Draft a medical device SaMD/AI compliance checklist and PMS plan.
- User
Draft a medical device SaMD/AI compliance checklist and PMS plan.
- Assistant
SaMD/AI Medical Device Compliance Checklist and Post-Market Surveillance (PMS) Plan
Draft prepared by Sherlock, AI by Trailblazer Labs
Date: [Insert Current Date]
Scope: This checklist and PMS plan are designed for Software as a Medical Device (SaMD) incorporating AI/ML elements. They align with key regulations including EU MDR 2017/745 (Annex I, II, III), FDA SaMD Framework and AI/ML SaMD Action Plan, ISO 13485, IEC 62304, IEC 62366-1, ISO 14971, IMDRF SaMD guidance, and emerging AI-specific guidances (e.g., MDCG 2023-7, FDA Good Machine Learning Practice). Customize for jurisdiction, risk class (e.g., Class I-III or I-IV per MDR), and SaMD intended use. Not legal advice—consult regulatory experts.
1. SaMD/AI Compliance Checklist
Use this checklist during pre-market development, submission (e.g., EU Technical Documentation, FDA 510(k)/De Novo/PMA), and audits. Mark as Compliant (Y/N/NA), with evidence references.
1.1 Regulatory Classification and General Requirements
Item Description Y/N/NA Evidence 1.1.1 SaMD defined per IMDRF/FDA (standalone software with medical purpose; no hardware integration unless specified). 1.1.2 Risk classification completed (e.g., FDA I/II/III; EU Class I/IIa/IIb/III; IMDRF a/b/c/d based on significance of state, info, clinical condition). 1.1.3 AI/ML classification addressed (locked vs. adaptive/continuous learning models). 1.1.4 GSPR/QSR conformity declared (EU Annex I; FDA QSR 21 CFR 820). 1.1.5 UDI assigned and implemented (Unique Device Identification). 1.2 Quality Management System (QMS)
Item Description Y/N/NA Evidence 1.2.1 ISO 13485:2016 QMS certified/implemented (design controls, CAPA, complaints, etc.). 1.2.2 Software lifecycle processes per IEC 62304 (Class A/B/C based on risk). 1.2.3 Change control for AI model updates (versioning, retraining protocols). 1.3 Technical Documentation (EU Annex II/III; FDA Design History File)
Item Description Y/N/NA Evidence 1.3.1 Detailed description of SaMD/AI architecture (algorithms, data inputs/outputs, training/validation datasets). 1.3.2 Bill of materials/software architecture diagram. 1.3.3 Cybersecurity design (encryption, access controls, vulnerability assessments per FDA cybersecurity guidance). 1.3.4 Performance specs, algorithms, and limitations documented. 1.4 Risk Management (ISO 14971:2019)
Item Description Y/N/NA Evidence 1.4.1 Risk Management File (RMF) with hazard analysis, FMEA/FTA for SaMD/AI-specific risks (e.g., bias, overfitting, adversarial attacks). 1.4.2 AI risks mitigated: data bias, dataset diversity, explainability (e.g., SHAP/LIME), robustness testing. 1.4.3 Residual risks acceptable and disclosed in IFU. 1.4.4 Post-market risk integration (benefit-risk analysis). 1.5 Clinical Evaluation (MEDDEV 2.7/1 Rev 4; FDA Clinical Guidance)
Item Description Y/N/NA Evidence 1.5.1 Clinical Evaluation Plan (CEP) and Report (CER) with literature review, clinical investigations, or equivalence. 1.5.2 AI performance metrics: sensitivity/specificity, AUC-ROC, calibration plots; stratified by subgroups (bias analysis). 1.5.3 Real-World Evidence (RWE) plan if applicable for adaptive AI. 1.5.4 Scientific validity, analytical/clinical validation demonstrated. 1.6 Verification, Validation, and Usability
Item Description Y/N/NA Evidence 1.6.1 Software V&V per IEC 62304 (unit/integration/system testing). 1.6.2 AI-specific testing: dataset splits (train/valid/test), cross-validation, external validation sets. 1.6.3 Usability engineering per IEC 62366-1 (formative/summative testing with users). 1.6.4 Simulated use, stress testing, interoperability testing. 1.7 AI/ML-Specific Considerations (FDA GMLP; MDCG AI Guidance)
Item Description Y/N/NA Evidence 1.7.1 Dataset governance: source, size, diversity, preprocessing documented; bias mitigation (e.g., reweighting, augmentation). 1.7.2 Model transparency/explainability (e.g., feature importance). 1.7.3 Algorithm transparency (pre-determined vs. ML; training process). 1.7.4 Change management for model updates (Predetermined Change Control Plan for adaptive AI). 1.7.5 Ethical AI: fairness, accountability documented. 1.8 Labeling, Instructions for Use (IFU), and Promotion
Item Description Y/N/NA Evidence 1.8.1 IFU includes intended use, limitations, warnings (e.g., AI confidence scores, operator training). 1.8.2 Symbols per ISO 15223; multilingual as required. 1.8.3 Promotional claims substantiated (no off-label). 1.9 Pre-Market Submission Readiness
Item Description Y/N/NA Evidence 1.9.1 Notified Body/competent authority engagement (if Class IIa+). 1.9.2 Clinical trials registered (if required, e.g., ClinicalTrials.gov). 1.9.3 PMS/Vigilance plan in place (see Section 2). Completion Notes: Total items compliant: /. Gaps actioned by [date]. Auditor sign-off: ________________.
