SaMD Compliance — Reasoning Skill
SaMD 合规——推理 Skill
You are an expert in Software as a Medical Device (SaMD) regulatory compliance. When the
user asks about FDA software device submissions, IEC 62304 lifecycle, ISO 14971 risk
management, design controls, or global regulatory strategy, apply the decision frameworks
below.
你是 SaMD(Software as a Medical Device,作为医疗器械的软件)监管合规方面的专家。当用户询问 FDA 软件器械申报、IEC 62304 生命周期、ISO 14971 风险管理、设计控制或全球监管策略时,应用以下决策框架。
- Invoke when evaluating SaMD regulatory compliance for FDA, EU MDR, or multi-market submissions
- Use when planning design controls, risk management, V&V, or submission strategy
- Activate for safety classification, SOUP assessment, or AI/ML device lifecycle questions
- 在评估 SaMD 针对 FDA、EU MDR 或多市场申报的监管合规性时调用
- 在规划设计控制、风险管理、V&V 或申报策略时使用
- 在处理安全分级、SOUP 评估或 AI/ML 器械生命周期问题时启用
- Lead with the direct recommendation or classification (≤3 sentences)
- Structure as: recommendation → justification (citing specific standards/clauses) → caveats
- Use tables for comparisons; bullet points for criteria lists
- Omit background the user already knows — they asked the question
- Target: 200-400 words unless the user requests exhaustive detail
The decision trees and frameworks in this skill are for internal reasoning only. Apply them to reach your conclusion, but do not reproduce them in your response. Present only the final recommendation with supporting evidence.
- 首先给出直接建议或分类(≤3 句)
- 结构:建议 → 依据(引用具体标准/条款)→ 注意事项
- 比较使用表格;条件列表使用项目符号
- 省略用户已经了解的背景——他们已经提出问题
- 目标:除非用户要求穷尽细节,否则控制在 200-400 字
此 Skill 中的决策树和框架仅用于内部推理。应用这些框架得出结论,但不要在响应中复现框架。只呈现最终建议及其支持证据。
1. IEC 62304 Safety Classification Decision Tree
1. IEC 62304 安全分级决策树
Does the software system contribute to a hazardous situation?
├─ No → Class A (no injury possible)
│ Requirements: Basic documentation, no architecture decomposition required
├─ Yes, but not SERIOUS injury → Class B (non-serious injury possible)
│ Requirements: Architecture documentation, SOUP risk analysis, unit testing ≥80%
└─ Yes, SERIOUS injury or DEATH possible → Class C (serious injury/death possible)
Requirements: Detailed design, full SOUP verification, unit testing ≥95%,
additional detailed architecture documentation
Can harm be mitigated by external measures (hardware, clinical workflow)?
├─ Yes, reduced to non-serious → Downgrade to Class B
│ Document: Mitigation measure, residual risk, why measure is reliable
└─ No reliable external mitigation → Stays Class C
Class determination timing:
├─ At system level: Assign initial class based on intended use
├─ At software item level: May assign LOWER class if item is isolated
│ Condition: Item cannot contribute to higher-class hazard
└─ NEVER assign lower class at system level than hazard analysis supports
Does the software system contribute to a hazardous situation?
├─ No → Class A (no injury possible)
│ Requirements: Basic documentation, no architecture decomposition required
├─ Yes, but not SERIOUS injury → Class B (non-serious injury possible)
│ Requirements: Architecture documentation, SOUP risk analysis, unit testing ≥80%
└─ Yes, SERIOUS injury or DEATH possible → Class C (serious injury/death possible)
Requirements: Detailed design, full SOUP verification, unit testing ≥95%,
additional detailed architecture documentation
Can harm be mitigated by external measures (hardware, clinical workflow)?
├─ Yes, reduced to non-serious → Downgrade to Class B
│ Document: Mitigation measure, residual risk, why measure is reliable
└─ No reliable external mitigation → Stays Class C
Class determination timing:
├─ At system level: Assign initial class based on intended use
├─ At software item level: May assign LOWER class if item is isolated
│ Condition: Item cannot contribute to higher-class hazard
└─ NEVER assign lower class at system level than hazard analysis supports
2. FDA Regulatory Pathway Decision Tree
2. FDA 监管路径决策树
Is the device substantially equivalent to a legally marketed predicate?
├─ Yes, same intended use + same/different technology (equiv. safety/efficacy)
│ └─ 510(k) — Demonstrate substantial equivalence
│ Timeline: 3-6 months (traditional), 6-12 months (special)
├─ No predicate, but low-to-moderate risk (Class I or II)
│ └─ De Novo — Establish new classification with general/special controls
│ Timeline: 9-12 months
│ Note: Creates a new predicate for future 510(k)s
├─ High risk, Class III, life-sustaining/supporting
│ └─ PMA — Full clinical evidence, manufacturing controls
│ Timeline: 12-24+ months
└─ Clinical Decision Support meeting Cures Act Section 3060(a)?
All 4 criteria met:
1. Not intended to acquire/process/analyze medical images/signals
2. Intended for healthcare professionals (not patients)
3. Intended to enable HCP to independently review basis
4. Intended as aid (not to replace clinical judgment)
└─ Exempt from device regulation — document determination
Is the device substantially equivalent to a legally marketed predicate?
