Full-Stack Cognition Architecture for Healthcare

Regulatory framework

Healthcare is the most heavily regulated AI deployment domain. HIPAA and the HITECH Act define privacy and breach-notification obligations for protected health information across covered entities, business associates, and downstream subcontractors. 21 CFR Part 11 governs the integrity, attribution, and audit-trail requirements for electronic records and electronic signatures used in FDA-regulated activity. 21 CFR 820, the Quality System Regulation (now harmonized toward ISO 13485 under the QMSR final rule), defines design-control, validation, and post-market surveillance obligations for medical devices including software-as-a-medical-device.

For clinical AI specifically, IEC 62304 defines the software lifecycle for medical device software, partitioning the system into safety classes (A, B, C) with increasing rigor for documentation, unit and integration verification, and configuration management. ISO 14971 defines the risk-management process: hazard identification, risk estimation, risk-control measures, residual risk evaluation, and post-production information feedback. IEC 60601-1 governs general safety and essential performance for medical electrical equipment, with collateral standards for usability (60601-1-6), alarm systems (60601-1-8), and cybersecurity expectations through AAMI TIR57 and the FDA premarket cybersecurity guidance.

The FDA AI/ML SaMD Action Plan and the Predetermined Change Control Plan framework recognize that learning systems evolve post-market. A PCCP defines, at clearance time, the specific modifications a model may undergo without triggering a new submission, the SaMD Pre-Specifications (SPS) that bound those modifications, and the Algorithm Change Protocol (ACP) that disciplines them. EU MDR raises the bar further with mandatory clinical evaluation, post-market clinical follow-up, and Notified Body oversight for higher-risk classes. ONC certification under the 21st Century Cures Act adds USCDI v3 data-element obligations, FHIR R4 APIs, CDS Hooks for decision-support invocation, and information-blocking prohibitions. CMS Conditions of Participation, NPI registration, and state medical-board licensure layer across all of this.

Architectural requirement

This regulatory surface implies a specific architectural shape, not a checklist. Patient identity must persist across encounters, providers, and care settings, because privacy obligations and clinical safety obligations both attach to the patient — not to any one record system. Clinical inference must be evaluated against the full patient context at the moment of generation, because liability, formulary, contraindication, and protocol constraints are contextual and cannot be safely litigated after the fact. Training data and model updates must carry evidence grades and provenance, because PCCPs, MDR clinical evaluation, and ISO 14971 residual-risk arguments all rest on traceable evidence chains. Provider state must be observable, because IEC 60601-1-6 usability and CMS staffing-quality measures depend on the human in the loop remaining cognitively coherent. Knowledge retrieval must be governed, because ONC information-blocking and Part 11 attribution obligations attach to every clinical assertion derived from literature.

No single AI capability satisfies this shape. A diagnostic imaging model satisfies one cell of the matrix. A clinical documentation tool satisfies another. The matrix as a whole is satisfied only by an integrated stack whose layers communicate through shared primitives.

The fragmented AI problem in healthcare

Health systems today deploy AI in isolated capabilities: a diagnostic imaging model, a clinical documentation tool, a patient scheduling optimizer, a sepsis-prediction alert, an ambient scribe, a coding assistant. Each system operates independently with its own data plane, its own governance binding, and its own representation of the patient. A patient interacting with five AI-assisted systems has five identity representations, five governance frameworks, five audit trails, and no continuity of cognitive state across interactions.

This fragmentation is a governance gap rather than an integration inconvenience. No single system holds sufficient context to govern its outputs appropriately because each sees only its slice of the patient's clinical picture. The sepsis alert does not know about the documentation assistant's prior summary. The imaging model does not know what the formulary checker has already flagged. The PCCP for one model cannot reason about drift introduced by another. ISO 14971 risk control becomes a per-tool exercise that misses the system-level hazards that emerge from tool interaction.

Why procedural compliance fails

The procedural answer to fragmented AI is documentary: produce a HIPAA risk analysis, a Part 11 validation package, an IEC 62304 software-development plan, an ISO 14971 risk management file, a 510(k) or De Novo submission with a PCCP, an MDR technical file, an ONC attestation, and a SOC 2 report. Each artifact is necessary. None is sufficient, because each describes intended behavior at a point in time while the deployed system continues to evolve.

Procedural compliance fails specifically at three seams. First, at the seam between models: a documented PCCP for model A does not constrain model B, and the institution carries the residual risk of their interaction without an artifact that describes it. Second, at the seam between training and inference: training-time evidence grades are not carried into inference-time decisions, so a recommendation grounded in a case report is presented to the clinician with the same surface confidence as one grounded in a Cochrane review. Third, at the seam between the AI and the clinician: usability obligations under 60601-1-6 and CMS quality measures presume a coherent clinician, but no procedural artifact observes whether the clinician downstream of the AI is in a state to receive its outputs safely. These seams are where adverse events originate, and they are precisely where documentation cannot reach.

