Global life sciences AI · Published 16 November 2026 · HERMS Consultancy DMCC
See how HERMS delivers integrated life sciences solutions.
AI-powered medical devices need evidence and controls across their lifecycle—not only a strong model-development score. Define the intended medical purpose, establish representative data and clinically meaningful acceptance criteria, evaluate the device in its intended context, control changes and monitor real-world performance. A model update is not automatically a permitted device change. Manufacturers must connect each version to its evidence, applicable authorization and risk assessment, while users need clear limitations, appropriate oversight and a reliable route for reporting problems. Artificial Intelligence-Enabled Medical Devices Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions MDCG 2020-1: Guidance on Clinical Evaluation (MDR) / Performance Evaluation (IVDR) of Medical Device Software
Start with intended use: what decision does the device support?
Describe the medical purpose, patient population, intended users, input data, output, clinical setting and role in care. Does the software prioritize images, suggest a diagnosis, estimate risk or drive an intervention? State whether users independently review the result and what happens if the output is wrong or unavailable.
FDA regulates medical devices, not AI as an abstract technology. It assesses intended use and technological characteristics under the applicable device framework. Some software functions are excluded from the US device definition; conversely, an AI feature embedded in hardware can still require device-level assessment. SaMD describes software with a medical purpose that performs that purpose without being part of a hardware medical device; it is not a synonym for every AI health application. Artificial Intelligence-Enabled Medical Devices
Treat promotional wording as part of the claims boundary. Evidence supporting prioritization does not automatically support autonomous diagnosis. A broadened population, setting or clinical role can change the evidentiary and regulatory assessment even if the model weights stay identical. Artificial Intelligence-Enabled Medical Devices MDCG 2020-1: Guidance on Clinical Evaluation (MDR) / Performance Evaluation (IVDR) of Medical Device Software
Build a lifecycle evidence map before collecting more data
Use a version-linked evidence map connecting clinical claims, hazards, requirements, tests, release decisions and monitoring. The table is a proposed planning framework; actual documentation and approval routes depend on the device, jurisdiction and applicable quality system.
| Lifecycle stage | Core question | Evidence/control | Accountable function |
|---|---|---|---|
| Definition | What is the device meant to do? | Intended use, claims and pathway rationale | Clinical / Regulatory / product owner |
| Data preparation | Are data appropriate and trustworthy? | Provenance, labels, subgroup coverage and separation | Data science / clinical experts |
| Development | Does the implementation meet specifications? | Configuration, software tests and risk controls | Engineering / quality |
| Validation | Does the device meet user needs in context? | Representative testing and clinical evaluation | Clinical / biostatistics / quality |
| Release | Is this exact version ready and authorized? | Acceptance, labeling, cybersecurity and release approval | Quality / Regulatory |
| Change | Is the modification within an allowed scope? | Impact assessment, regression and applicable PCCP | Cross-functional change board |
| Post-market | Is performance maintained in use? | PMS signals, real-world evaluation and corrective action | PMS / clinical safety / quality |
| Retirement | Can discontinuation occur safely? | User communication, records and transition arrangements | Manufacturer / deployment owner |
FDA’s January 2025 AI-device lifecycle document remains draft, nonbinding and not for implementation. Its lifecycle proposals are useful context, not a final universal checklist. Existing applicable device obligations and authorized conditions continue to govern the product. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations
Verification, validation and clinical evaluation are different questions
Verification asks whether a specified implementation meets its requirements. Validation asks whether the device meets intended user needs and intended use. Clinical evaluation considers whether the evidence supports clinical safety, performance and the claimed benefit in context. A passing unit test, retrospective accuracy figure or cybersecurity scan cannot substitute for the other evidence.
