IEC 62304 Medical Device Software Lifecycle

by Nick Clark | Published April 25, 2026 | PDF

IEC 62304 governs the lifecycle of medical-device software with a rigor that increases with the safety class assigned to each software item, and its expectations were originally formed around statically engineered codebases that change only through controlled releases. Continuously improving medical AI breaks that assumption: model behavior evolves in deployment, adaptation may be conditioned on local data, and clinical environments diverge from the development reference. Spatial adaptation supplies the architectural primitive that allows evolving medical software to remain inside the IEC 62304 envelope, providing runtime signed artifacts, sandbox pre-activation certification, cross-model portability, and regulatory-aware activation as structural properties rather than process commitments.


Regulatory Framework

IEC 62304:2006 with Amendment 1:2015 ("Medical device software — Software life cycle processes") defines the international consensus framework for medical-software development, maintenance, configuration management, problem resolution, and risk management. The standard partitions software into items and units, requires the manufacturer to assign a safety class of A, B, or C to each item based on the possible harm if the item fails, and scales the depth of required activities accordingly. Class A applies where no injury or damage to health is possible; Class B where non-serious injury is possible; and Class C where death or serious injury is possible. The standard is harmonized under the EU Medical Device Regulation 2017/745 (MDR) through Annex Z mappings, and is recognized by the FDA as a consensus standard for software lifecycle support of premarket submissions.

IEC 62304 does not exist in isolation. Section 4.2 explicitly requires that a risk-management process conforming to ISO 14971 be in place, and the standard's hazard-analysis hooks feed directly into ISO 14971 risk files. Configuration management aligns with ISO 13485:2016 and, in the United States, with the FDA's harmonized Quality Management System Regulation (QMSR). For software with continuously evolving behavior, the FDA's Predetermined Change Control Plan guidance and the EU MDCG guidance on software qualification and classification (MDCG 2019-11) provide complementary expectations about how change is to be controlled. The 2026 revision of IEC 62304 is expected to expand explicit treatment of machine-learning-enabled software, post-market software updates, and SOUP (software of unknown provenance) handling for foundation-model components.

Within the lifecycle proper, IEC 62304 mandates software development planning (Clause 5.1), requirements analysis (5.2), architectural design (5.3), detailed design (5.4), unit implementation and verification (5.5), integration and integration testing (5.6), system testing (5.7), and release (5.8), with traceability among hazards, requirements, design elements, code, and tests required throughout. Maintenance (Clause 6), configuration management (Clause 8), and problem resolution (Clause 9) extend the discipline beyond initial release. Each clause assumes that change occurs through a deliberate, documented process with verification appropriate to the affected safety class, an assumption that is increasingly difficult to honor when adaptation is intrinsic to the device's clinical value.

Architectural Requirement

The structural challenge presented by adaptive medical software is that IEC 62304 traceability is grounded in identified, versioned software items, while adaptation by its nature produces new effective behavior without producing new versioned items in the conventional sense. If a Class C item adapts in deployment, the manufacturer must still be able to demonstrate that every behavior the item exhibits in the field is covered by the verification activities prescribed for Class C. That demonstration cannot rest on retrospective analysis alone, because Clause 5.5.2 requires verification before integration, and Clause 5.6 requires integration testing before system release. The architecture must therefore make adaptation events first-class lifecycle objects with their own identity, integrity, and verification status.

A second architectural requirement arises from Clause 7 (software risk management process), which requires that risk control measures be implemented in the software and verified as effective. When adaptation alters the behavior on which a risk control depends, the verification of the risk control must be re-established before the adapted behavior is permitted to execute clinically. This implies a sandbox in which adapted behavior can be exercised against the risk-control evidence base before activation. A third requirement emerges from Clause 8 (configuration management), which requires that the configuration of every released software item be reproducible: adaptive artifacts must therefore carry signed, retrievable identity that survives device service events and supports reconstruction during MDR investigations under EU MDR Article 87 or FDA 21 CFR Part 803.

Why Procedural Compliance Fails

A procedural approach to IEC 62304 conformance in adaptive software typically converts every adaptation into a "release" through document workflow, requiring change-control board review, regression testing, and formal release notes for each adaptation event. This approach fails on two axes. Operationally, the cadence of clinically meaningful adaptation in modern AI-enabled devices is incompatible with manual release governance: adaptation events that should occur on the timescale of clinical deployment cannot wait for weekly or monthly review cycles. Evidentially, the procedural model produces release artifacts whose connection to the actually-deployed software is mediated by human attestation, and an attestation chain cannot answer Clause 8.1.2's reproducibility requirement when the adapted artifact is not itself a controlled object.