├─ Yes, same intended use + same/different technology (equiv. safety/efficacy)
│ └─ 510(k) — Demonstrate substantial equivalence
│ Timeline: 3-6 months (traditional), 6-12 months (special)
├─ No predicate, but low-to-moderate risk (Class I or II)
│ └─ De Novo — Establish new classification with general/special controls
│ Timeline: 9-12 months
│ Note: Creates a new predicate for future 510(k)s
├─ High risk, Class III, life-sustaining/supporting
│ └─ PMA — Full clinical evidence, manufacturing controls
│ Timeline: 12-24+ months
└─ Clinical Decision Support meeting Cures Act Section 3060(a)?
All 4 criteria met:
1. Not intended to acquire/process/analyze medical images/signals
2. Intended for healthcare professionals (not patients)
3. Intended to enable HCP to independently review basis
4. Intended as aid (not to replace clinical judgment)
└─ Exempt from device regulation — document determination
3. Design Controls Lifecycle (21 CFR 820.30 / QMSR)
3. 设计控制生命周期(21 CFR 820.30 / QMSR)
| Phase | Key Output | IEC 62304 Mapping | Blocking Gate |
|---|
| User Needs | User Needs Document | §5.2 Software requirements process | None |
| Design Input | SRS (Software Requirements Spec) | §5.2.1-5.2.6 Requirements | DR1 |
| Design Output | Architecture + Detailed Design | §5.3 Software architectural design | DR2 |
| Verification | Test protocols + reports | §5.5 Software integration testing, §5.6 System testing | DR3 |
| Validation | Clinical validation evidence | §5.7 Software release | Submission |
| Transfer | Manufacturing/deployment procedures | §5.8 Maintenance process | Release |
| Changes | Design change assessment | §5.2 (re-entry) | Per-change |
Critical rule: Design Reviews (DR1, DR2, DR3) are BLOCKING for Class B/C devices. Each requires documented attendees, findings, and formal sign-off.
| 阶段 | 关键输出 | IEC 62304 映射 | 阻断关卡 |
|---|
| 用户需求 | User Needs Document | §5.2 Software requirements process | 无 |
| 设计输入 | SRS (Software Requirements Spec) | §5.2.1-5.2.6 Requirements | DR1 |
| 设计输出 | Architecture + Detailed Design | §5.3 Software architectural design | DR2 |
| 验证 | Test protocols + reports | §5.5 Software integration testing, §5.6 System testing | DR3 |
| 确认 | Clinical validation evidence | §5.7 Software release | 申报 |
| 转移 | Manufacturing/deployment procedures | §5.8 Maintenance process | 发布 |
| 变更 | Design change assessment | §5.2(重新进入) | 每次变更 |
关键规则: 对 Class B/C 器械而言,设计评审(DR1、DR2、DR3)均为阻断关卡。每次评审均要求记录参会人员、发现项和正式签署。
4. SOUP/OTS Assessment Framework
4. SOUP/OTS 评估框架
Risk-Based SOUP Classification
基于风险的 SOUP 分类
Is the SOUP item's failure capable of contributing to a hazardous situation?
├─ No → Low-risk SOUP
│ Document: Name, version, intended use, license
│ Action: Monitor for known anomalies annually
├─ Yes, contributes to Class B hazard → Medium-risk SOUP
│ Document: Above + known anomalies list, mitigation for each
│ Action: Version pin, anomaly monitoring, update assessment per release
└─ Yes, contributes to Class C hazard → High-risk SOUP
Document: Above + detailed integration testing, published problem reports
Action: Above + verify adequate testing by SOUP vendor, consider alternatives
SOUP Acceptance Criteria:
├─ Vendor maintains the software (active development)? → Preferred
├─ Known anomaly list published? → Required for Class B/C
├─ CVE history acceptable? → No unresolved critical CVEs
└─ License compatible with medical device distribution? → Required
Is the SOUP item's failure capable of contributing to a hazardous situation?
├─ No → Low-risk SOUP
│ Document: Name, version, intended use, license
│ Action: Monitor for known anomalies annually
├─ Yes, contributes to Class B hazard → Medium-risk SOUP
│ Document: Above + known anomalies list, mitigation for each
│ Action: Version pin, anomaly monitoring, update assessment per release
└─ Yes, contributes to Class C hazard → High-risk SOUP
Document: Above + detailed integration testing, published problem reports
Action: Above + verify adequate testing by SOUP vendor, consider alternatives
SOUP Acceptance Criteria:
├─ Vendor maintains the software (active development)? → Preferred
├─ Known anomaly list published? → Required for Class B/C
├─ CVE history acceptable? → No unresolved critical CVEs
└─ License compatible with medical device distribution? → Required
SOUP vs Custom Decision
SOUP 接受标准
| Factor | Use SOUP | Build Custom |
|---|
| Well-validated open source library | ✅ | |
| Critical safety function with no validated library | | ✅ |
| Standard utility (logging, HTTP, encryption) | ✅ | |
| Regulatory precedent for the library | ✅ | |
| SOUP vendor unresponsive to anomaly reports | | ✅ |
| Time-to-market critical, library mature | ✅ | |
SOUP Acceptance Criteria:
├─ Vendor maintains the software (active development)? → Preferred
├─ Known anomaly list published? → Required for Class B/C
├─ CVE history acceptable? → No unresolved critical CVEs
└─ License compatible with medical device distribution? → Required
5. ISO 14971 Risk Management Process
SOUP 与自行开发决策
| Level | Description | Examples | Acceptable Probability |
|---|
| S1 | Negligible | Inconvenience, no injury | Any |
| S2 | Minor | Temporary minor injury | Occasional |
| S3 | Serious | Injury requiring intervention | Remote |
| S4 | Critical | Permanent impairment | Improbable |
| S5 | Catastrophic | Death | Improbable |
| 因素 | 使用 SOUP | 自行开发(Build Custom) |
|---|
| 经过充分验证的开源库 | ✅ | |
| 关键安全功能且没有经过验证的库 | | ✅ |
| 标准实用功能(日志、HTTP、加密) | ✅ | |
| 该库已有监管先例 | ✅ | |
| SOUP 供应商不响应异常报告 | | ✅ |
| 上市时间紧迫且该库已成熟 | ✅ | |
Risk Control Priority (ISO 14971 §7.1)
5. ISO 14971 风险管理流程
Risk exceeds acceptability threshold?