EU guidance for medical device software organizes evaluation around valid clinical association or scientific validity, technical or analytical performance, and clinical performance. The exact terminology and route depend on whether MDR or IVDR applies. Each component needs evidence proportionate to the intended purpose and risk; do not simply rename a machine-learning test report a clinical evaluation report. MDCG 2020-1: Guidance on Clinical Evaluation (MDR) / Performance Evaluation (IVDR) of Medical Device Software
| Evidence layer | Question | Example limitation |
|---|---|---|
| Clinical association / scientific validity | Is the output meaningfully related to the target clinical condition or state? | A plausible correlation may not support the intended medical claim |
| Technical / analytical performance | Does the software reliably generate the intended output from the input? | Strong results may depend on a narrow scanner or laboratory configuration |
| Clinical performance | Does it achieve intended performance in the target population and setting? | Retrospective testing may not reflect actual users, prevalence or workflow |
Design representative, independent performance testing
Define acceptance criteria and the statistical analysis before final testing. Specify the deployment population, sites, equipment, acquisition protocols, reference standard and relevant subgroups. Justify sample size and uncertainty estimates in relation to clinical risk rather than choosing the largest available convenience dataset. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations MDCG 2020-1: Guidance on Clinical Evaluation (MDR) / Performance Evaluation (IVDR) of Medical Device Software
Separate development, tuning and final evaluation data. Prevent leakage through repeated patients, related images, duplicated records, site-specific artifacts or label-derived features. External and temporal evaluation can challenge generalizability, but independence must be demonstrated rather than assumed from a different file name or institution.
Examine label quality and disagreement. A reference standard may be imperfect, influenced by prior care or inconsistently applied. Document adjudication, blinding where appropriate and how uncertainty affects interpretation. Evaluate missing, corrupted and unsupported inputs; the device needs safe handling when it cannot provide a reliable result. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations MDCG 2020-1: Guidance on Clinical Evaluation (MDR) / Performance Evaluation (IVDR) of Medical Device Software
Choose measures matched to purpose. Sensitivity and specificity can be relevant to detection; predictive values vary with prevalence. Calibration matters when estimated probabilities inform decisions. Workflow tools may also need evidence concerning delay, alarm burden or user performance. An aggregate AUC alone does not show clinical usefulness, acceptable subgroup performance or safe deployment. MDCG 2020-1: Guidance on Clinical Evaluation (MDR) / Performance Evaluation (IVDR) of Medical Device Software
Evaluate the human–AI team and the deployment environment
An assistive device is used by people within a process. Test whether intended users understand the output, limitations, uncertainty and appropriate response. Consider automation bias, alert fatigue, inappropriate reassurance and use beyond the supported population. The relevant comparison may concern the clinician–device workflow, not the algorithm in isolation.
Assess local integration: input formatting, units, patient matching, image transfer, latency, downtime and output display. A validated model can be unsafe if the wrong patient record receives its result. Proposed deployment checks include interface testing, user training, local compatibility verification and a defined fallback when the device is unavailable. Do not change the authorized intended use through an informal local configuration. Artificial Intelligence-Enabled Medical Devices MDCG 2020-1: Guidance on Clinical Evaluation (MDR) / Performance Evaluation (IVDR) of Medical Device Software
Control the whole configuration—not only the model weights
A device version includes more than an algorithm checkpoint. Preprocessing, thresholds, feature extraction, user interface, dependencies, input connectors and hardware can alter clinical behavior. For generative systems, prompts, retrieval sources, decoding configuration and external model services can also be consequential.