The procedural failure becomes acute under post-market surveillance. When an incident is reported, the manufacturer must retrieve the exact configuration of the device at the time of the incident, including any adapted state. If the adaptation pipeline is procedural, the retrieval depends on log correlation across disparate systems, and the connection between the retrieved logs and the actually executed behavior cannot be established to the standard required by ISO 14971 risk file updates. A Class C software item whose adaptation history cannot be reconstructed places the entire device's regulatory status at risk, because Clause 9 (problem resolution) requires that problems be analyzed against the configuration that produced them.

What the Spatial-Adaptation Primitive Provides

Spatial adaptation reframes adaptation events as architectural objects: each adaptation produces a runtime signed artifact whose identity, provenance, parameters, and verification status are bound into a single signed structure. The signature attaches the artifact to a specific configuration baseline, satisfying the Clause 8 reproducibility requirement without requiring procedural release ceremony. Sandbox pre-activation certification ensures that no signed artifact reaches the clinical execution path until it has been exercised against the risk-control evidence base in a structurally isolated environment. The certification is itself a signed record, so the lifecycle audit trail is closed under inspection.

Cross-model portability addresses the SOUP and component-substitution problem that arises when foundation-model components or third-party model layers are updated. The primitive isolates the adaptation envelope from the underlying model identity, so that a change in the SOUP component triggers re-certification of the adaptation under the new substrate without invalidating the architectural commitments. Regulatory-aware activation, finally, gates artifact deployment on the regulatory state of the destination jurisdiction: an artifact certified under FDA QMSR and IEC 62304 Class C requirements activates only in jurisdictions where those certifications are valid, and an artifact certified to a more restrictive envelope (for example MDR Class IIb additional Annex VIII rules) activates correspondingly. Together these four elements convert continuous improvement from a lifecycle exception into a structurally lawful operating mode.

Compliance Mapping

Mapping spatial adaptation to IEC 62304 clauses is direct. Clause 5.1 (development planning) accommodates adaptation pipelines as planned activities with defined deliverables, namely the signed artifact and its certification record. Clause 5.2 (requirements) is unchanged, because the adaptation envelope is itself a requirement. Clause 5.3 (architectural design) explicitly identifies the adaptation manager, the sandbox, and the activation gate as architectural items with defined safety classes, typically inheriting the highest class of the items they affect. Clauses 5.5 through 5.7 (verification, integration, and system testing) attach to the sandbox certification protocol, ensuring that the verification evidence required for the relevant safety class is produced before activation.

Clause 7 (risk management) ingests the adaptation envelope and risk-control set, and the activation gate enforces that no adaptation outside the envelope can reach clinical execution. Clause 8 (configuration management) is satisfied by the signed artifact, which is a configuration item in the strict sense and supports retrieval during incident analysis. Clause 9 (problem resolution) is supported by the immutable adaptation log, which permits exact reconstruction of the configuration at the time of any reported event. The mapping extends naturally to ISO 14971 risk files and to FDA PCCP commitments, because the modification protocol can be expressed as a constraint on the adaptation envelope, the change classes correspond to envelope expansions or contractions, and the verification methods correspond to the sandbox certification protocol.

Adoption Pathway

Manufacturers can adopt spatial adaptation incrementally without rewriting their existing IEC 62304 lifecycle infrastructure. The first step is to inventory the existing adaptation surface — every place where deployed device behavior depends on data, parameters, or models that can change post-release — and to express each surface as a candidate adaptation envelope. The second step is to instrument the runtime to produce signed artifacts whenever the envelope is exercised, even before formal certification is in place, which immediately produces the audit substrate that Clauses 8 and 9 require. The third step is to construct a sandbox that exercises each candidate artifact against the existing verification suite for the relevant safety class.

The fourth step is to integrate the activation gate with the manufacturer's regulatory configuration, which includes jurisdiction, declared safety class, MDR or QMSR commitment, and any PCCP scope. The fifth step is to revise the IEC 62304 software development plan and architectural design documents to reflect the primitive as a first-class architectural element, which enables Notified Body or FDA review to proceed against a coherent technical narrative. Sponsors with active MDR Notified Body engagement should expect that the adaptation envelope and certification protocol become focal points of technical documentation review, and sponsors with FDA-cleared devices should expect that PCCP authoring becomes substantially easier once the primitive is in place. Across both regimes, the adoption pathway is incremental and evidence-positive: each step produces evidence that strengthens the existing lifecycle file rather than displacing it.

Nick Clark Invented by Nick Clark Founding Investors:
Anonymous, Devin Wilkie
72 28 14 36 01