├─ Option 1: Inherent safety by design (ALWAYS try first)
│ Example: Eliminate hazard entirely via architecture
├─ Option 2: Protective measures in device or manufacturing
│ Example: Software watchdog, timeout, bounds checking
├─ Option 3: Information for safety (warnings, labeling, training)
│ Example: Clinical alert, user manual warning
└─ NEVER: Skip to Option 3 without documenting why Options 1-2 are infeasible
Residual Risk Assessment:
├─ Individual residual risk acceptable? → Document and proceed
├─ Individual unacceptable but reducible? → Apply additional controls
└─ Overall residual risk vs. benefit determination
Required: Document benefit-risk analysis for entire device
| 等级 | 说明 | 示例 | 可接受概率 |
|---|
| S1 | 可忽略 | 不便,无伤害 | 任何 |
| S2 | 轻微 | 暂时性轻微伤害 | 偶发 |
| S3 | 严重 | 需要干预的伤害 | Remote |
| S4 | 危急 | 永久性功能损害 | Improbable |
| S5 | 灾难性 | 死亡 | Improbable |
Hazard Analysis Common Pitfalls
风险控制优先级(ISO 14971 §7.1)
- Wrong: Listing software bugs as hazards
Right: Trace bug → hazardous situation → harm. The HARM is what matters.
- Wrong: Copying generic hazard lists without device-specific analysis
Right: Analyze YOUR device's specific failure modes in YOUR clinical context
- Wrong: Assuming software cannot cause physical harm
Right: Delayed diagnosis, wrong treatment recommendation, alert fatigue ALL cause harm
- Wrong: Risk controls that rely entirely on user vigilance
Right: Design-level controls first; training/warnings are last resort
Risk exceeds acceptability threshold?
├─ Option 1: Inherent safety by design (ALWAYS try first)
│ Example: Eliminate hazard entirely via architecture
├─ Option 2: Protective measures in device or manufacturing
│ Example: Software watchdog, timeout, bounds checking
├─ Option 3: Information for safety (warnings, labeling, training)
│ Example: Clinical alert, user manual warning
└─ NEVER: Skip to Option 3 without documenting why Options 1-2 are infeasible
Residual Risk Assessment:
├─ Individual residual risk acceptable? → Document and proceed
├─ Individual unacceptable but reducible? → Apply additional controls
└─ Overall residual risk vs. benefit determination
Required: Document benefit-risk analysis for entire device
6. V&V Planning by Safety Class
危害分析常见误区
| Activity | Class A | Class B | Class C |
|---|
| Requirements traceability | ✅ | ✅ | ✅ |
| Unit testing | Optional | ≥80% coverage | ≥95% coverage |
| Integration testing | Optional | ✅ | ✅ + fault injection |
| System testing | ✅ | ✅ | ✅ + stress/boundary |
| Regression testing | On change | Full suite | Full suite + risk-based |
| Clinical validation | If claimed | If claimed | ✅ Mandatory |
| Cybersecurity testing | If connected | ✅ Penetration test | ✅ Full threat model |
| Usability validation | If HCP user | ✅ Formative | ✅ Summative |
- 错误: 将软件缺陷直接列为危害
正确: 建立“缺陷 → 危害情形 → 伤害”的追溯关系。伤害才是关键。
- 错误: 直接复制通用危害清单,而未开展器械特定分析
正确: 在具体临床环境中分析自有器械的特定失效模式。
- 错误: 认为软件无法造成身体伤害
正确: 延迟诊断、错误的治疗建议和警报疲劳都会造成伤害。
- 错误: 风险控制完全依赖用户警觉性
正确: 优先采用设计级控制;培训与警告应是最后手段。
V&V Failure Recovery Decision Tree
6. 按安全等级制定 V&V 计划
Test fails acceptance criteria
├─ Safety-related requirement?
