deployment decision)

1. textbfThe AEGIS Infrastructure for Adaptive Governance (DARM, MMM, CDM): The system can operationalize FDA PCCP requirements and European Union (EU) AI Act Article 43(4) provisions for predetermined changes. This means the system provides a structured approach to implementing governance policies by specifying what must be monitored and how decisions must be made.

2. textbfFour-Category Deployment Taxonomy with Independent PMS ALARM Signal: The system can manage complex release states, such as the critical governance state in which no deployable new model is available and the currently released model is simultaneously assessed as at risk. This includes co-issuing signals like APPROVE + ALARM or the critical REJECT + ALARM composite state.

3. textbfFixed Performance Reference Comparison Mechanism: The system can enforce regulatory Substantial Equivalence (SE) by implementing an analogous mechanism for model updates, ensuring New iterations must perform within defined bounds of the fixed performance reference metrics or demonstrate improvement. This is achieved by comparing candidate model performance against the fixed performance reference comparison mechanism.

4. textbfContext-Dependent Configuration: The system can be adapted to different clinical needs by allowing configuration of parameters such as data modality, model architecture, primary metrics, MLcps weighting, drift detection method, enabling it to handle diverse applications like tabular time-series (sepsis) and volumetric data (brain tumor segmentation).

5. textbfDeterministic Decision Logic with Priority Hierarchy: The system can produce auditable governance decisions by using a formal decision function where If f evaluates to True, decision D1 is selected; otherwise, D2, ensuring that safety violations are detected first according to the priority order defined in Table IV.

Machine learning systems deployed in medical devices require governance frameworks that ensure safety while enabling continuous improvement. Regulatory bodies including the FDA and European Union have introduced mechanisms such as the Predetermined Change Control Plan (PCCP) and Post-Market Surveillance (PMS) to manage iterative model updates without repeated submissions. This paper presents AI/ML Evaluation and Governance Infrastructure for Safety (AEGIS), a governance framework applicable to any healthcare AI system. AEGIS comprises three modules, i.e., dataset assimilation and retraining, model monitoring, and conditional decision, that operationalize FDA PCCP and EU AI Act Article 43(4) provisions. We implement a four-category deployment decision taxonomy (APPROVE, CONDITIONAL APPROVAL, CLINICAL REVIEW, REJECT) with an independent PMS ALARM signal, enabling detection of the critical state in which no deployable model exists while the released model is simultaneously at risk. To illustrate how AEGIS can be instantiated across heterogeneous clinical contexts, we provide two examples: sepsis prediction from electronic health records and brain tumor segmentation from medical imaging. Both cases use identical governance architecture, differing only in configuration. Across 11 simulated iterations on the sepsis example, AEGIS yielded 8 APPROVE, 1 CONDITIONAL APPROVAL, 1 CLINICAL REVIEW, and 1 REJECT decision, exercising all four categories. ALARM signals were co-issued at iterations 8 and 10, including the critical state where no deployable model exists and the released model is simultaneously failing. AEGIS detected drift before observable performance degradation. These results demonstrate that AEGIS translates regulatory change-control concepts into executable governance procedures, supporting safe continuous learning for adaptive medical AI across diverse clinical applications.