Maintain a controlled configuration record and assess changes to each relevant component. Evaluate third-party update practices, cybersecurity, access, compatibility and supply-chain dependencies. “The model is unchanged” is not a sufficient conclusion when an interface change reverses a measurement or a retrieval update changes medical content. This is a proposed engineering discipline aligned with lifecycle risk management. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions
How a predetermined change control plan works
FDA’s August 2025 final PCCP guidance describes an optional mechanism for specified planned modifications to AI-enabled device software functions. A PCCP is reviewed within the marketing submission. Changes consistent with an authorized plan can be implemented without an additional submission for each described modification, subject to the applicable conditions. It is not blanket permission for unrestricted self-learning or a change in intended use. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions
| PCCP component | What it should make explicit | Implementation evidence |
|---|---|---|
| Description of Modifications | Specific planned changes and their boundaries | Which inputs, performance characteristics or configurations may change |
| Modification Protocol | How each change will be developed, tested and implemented | Data management, retraining, evaluation, acceptance criteria and update procedures |
| Impact Assessment | Benefits, risks, mitigations and cumulative effects | Comparison to the prior version and assessment of interactions |
Check authorization scope before relying on a PCCP. A sponsor-written plan is not the same as a plan cleared or approved through the applicable route. Changes outside the authorized boundary still need assessment under applicable modification rules and may require a new submission. Joint FDA, Health Canada and MHRA principles support a focused, risk-based, evidence-based lifecycle approach but do not create one international authorization. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions Predetermined Change Control Plans for Machine Learning-Enabled Medical Devices: Guiding Principles
A practical change-control decision tree
- Identify the exact change: model, data, threshold, interface, intended population, medical claim or environment.
- Assess whether intended use, benefit–risk, classification, performance or safety could be affected.
- Determine whether an authorized PCCP exists in the relevant market and covers this exact modification and implementation method.
- Define required data, testing, subgroup analysis, regression, cybersecurity and human-factors reassessment.
- Document the regulatory route, required external review and internal release authority before deployment.
- Release only the approved configuration; communicate material changes and ensure version traceability.
- Monitor the update, reconcile deployed versions and invoke containment or rollback when predefined triggers are met.
These steps are a proposed operating checklist, not a substitute for device-specific legal assessment. A modification may be acceptable in one market and require additional review in another. Cumulative small changes deserve scrutiny even when each appears limited on its own. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions Predetermined Change Control Plans for Machine Learning-Enabled Medical Devices: Guiding Principles
Post-market monitoring: detect harm, not just data drift
Plan monitoring before release. Define which signals matter, how they will be collected, who assesses them and what triggers action. Input drift, performance deterioration, safety complaints, subgroup disparities, failed runs and unexpected usage provide different information; none alone is a universal proxy for clinical safety. Medical devices: post-market surveillance Why post-market surveillance is a priority for the MHRA AI Airlock programme
| Monitoring domain | Example signal | Response to define in advance |
|---|---|---|
| Input compatibility | New scanner or changed measurement distribution | Verify supported environment and restrict unsupported inputs |
| Clinical performance | Sensitivity or calibration changes against reliable outcomes | Investigate measurement, case mix and model performance |
| Subgroups | Deterioration in a clinically relevant population | Targeted review; assess labeling or use restrictions |
| Workflow | Delayed outputs, ignored alerts or misuse | Human-factors assessment and operational mitigation |
| Safety / vigilance | Potential serious incident or emerging trend | Investigate and meet applicable reporting/action requirements |
| Configuration | Unintended version or unauthorized update | Contain deployment, assess impact and reconcile records |
Outcome labels may arrive late or remain unavailable. Track leading operational indicators while establishing a defensible route to meaningful clinical evaluation. Drift detection without investigation does not establish whether a device is safe; stable inputs likewise do not prove continued clinical performance. Why post-market surveillance is a priority for the MHRA AI Airlock programme
For Great Britain, strengthened post-market surveillance requirements took effect on 16 June 2025. MHRA distinguishes GB from Northern Ireland, where EU device PMS rules apply. The GB framework includes risk-proportionate PMS systems and appropriate safety reporting and action. Avoid generalizing GB requirements to every country or describing regulatory sandbox findings as new law. Medical devices: post-market surveillance The Medical Devices (Post-market Surveillance Requirements) (Amendment) (Great Britain) Regulations 2024
Hypothetical example: an imaging update across hospital sites
A manufacturer proposes retraining an image-analysis device using additional hospital data. Its intended purpose remains a bounded assistive task. Before implementation, the team confirms whether the modification falls within the authorized PCCP in the relevant market, checks data provenance and label quality, and preserves independent evaluation data.