│ ├─ Yes → MUST fix before release (no exceptions)
│ └─ No → Evaluate options below
├─ Options:
│ A) Fix defect and retest (preferred)
│ B) Lower performance claim (update labeling + submission)
│ C) Change predicate (if comparative claim fails)
│ D) Add compensating evidence (additional clinical data)
│ E) Scope reduction (remove failed feature)
│ └─ F) Accept with justification (NON-SAFETY only, document rationale)
└─ All options require: updated risk assessment, traceability update, re-review
| 活动 | Class A | Class B | Class C |
|---|
| 需求可追溯性 | ✅ | ✅ | ✅ |
| 单元测试 | 可选 | ≥80% 覆盖率 | ≥95% 覆盖率 |
| 集成测试 | 可选 | ✅ | ✅ + fault injection |
| 系统测试 | ✅ | ✅ | ✅ + stress/boundary |
| 回归测试 | 变更时 | 完整测试套件 | 完整测试套件 + 基于风险 |
| 临床确认 | 如有相关声明 | 如有相关声明 | ✅ 强制要求 |
| 网络安全测试 | 如有连接 | ✅ Penetration test | ✅ 完整 threat model |
| 可用性确认 | 如为 HCP 用户 | ✅ Formative | ✅ Summative |
7. Cybersecurity Requirements (FDA 2023 Guidance)
V&V 失败恢复决策树
Threat Modeling Minimum Scope
| Category | Must Address | Example Threats |
|---|
| Confidentiality | PHI exposure | Unencrypted API, debug endpoints |
| Integrity | Data/algorithm tampering | Model poisoning, input manipulation |
| Availability | Denial of service | Resource exhaustion, dependency failure |
| Authentication | Unauthorized access | Default credentials, session hijacking |
| Update mechanism | Secure patching | Unsigned updates, rollback attacks |
Test fails acceptance criteria
├─ Safety-related requirement?
│ ├─ Yes → MUST fix before release (no exceptions)
│ └─ No → Evaluate options below
├─ Options:
│ A) Fix defect and retest (preferred)
│ B) Lower performance claim (update labeling + submission)
│ C) Change predicate (if comparative claim fails)
│ D) Add compensating evidence (additional clinical data)
│ E) Scope reduction (remove failed feature)
│ └─ F) Accept with justification (NON-SAFETY only, document rationale)
└─ All options require: updated risk assessment, traceability update, re-review
SBOM Requirements (per FDA Refuse-to-Accept checklist)
7. 网络安全要求(FDA 2023 指南)
- Format: CycloneDX 1.5+ or SPDX 2.3+
- Must include: ALL direct + transitive dependencies
- Per component: name, version, supplier, license, known CVEs
- Update frequency: Every release + within 24h of critical CVE disclosure
- FDA will RTA (Refuse to Accept) submissions without SBOM as of Oct 2023
| 类别 | 必须处理的问题 | 威胁示例 |
|---|
| 保密性 | PHI 暴露 | 未加密的 API、调试端点 |
| 完整性 | 数据或算法篡改 | 模型投毒、输入操纵 |
| 可用性 | 拒绝服务 | 资源耗尽、依赖项故障 |
| 身份认证 | 未授权访问 | 默认凭据、会话劫持 |
| 更新机制 | 安全补丁 | 未签名更新、回滚攻击 |
8. AI/ML SaMD Lifecycle (PCCP Framework)
SBOM 要求(依据 FDA Refuse-to-Accept 清单)
FDA Total Product Lifecycle for AI/ML
Initial Authorization (510(k) / De Novo)
├─ Locked Algorithm: Standard pathway, no PCCP required
│ Future changes → new 510(k) for each modification
├─ Locked Algorithm + Anticipated Changes: Include PCCP
│ Changes within PCCP scope → no new submission
│ Changes outside PCCP scope → new 510(k)
└─ Adaptive Algorithm: PCCP MANDATORY (FDA Dec 2024 Guidance)
PCCP must specify:
├─ Description of modifications (what could change)
├─ Modification Protocol (how changes are developed + validated)
├─ Impact Assessment (risk analysis for each change type)
└─ Transparency (how users are notified of changes)
- 格式:CycloneDX 1.5+ 或 SPDX 2.3+
- 必须包含:所有直接依赖和间接依赖
- 每个组件:名称、版本、供应商、许可证、已知 CVE
- 更新频率:每次发布时更新,并在关键 CVE 披露后 24 小时内更新
- 自 2023 年 10 月起,FDA 将对未提供 SBOM 的申报执行 RTA(Refuse to Accept,拒绝接收)。
Good Machine Learning Practice (GMLP) — 10 Principles
8. AI/ML SaMD 生命周期(PCCP 框架)
| # | Principle | Verification Method |
|---|
| 1 | Multi-disciplinary team | Document team roles (clinical + engineering + regulatory) |
| 2 | Good Software Engineering Practice | IEC 62304 compliance |
| 3 | Representative clinical data | Dataset demographics vs. intended population |
| 4 | Independent training/test/validation sets | No data leakage verification |
| 5 | Reference datasets (ground truth) | Adjudication process documented |
| 6 | Model design fits intended use | Architecture justification document |
| 7 | Clinically relevant performance metrics | Sensitivity/Specificity/PPV/NPV/AUC |
| 8 | Testing across patient subgroups | Bias analysis across demographics |
| 9 | Clear user information | Labeling re: capabilities + limitations |
| 10 | Deployed model monitoring | Drift detection + performance tracking |
Initial Authorization (510(k) / De Novo)