2. Post-Market Surveillance (PMS) Plan
Document Reference: [e.g., PMS-001 v1.0]
Device: [SaMD/AI Name, Model, UDI] | Risk Class: [e.g., IIb] | PMS Responsible: [QRA/Regulatory Lead]
Review Frequency: Annual (or post-update); integrate with PSUR (EU) / PMSR (FDA).
Objective: Proactively monitor safety, performance, and effectiveness post-market to detect signals, mitigate risks, and support lifecycle changes (MDR Art. 83-86, Annex III; FDA 21 CFR 803/822). For AI: Monitor drift, bias emergence, real-world performance.2.1 PMS Objectives and Scope
- Collect/analyze data on foreseeable risks, usability, clinical performance/benefits.
- Detect new/ emerging risks (esp. AI: dataset shift, adversarial inputs).
- Support PSUR/PMSR, CAPA, field safety notices, recalls.
- Metrics: Complaint rates, adverse events (MAUDE/EUDAMED), performance KPIs (e.g., AUC drift <5%).
2.2 Data Sources and Collection Methods
Data Source Collection Method Frequency Responsible Analysis Tool User complaints/helpdesk Automated ticketing (e.g., Zendesk), hotline Real-time Customer Support Trend analysis (Excel/Tableau) Adverse events/vigilance EUDAMED/FDA MedWatch/MDR reporting Real-time Pharmacovigilance Signal detection software (e.g., Oracle Argus) Registry/RWE studies EHR integration, patient registries Quarterly Clinical/HEOR Statistical software (R/Python) Literature/scientific PubMed/Google Scholar alerts Monthly Medical Affairs Review database User analytics/telemetry Usage logs, error rates, AI confidence scores Continuous Data Analytics ML monitoring (e.g., Evidently AI) Audits/capabilities Supplier audits, cybersecurity scans Annual QMS Audit reports PSUR inputs Sales/volume data, market feedback Annual Regulatory PMS database 2.3 Signal Detection and Risk Assessment
- Thresholds: Complaint rate >0.1%; performance degradation >10%; new risk probability >1/10,000.
- Methods: Statistical process control (SPC), Bayesian anomaly detection for AI drift; FMEA refresh annually.
- Escalation: Signals → Risk Review Board (within 7 days) → CAPA if confirmed.
- AI-Specific: Monitor input data drift (e.g., KS-test), model retraining triggers.
2.4 Reporting and Communication
Report Content Frequency Recipients PMS Report Data summary, signals, actions Annual NB/RA/Management PSUR (EU Class IIa+) Benefit-risk update, clinical data Every 2 years (post-CE mark) NB/EUDAMED Field Safety Notice Confirmed issues As needed Users/RA Trend Report AI performance KPIs Quarterly Internal 2.5 Resource Allocation and Training
- Team: QRA (lead), Data Scientist (AI), Clinical Expert.
- Budget: [e.g., $XXK/year for tools/studies].
- Training: Annual PMS/AI ethics for staff; SOPs: PMS-001, AI-MON-001.
2.6 Plan Review and Updates
- Review annually or post-model update/change.
- Metrics for PMS effectiveness: Signal detection rate >95%, time-to-resolution <30 days.
- Approval: ________________ Date: __________ Next Review: [Date +1yr].
Annexes: [PMS SOPs, Data Collection Forms, Signal Log Template].
This draft provides a robust starting point. For implementation, integrate into your QMS and validate with a mock audit. Questions? Provide specifics for tailoring (e.g., jurisdiction, AI type).
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.