Testing considers relevant equipment, acquisition protocols, demographic groups and difficult cases, with prespecified acceptance criteria. A higher aggregate metric does not justify release if a clinically important subgroup deteriorates beyond an acceptable boundary. Quality, clinical and regulatory reviewers evaluate the complete evidence and applicable route.
After controlled release, the manufacturer tracks installed versions, operational failures, complaints and suitable performance signals. A site-specific issue triggers investigation and containment; an unsupported scanner is not silently solved by changing the model locally. This scenario illustrates a proposed workflow, not a HERMS deployment or reported performance result.
Regulatory landscape: keep law, guidance and proposals separate
| Jurisdiction / resource | Status | Practical distinction |
|---|---|---|
| US FDA device regulation | Binding law/regulations and device-specific authorization; supporting guidance is generally nonbinding | 510(k) clearance, De Novo classification and PMA approval are distinct pathways |
| FDA AI lifecycle guidance | January 2025 draft; not for implementation | Do not describe draft recommendations as final requirements |
| FDA AI PCCP guidance | August 2025 final nonbinding guidance | Authorized plan boundaries matter; not unrestricted updates |
| EU MDR / IVDR and MDCG software guidance | Binding regulations; MDCG documents are nonbinding guidance | Clinical/performance evaluation depends on device and purpose |
| EU AI Act interaction | Separate risk-based regime; June 2025 AIB/MDCG guidance explains interaction | Assess actor, scope and current applicability; no timetable inferred from older guidance |
| Great Britain PMS | Strengthened binding rules effective June 2025; MHRA explanatory guidance | Separate GB/NI positions; monitoring continues after market entry |
Not all healthcare AI is a regulated device, and not every device-related AI function receives the same classification or pathway. EU AI Act obligations do not replace MDR/IVDR obligations. The June 2025 interaction guidance is useful for scope and coordination; confirm the current amended implementation timetable separately rather than copying historical dates. Artificial Intelligence-Enabled Medical Devices AIB 2025-1 / MDCG 2025-6: Interplay between the Medical Devices Regulation (MDR) & In vitro Diagnostic Medical Devices Regulation (IVDR) and the Artificial Intelligence Act (AIA) Medical devices: post-market surveillance
Release-readiness and monitoring checklist
- Intended use and claims match the device evidence and target markets.
- Data provenance, reference standards and independent evaluation are documented.
- Prespecified criteria address clinically meaningful outcomes and relevant subgroups.
- Human factors, integrations, failure handling and cybersecurity have been assessed.
- The exact configuration is controlled and release authority is documented.
- Changes are evaluated against the authorized scope and relevant market rules.
- PMS roles, thresholds, reporting pathways and corrective actions are operational.
- Version traceability, records and end-of-support arrangements are defined.
How HERMS fits an evidence-led device lifecycle
HERMS describes regulatory-strategy, manufacturing-quality and evidence-integrity solution areas within its portfolio. For medical-device projects, the useful evaluation question is whether the proposed workflow can connect claims, test evidence, change decisions and accountable approvals. This article does not independently validate HERMS medical-device functionality, regulatory acceptance, integrations or performance. AI Life Sciences Software | 25 HERMS Solutions
Explore HERMS regulatory, quality and evidence-integrity solutions to discuss a demonstration centered on your intended use, validation evidence and change-control requirements.
Frequently asked questions
What is an AI-powered medical device?
It is a medical device incorporating AI for a function within its intended purpose. The medical claims, actual function and applicable legal definition matter more than the AI label. A general administrative assistant is not automatically a device; software supporting diagnosis or treatment may require device assessment. Artificial Intelligence-Enabled Medical Devices
Is an accurate model a validated device?
No. Model performance is one part of the evidence. Device assessment also considers intended users, population, workflow, integrations, risks and clinical/performance evidence. Validation must address the configured device and intended context, not just a development dataset or isolated benchmark. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations MDCG 2020-1: Guidance on Clinical Evaluation (MDR) / Performance Evaluation (IVDR) of Medical Device Software
What is a PCCP?