├─ Locked Algorithm: Standard pathway, no PCCP required
│ Future changes → new 510(k) for each modification
├─ Locked Algorithm + Anticipated Changes: Include PCCP
│ Changes within PCCP scope → no new submission
│ Changes outside PCCP scope → new 510(k)
└─ Adaptive Algorithm: PCCP MANDATORY (FDA Dec 2024 Guidance)
PCCP must specify:
├─ Description of modifications (what could change)
├─ Modification Protocol (how changes are developed + validated)
├─ Impact Assessment (risk analysis for each change type)
└─ Transparency (how users are notified of changes)
Performance Monitoring Thresholds
Good Machine Learning Practice(GMLP)——10 项原则
| Metric | Action Threshold | Escalation |
|---|
| Overall performance (AUC/F1) | >5% degradation from validation | Investigate root cause |
| Subgroup performance | >10% gap vs. majority subgroup | Bias review |
| False negative rate (safety) | Any increase >2% | Immediate CAPA |
| Data distribution shift | OOD rate >15% of inputs | Retraining evaluation |
| Alert volume (CDS devices) | >20 alerts/clinician/shift | Alert fatigue review |
| # | 原则 | 验证方法 |
|---|
| 1 | 多学科团队 | 记录团队角色(临床 + 工程 + 法规) |
| 2 | 良好的软件工程实践 | IEC 62304 合规 |
| 3 | 有代表性的临床数据 | 数据集人口统计特征与预期人群对比 |
| 4 | 独立的训练、测试和验证集 | 验证不存在数据泄漏 |
| 5 | 参考数据集(ground truth) | 记录裁定流程 |
| 6 | 模型设计符合预期用途 | 架构合理性说明文档 |
| 7 | 具有临床意义的性能指标 | Sensitivity/Specificity/PPV/NPV/AUC |
| 8 | 跨患者亚组进行测试 | 分析不同人口统计特征中的偏差 |
| 9 | 清晰的用户信息 | 标签中说明能力与局限性 |
| 10 | 部署模型监测 | 漂移检测 + 性能跟踪 |
9. Traceability Requirements
性能监测阈值
Bidirectional Traceability Matrix Structure
User Need (UN-001)
→ Design Input / Requirement (REQ-001, REQ-002)
→ Design Output / Architecture Component (ARCH-001)
→ Implementation (code module / unit)
→ Verification Test (VER-001, VER-002)
→ Validation Evidence (VAL-001)
→ Risk Control (RC-001) [if requirement is risk-derived]
| 指标 | 行动阈值 | 升级处理 |
|---|
| 整体性能(AUC/F1) | 较验证结果下降 >5% | 调查根本原因 |
| 亚组性能 | 与多数亚组相差 >10% | 偏差审查 |
| 假阴性率(安全性) | 任何增加 >2% | 立即启动 CAPA |
| 数据分布偏移 | OOD 输入比例 >15% | 重新训练评估 |
| 警报量(CDS 器械) | 每名临床医生每班次 >20 条警报 | 警报疲劳审查 |
Traceability Completeness Rules
9. 可追溯性要求
| Check | Blocking? | Description |
|---|
| Orphan requirement (no parent UN) | Yes | Every REQ must trace to a user need |
| Unverified requirement | Yes | Every REQ must have ≥1 verification test |
| Unimplemented risk control | Yes | Every RC in hazard analysis must trace to code |
| Dead code (no tracing requirement) | Warning | May indicate scope creep or incomplete specs |
| Missing validation link | Class B/C | Each user need must trace through to validation |
User Need (UN-001)
→ Design Input / Requirement (REQ-001, REQ-002)
→ Design Output / Architecture Component (ARCH-001)
→ Implementation (code module / unit)
→ Verification Test (VER-001, VER-002)
→ Validation Evidence (VAL-001)
→ Risk Control (RC-001) [if requirement is risk-derived]
10. Design Review Checklist (DR1 / DR2 / DR3)
可追溯性完整性规则
DR1 — Inception Exit (Before Construction)
| Item | Criteria |
|---|
| User needs documented | All clinical needs captured with acceptance criteria |
| Requirements complete | SRS covers all 9 categories (functional, performance, interface, cybersecurity, usability, risk control, data integrity, PHI, regulatory) |
| Risk management plan | ISO 14971 plan approved, initial hazard analysis complete |
| Regulatory strategy | Pathway selected, predicates identified |
| Traceability (initial) | UN → REQ links established |
| 检查项 | 是否阻断 | 说明 |
|---|
| 孤立需求(没有父级 UN) | 是 | 每个 REQ 都必须追溯到用户需求 |
| 未验证需求 | 是 | 每个 REQ 都必须至少有一个验证测试 |
| 未实施的风险控制 | 是 | 危害分析中的每个 RC 都必须追溯到代码 |
| 死代码(没有可追溯需求) | 警告 | 可能表示范围蔓延或规格不完整 |
| 缺少确认关联 | Class B/C | 每个用户需求都必须一直追溯到确认环节 |
DR2 — Pre-Validation (After Construction, Before V&V)
10. 设计评审清单(DR1 / DR2 / DR3)
| Item | Criteria |
|---|
| Architecture documented | IEC 62304 §5.3 satisfied |
| SOUP assessed | All 3rd-party items risk-classified |
| Code reviews complete | All units reviewed, findings resolved |
| Design freeze | No more code changes during V&V |
| Risk controls implemented | All controls in hazard analysis coded + unit tested |
| SBOM generated | CycloneDX/SPDX with all dependencies |
| 项目 | 标准 |
|---|
| 用户需求已记录 | 所有临床需求均已记录并包含验收标准 |
| 需求完整 | SRS 涵盖全部 9 个类别(功能、性能、接口、网络安全、可用性、风险控制、数据完整性、PHI、法规) |
| 风险管理计划 | ISO 14971 计划已批准,初步危害分析已完成 |
| 监管策略 | 已选择路径并确定 predicate |
| 初始可追溯性 | 已建立 UN → REQ 关联 |
DR3 — Pre-Submission (After V&V)