What the AQ primitive provides

The cognition architecture supplies five primitives that together close the seams. Biological identity provides patient continuity through behavioral trajectory rather than through opaque MRNs that fragment across systems. The patient's identity persists across hospital, rehabilitation, ambulatory, and home settings because the trust slope binding accumulated context travels with the patient rather than being re-established at each site. This satisfies HIPAA's minimum-necessary rule structurally — context is composed, not copied — and gives ONC's USCDI v3 and FHIR R4 surfaces a stable referent.

Inference control governs clinical AI at the point of generation. Every candidate recommendation transitions against the patient's persistent state — current medications, allergies, contraindications, active care plan, formulary, prior authorizations, and institutional protocol — before the transition is committed. Unsafe recommendations are not generated and then filtered; they are structurally prevented. This converts the ISO 14971 risk-control argument from a probabilistic claim about a black-box model into a deterministic claim about an admissibility gate, and it gives a PCCP a stable surface to constrain because the gate, not the model, is the regulated boundary.

Training governance stratifies clinical AI learning by evidence depth. Randomized-trial findings train at foundational depth. Guideline statements train at protocol depth. Case reports train at recognition depth only. Provenance tracing connects every model-mediated recommendation to its evidentiary basis, satisfying Part 11 attribution and providing the traceable chain that MDR clinical evaluation and FDA PCCP submissions require. The IEC 62304 lifecycle becomes auditable end-to-end because the artifacts of training are themselves first-class governed objects rather than spreadsheets in a quality folder.

Disruption modeling observes provider coherence. Nursing units, physician practices, and care teams are assessed for trajectories that indicate developing burnout, alarm fatigue, or team dysfunction. The signal is not used punitively; it modulates AI behavior. A provider whose coherence trajectory is deteriorating triggers enhanced confirmation requirements for high-risk recommendations and routes lower-acuity decisions to peer review. This satisfies the human-factors expectation embedded in IEC 60601-1-6 and the staffing-quality expectations under CMS Conditions of Participation in a way that is observable rather than asserted.

Semantic discovery provides governed clinical knowledge retrieval. Clinicians query the literature through persistent discovery objects that accumulate clinical context and return evidence-graded, provenance-traced results rather than keyword-ranked document lists. The same evidence grades used in training are the trust weights used in retrieval, so the model and the librarian agree on what counts as foundational. Information-blocking obligations under the Cures Act are satisfied because retrieval is auditable end-to-end.

Compliance mapping

Each primitive lands on specific regulatory cells. Biological identity addresses HIPAA minimum-necessary, the Privacy Rule's accounting-of-disclosures obligation, and the ONC patient-matching expectation under USCDI v3. Inference control addresses ISO 14971 risk control, FDA 21 CFR 820 design output verification, IEC 62304 Class B/C software unit and integration verification, the PCCP Algorithm Change Protocol surface, and EU MDR clinical-performance requirements. Training governance addresses 21 CFR Part 11 attribution and audit trails, the FDA AI/ML SaMD Action Plan's Good Machine Learning Practice principles, MDR Annex II technical-documentation requirements, and ONC's real-world-testing obligations. Disruption modeling addresses IEC 60601-1-6 usability engineering, CMS Conditions of Participation staffing-quality measures, and the human-factors expectations of the FDA premarket guidance for AI-enabled devices. Semantic discovery addresses Part 11 record integrity, ONC information-blocking prohibitions, and the citation-traceability expectations that NIH and AHRQ guidance increasingly impose on clinical AI.

The mapping is not one-to-one. The architectural value emerges from cross-layer integration. Biological identity informs inference control: the patient's identity trajectory is the clinical context against which inference governance operates. Training governance informs semantic discovery: the evidence grades used in training are the same trust weights used in retrieval. Disruption modeling connects to inference control: a deteriorating clinician triggers enhanced admissibility gates. The shared primitives — trust slopes, governance bindings, persistent state objects — are what allow a single MDR technical file or a single PCCP to coherently describe the system rather than fragments of it.

Adoption pathway

A health system does not adopt the full stack at once. The adoption pathway begins where the regulatory exposure is highest and the integration cost is lowest. Inference control is typically first, deployed as a wrapper around an existing clinical decision support surface (CDS Hooks endpoints, sepsis alerts, drug-interaction checkers) so that admissibility gating becomes the regulated boundary the FDA submission and ISO 14971 file describe. The wrapper requires no model retraining and gives the institution an immediate auditable surface.

Training governance is second, deployed for new model development and PCCP submissions. Existing models are brought under governance retroactively as their PCCPs are renewed. Biological identity is third, deployed at care-transition seams (admit, discharge, referral) where MRN fragmentation produces the most acute safety and privacy exposure. Semantic discovery follows, replacing ad-hoc literature retrieval in clinical pathways and quality-improvement work. Disruption modeling is last, because it requires the longest baseline and the most careful labor-relations framing; it is deployed first as an aggregate signal to nursing leadership and only later wired into inference-time confirmation logic.

Through every stage, the stack interoperates with existing clinical systems rather than replacing them. The EHR remains the system of record. CDS Hooks remain the invocation surface. FHIR R4 remains the wire format. The cognition architecture supplies the governance, identity, and coherence infrastructure that those systems lack, so that AI-assisted care is governed by construction across every interaction rather than by attestation after the fact.