A Predetermined Change Control Plan describes specified modifications, their development and validation methods, and an impact assessment. In the US, the authorized plan can permit covered modifications without an additional submission for each one. It does not authorize changes outside its scope or replace quality-system controls. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions
Can an AI device learn continuously after deployment?
Automatic adaptation is not unrestricted permission to modify a regulated product. Whether and how updates are permitted depends on device authorization, applicable rules and any authorized change plan. Define boundaries, evaluation, monitoring and intervention rather than assuming all online learning is acceptable. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions
Does every software update require a new submission?
No universal answer applies. Assess the modification, its potential effect on safety/effectiveness or intended use, the device pathway and any authorized PCCP. A change outside a plan does not automatically need a submission, but it does require assessment under the relevant modification rules. Artificial Intelligence-Enabled Medical Devices Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions
What should post-market monitoring include?
Use risk-proportionate measures of safety, performance, compatibility, usage and relevant subgroups, plus complaint and incident processes. Assign owners and response thresholds. Monitoring should support investigation and corrective action; simply plotting input drift is not sufficient evidence of clinical safety. Medical devices: post-market surveillance Why post-market surveillance is a priority for the MHRA AI Airlock programme
Does FDA authorization apply in Europe or the UK?
No single authorization automatically grants worldwide access. US, EU and UK pathways and responsibilities differ, and GB/NI positions require care. International principles can align approaches but do not replace local classification, conformity assessment, registration or surveillance obligations. Artificial Intelligence-Enabled Medical Devices Predetermined Change Control Plans for Machine Learning-Enabled Medical Devices: Guiding Principles AIB 2025-1 / MDCG 2025-6: Interplay between the Medical Devices Regulation (MDR) & In vitro Diagnostic Medical Devices Regulation (IVDR) and the Artificial Intelligence Act (AIA) Medical devices: post-market surveillance
Can generative AI be part of a medical device?
Potentially. Its function and medical purpose determine the assessment. Evaluate output variability, unsupported content, prompts, retrieval sources, external dependencies and user interpretation within the complete device. Do not assume a general-purpose model benchmark establishes safety or effectiveness for a medical application. Artificial Intelligence-Enabled Medical Devices Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations
A defensible AI-device lifecycle connects every medical claim to evidence, every change to a documented decision and every deployed version to monitoring. The objective is not frozen software or uncontrolled improvement; it is justified change within a maintained safety and performance framework.
Discuss your lifecycle evidence and change-control priorities through HERMS Solutions , or explore related HERMS medical-device and life-sciences insights .
Relevant HERMS solutions
Explore HERMS solutions relevant to the workflows discussed in this article: VERIQIS — evidence integrity and assurance; REGAXIS — registration lifecycle intelligence; AXION — manufacturing quality. Evaluate each solution against your intended use, evidence and approval requirements.
Sources and further reading
- Artificial Intelligence-Enabled Medical Devices
- Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions
- MDCG 2020-1: Guidance on Clinical Evaluation (MDR) / Performance Evaluation (IVDR) of Medical Device Software
- Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations
- Predetermined Change Control Plans for Machine Learning-Enabled Medical Devices: Guiding Principles
- Medical devices: post-market surveillance
- Why post-market surveillance is a priority for the MHRA AI Airlock programme
- The Medical Devices (Post-market Surveillance Requirements) (Amendment) (Great Britain) Regulations 2024
- AIB 2025-1 / MDCG 2025-6: Interplay between the Medical Devices Regulation (MDR) & In vitro Diagnostic Medical Devices Regulation (IVDR) and the Artificial Intelligence Act (AIA)
Explore the pharmaceutical AI series
- AI decision authority and human oversight
- Agentic versus generative AI in pharmaceuticals
- AI governance in life sciences
- Human approval controls for pharmaceutical AI
- Real-world evidence for regulatory and market access
- Global pharmaceutical market access, HTA and pricing evidence
Explore HERMS life sciences solutions and AI governance and traceability, or request a workflow-focused demonstration.