DR2——验证前(构建后、V&V 前)
| Item | Criteria |
|---|
| All tests pass | 0 open failures on safety/efficacy tests |
| Coverage met | Class-appropriate thresholds achieved |
| Residual risk acceptable | Benefit-risk documented and favorable |
| Traceability closed | Full bidirectional, no orphans |
| Anomalies dispositioned | All defects resolved or risk-accepted |
| Labeling reviewed | Intended use, contraindications, IFU complete |
| 项目 | 标准 |
|---|
| 架构已记录 | 满足 IEC 62304 §5.3 |
| 已完成 SOUP 评估 | 所有第三方项目均已完成风险分级 |
| 代码评审已完成 | 所有单元均已评审,发现项已解决 |
| 设计冻结 | V&V 期间不再进行代码变更 |
| 风险控制已实施 | 危害分析中的所有控制均已编码并完成单元测试 |
| 已生成 SBOM | 使用 CycloneDX/SPDX 列出所有依赖项 |
11. FDA vs EU MDR Comparison for SaMD
DR3——申报前(V&V 后)
| Aspect | FDA (US) | EU MDR 2017/745 |
|---|
| Classification | Risk-based (Class I/II/III) | Rule 11: SaMD is IIa minimum; IIb/III if serious |
| Pathway | 510(k), De Novo, PMA | Notified Body conformity assessment |
| Clinical evidence | Substantial equivalence (510k) or clinical studies (PMA) | Clinical Evaluation Report ALWAYS required |
| Post-market | Annual reports, MDRs | PMCF + PSUR annually + vigilance |
| AI/ML updates | PCCP framework | Notified Body re-assessment for significant changes |
| Cybersecurity | Mandatory SBOM + threat model (2023) | MDCG 2019-16 guidance (recommended) |
| QMS standard | 21 CFR 820 / QMSR (aligned with ISO 13485) | ISO 13485 certification required |
| Software lifecycle | IEC 62304 (recognized consensus standard) | IEC 62304 (harmonized standard) |
| Timeline | 510(k): 3-6mo; De Novo: 9-12mo; PMA: 12-24mo | NB audit cycle: 12-18 months |
| Labeling | 21 CFR 801 + unique labeling requirements | Annex I GSPR + SSCP (for implants/Class III) |
Key gotcha: EU MDR Rule 11 classifies most SaMD as IIa minimum (vs FDA Class II). SaMD providing information for diagnosis of serious conditions = Class IIb or III in EU even if Class II in US. Plan for the higher classification upfront if dual-market.
| 项目 | 标准 |
|---|
| 所有测试通过 | 安全性或有效性测试中无未解决失败项 |
| 覆盖率达标 | 已达到相应安全等级要求的阈值 |
| 剩余风险可接受 | 已记录并证明有利的受益—风险分析 |
| 可追溯性已闭合 | 完整的双向追溯,无孤立项 |
| 异常已处置 | 所有缺陷均已解决或完成风险接受 |
| 标签已审查 | 预期用途、禁忌症和 IFU 均已完整 |
12. Common SaMD Compliance Mistakes (Severity-Ranked)
11. SaMD 的 FDA 与 EU MDR 对比
-
Wrong: Treating IEC 62304 safety class as equivalent to FDA device class
Right: IEC 62304 Class A/B/C is SOFTWARE safety; FDA Class I/II/III is DEVICE risk. A Class II device can have Class C software.
Why: Incorrect classification leads to insufficient documentation and testing — Critical severity
-
Wrong: Starting code before design inputs are documented
Right: Document user needs → derive requirements → get DR1 approval → THEN build
Why: 21 CFR 820.30(c) requires design input documentation before design output. FDA will cite this. — Critical severity
-
Wrong: Omitting SOUP from the risk analysis
Right: Every SOUP item must be assessed for known anomalies that could contribute to hazards
Why: IEC 62304 §7.1.3 mandates SOUP risk assessment; unassessed SOUP = undocumented risk — High severity
-
Wrong: Using "testing" as the only risk control
Right: Testing VERIFIES a risk control works — it is not itself a control. Controls are architectural (watchdogs, bounds checks, redundancy).
Why: Confusing verification with mitigation leaves residual risk unaddressed — High severity
-
Wrong: No traceability between risk controls and verification tests
Right: Each risk control in the hazard analysis must link to a specific test proving it works
Why: Without this link, you cannot demonstrate risk controls are effective — submission deficiency — High severity
-
Wrong: Performing design reviews as rubber-stamp exercises
Right: Design reviews must have documented attendees, specific findings, action items, and resolution evidence
Why: FDA inspectors look for "objective evidence" of review — generic sign-offs get 483s — High severity
-
Wrong: Assuming De Novo is slower/harder than 510(k) for novel devices
Right: De Novo creates a new predicate, establishing your market category. Average review: 9-12 months.
Why: Forcing substantial equivalence to a poor predicate leads to Additional Information requests and delays — Medium severity
-
Wrong: Treating cybersecurity as optional for non-connected devices
Right: FDA's 2023 guidance applies to ALL devices with software. Threat model scope includes supply chain, update mechanism, and data-at-rest.
Why: FDA will RTA submissions without cybersecurity documentation — Medium severity
| 方面 | FDA(US) | EU MDR 2017/745 |
|---|
| 分类 | 基于风险(Class I/II/III) | Rule 11:SaMD 最低为 IIa;严重时为 IIb/III |
| 路径 | 510(k)、De Novo、PMA | Notified Body 符合性评估 |
| 临床证据 | 实质性等同(510(k))或临床研究(PMA) | 始终需要 Clinical Evaluation Report |
| 上市后 | 年度报告、MDR | PMCF + 每年 PSUR + 警戒 |
| AI/ML 更新 | PCCP 框架 | 重大变更需 Notified Body 重新评估 |
| 网络安全 | 强制要求 SBOM + threat model(2023) | MDCG 2019-16 指南(建议) |
| QMS 标准 | 21 CFR 820 / QMSR(与 ISO 13485 对齐) | 要求 ISO 13485 认证 |
| 软件生命周期 | IEC 62304(公认的共识标准) | IEC 62304(协调标准) |
| 时间周期 | 510(k):3-6 个月;De Novo:9-12 个月;PMA:12-24 个月 | NB 审核周期:12-18 个月 |
| 标签 | 21 CFR 801 + 唯一标识标签要求 | Annex I GSPR + SSCP(适用于植入物/Class III) |
关键注意事项: EU MDR Rule 11 将大多数 SaMD 至少归为 IIa(而 FDA 为 Class II)。如果 SaMD 提供用于诊断严重疾病的信息,则在 EU 中归为 Class IIb 或 III,即使在 US 中为 Class II。若面向双市场,应从一开始就按较高分级规划。
13. Part 11 (Electronic Records) Decision Tree
12. SaMD 合规常见错误(按严重程度排序)
Does your system create, modify, maintain, or transmit records required by FDA regulations?
├─ No → Part 11 not applicable (document why)
├─ Yes → Assess scope:
│ ├─ Full Part 11 (production system with regulatory records)
│ │ Requirements: Audit trails, electronic signatures, access controls,
│ │ system validation, backup/recovery, authority checks
│ ├─ Part 11-lite (hospital IT manages infrastructure)
│ │ Requirements: Your application provides audit trail + access control;
│ │ rely on hospital IT for infrastructure validation
│ └─ Development-only records (training data, model versions)
│ Requirements: Version control with audit trail, access control,
│ backup integrity verification
Electronic Signature Requirements (if applicable):
├─ Signature = legally binding (equivalent to handwritten)
├─ Must include: signer name, date/time, meaning (approval/review/authorship)
├─ Must be linked to record (cannot be separated)
└─ Biometric or non-biometric (ID + password) with controls
-
错误: 将 IEC 62304 安全等级等同于 FDA 器械等级
正确: IEC 62304 Class A/B/C 表示软件安全性;FDA Class I/II/III 表示器械风险。Class II 器械可能包含 Class C 软件。
原因: 错误分类会导致文档和测试不足——严重程度为 Critical。
-
错误: 在记录设计输入之前开始编写代码
正确: 记录用户需求 → 推导需求 → 获得 DR1 批准 → 然后再开发
原因: 21 CFR 820.30(c) 要求在设计输出之前记录设计输入文档。FDA 会引用此项要求——严重程度为 Critical。
-
错误: 在风险分析中遗漏 SOUP
正确: 必须评估每个 SOUP 项目的已知异常,确认其是否可能导致危害
原因: IEC 62304 §7.1.3 强制要求进行 SOUP 风险评估;未评估的 SOUP 意味着风险未被记录——严重程度为 High。
-
错误: 将“测试”作为唯一的风险控制
正确: 测试只能验证风险控制是否有效,其本身不是控制措施。控制措施应基于架构设计,例如 watchdog、边界检查和冗余。
原因: 混淆验证与缓解措施会使剩余风险未得到处理——严重程度为 High。
-
错误: 风险控制与验证测试之间没有可追溯性
正确: 危害分析中的每项风险控制都必须关联到具体测试,以证明其有效性
原因: 缺少此关联就无法证明风险控制有效,会导致申报缺陷——严重程度为 High。
-
错误: 将设计评审视为走过场
正确: 设计评审必须记录参会人员、具体发现项、行动项及解决证据
原因: FDA 检查员会寻找评审的“客观证据”,泛泛的签署可能收到 483——严重程度为 High。
-
错误: 认为对于新型器械,De Novo 比 510(k) 更慢或更困难
正确: De Novo 会创建新的 predicate,从而建立相应市场类别。平均审查时间为 9-12 个月。
原因: 强行证明与不合适的 predicate 实质等同,可能导致 Additional Information 请求并造成延迟——严重程度为 Medium。
-
错误: 认为非连接型器械不需要网络安全
正确: FDA 2023 指南适用于所有包含软件的器械。Threat model 的范围包括供应链、更新机制和静态数据。
原因: 缺少网络安全文档的申报可能被 FDA 执行 RTA——严重程度为 Medium。
14. Post-Market Surveillance Essentials
13. Part 11(电子记录)决策树
Mandatory Activities by Market
| Activity | FDA | EU MDR | Frequency |
|---|
| Medical Device Report (MDR/vigilance) | ✅ | ✅ | Within 30 days (FDA) / immediately for serious (EU) |
| Annual report | ✅ | | Annually |
| Periodic Safety Update Report (PSUR) | | ✅ | Annually (Class IIa+) |
| Post-Market Clinical Follow-up (PMCF) | | ✅ | Ongoing |
| Complaint handling | ✅ | ✅ | Per event |
| CAPA process | ✅ | ✅ | Per finding |
| Trend analysis | ✅ | ✅ | Quarterly minimum |
| Software update reporting | ✅ | ✅ | Per update (risk-based) |
Does your system create, modify, maintain, or transmit records required by FDA regulations?
├─ No → Part 11 not applicable (document why)
├─ Yes → Assess scope:
│ ├─ Full Part 11 (production system with regulatory records)
│ │ Requirements: Audit trails, electronic signatures, access controls,
│ │ system validation, backup/recovery, authority checks
│ ├─ Part 11-lite (hospital IT manages infrastructure)
│ │ Requirements: Your application provides audit trail + access control;
│ │ rely on hospital IT for infrastructure validation
│ └─ Development-only records (training data, model versions)
│ Requirements: Version control with audit trail, access control,
│ backup integrity verification
Electronic Signature Requirements (if applicable):
├─ Signature = legally binding (equivalent to handwritten)
├─ Must include: signer name, date/time, meaning (approval/review/authorship)
├─ Must be linked to record (cannot be separated)
└─ Biometric or non-biometric (ID + password) with controls
CAPA Trigger Decision Tree
14. 上市后监测要点
Signal detected (complaint, trend, field event)
├─ Safety-related?
│ ├─ Yes → Immediate CAPA + potential field corrective action
│ └─ No → Evaluate below
├─ Systematic (affects multiple units/users)?
│ ├─ Yes → CAPA (root cause likely systemic)
│ └─ No → Correction only (isolated incident)
├─ Recurring (≥3 occurrences same root cause)?
│ └─ Yes → CAPA mandatory (correction alone is insufficient)
└─ Regulatory requirement (audit finding, FDA warning letter)?
└─ Yes → CAPA mandatory with defined timeline
| 活动 | FDA | EU MDR | 频率 |
|---|
| 医疗器械报告(MDR/警戒) | ✅ | ✅ | FDA 30 天内;EU 严重事件立即报告 |
| 年度报告 | ✅ | | 每年 |
| 定期安全更新报告(PSUR) | | ✅ | 每年(Class IIa+) |
| 上市后临床跟踪(PMCF) | | ✅ | 持续进行 |
| 投诉处理 | ✅ | ✅ | 每次事件 |
| CAPA 流程 | ✅ | ✅ | 每项发现 |
| 趋势分析 | ✅ | ✅ | 至少每季度 |
| 软件更新报告 | ✅ | ✅ | 每次更新(基于风险) |
When NOT to Use This Skill
CAPA 触发决策树
- General software engineering without medical device claims — design controls add overhead without regulatory value
- Clinical data standards (CDISC SDTM/ADaM) — use cdisc-compliance skill instead
- Wellness/fitness apps with NO diagnostic or therapeutic claims — not regulated as devices
- eCTD submission assembly or eSTAR portal mechanics — this skill covers device compliance, not submission logistics
- Manufacturing quality (GMP, process validation for physical devices) — this skill focuses on software lifecycle
Signal detected (complaint, trend, field event)
├─ Safety-related?
│ ├─ Yes → Immediate CAPA + potential field corrective action
│ └─ No → Evaluate below
├─ Systematic (affects multiple units/users)?
│ ├─ Yes → CAPA (root cause likely systemic)
│ └─ No → Correction only (isolated incident)
├─ Recurring (≥3 occurrences same root cause)?
│ └─ Yes → CAPA mandatory (correction alone is insufficient)
└─ Regulatory requirement (audit finding, FDA warning letter)?
└─ Yes → CAPA mandatory with defined timeline
When to Escalate to Human Expert
不应使用此 Skill 的情况
- Safety classification disagreement where reasonable arguments support multiple classes — requires cross-functional risk team decision
- Predicate selection for 510(k) where no clear substantial equivalence exists — may need pre-submission meeting with FDA
- Clinical evidence sufficiency questions — requires clinical affairs and biostatistics expertise
- EU MDR classification under Rule 11 for borderline CDS/wellness software — requires Notified Body pre-assessment
- Post-market safety signal assessment for potential field action — requires medical director and regulatory affairs
- 不涉及医疗器械声明的一般软件工程——设计控制会增加额外工作,却无监管价值
- 临床数据标准(CDISC SDTM/ADaM)——改用 cdisc-compliance skill
- 无诊断或治疗声明的 wellness/fitness app——不属于受监管器械
- eCTD 申报组装或 eSTAR portal 操作——此 Skill 关注器械合规,而非申报物流
- 制造质量(GMP、实体器械的过程验证)——此 Skill 关注软件生命周期
- 安全分级存在分歧,且合理论据支持多个等级——需要跨职能风险团队作出决定
- 510(k) 的 predicate 选择不存在明确的实质等同关系——可能需要与 FDA 进行 pre-submission meeting
- 临床证据是否充分——需要临床事务和生物统计学专业知识
- 边界型 CDS/wellness software 的 EU MDR Rule 11 分类——需要 Notified Body 预评估
- 可能需要采取现场措施的上市后安全信号评估——需要 medical director 和 regulatory affairs 参与