第 13 課：AI 醫療設備的 FDA 和監管合規性

健康與醫療保健中的人工智慧：實戰應用

第 4 部分：生產與合規性

亞洲開發網

「建立一個好的模型只是工作的 10%。剩下的 90% 是監管、臨床驗證和部署。」— 醫療人工智慧產業的現實。

---

1. SaMD — 軟體即醫療設備

FDA 將 SaMD 定義為：

旨在用於一種或多種醫療目的的軟體，無需成為硬體醫療設備的一部分即可執行這些目的。

**用於支援診斷的 AI/ML 模型 = SaMD。 **

風險分類矩陣

薩MD級	病患風險	FDA 途徑	範例
I 類	低	510(k) 豁免或加急	基礎心率監測
II 類	平均值	510(k) 或 De Novo	AI檢測糖尿病視網膜病變
等級 III	PMA（上市前核准）	AI控制胰島素幫浦

確定類別的規則：

Healthcare situation severity × Significance of information

CRITICAL × CRITICAL = Class III (highest, PMA)
CRITICAL × SERIOUS = Class III
SERIOUS × CRITICAL = Class III
SERIOUS × SERIOUS = Class IIb (De Novo)
NON-SERIOUS × any = Class I/IIa (510k)

---

2. FDA 510(k) 途徑 — 詳細信息

510(k) = 上市前通知。需證明與謂詞設備實質等同。

# Tài liệu cần chuẩn bị cho 510(k) AI/ML:
SUBMISSION_CHECKLIST = {
    "device_description": [
        "Software architecture diagram",
        "Training data description (source, size, demographics)",
        "Algorithm description (model type, hyperparameters)",
        "Intended use statement",
        "Contraindications and limitations",
    ],
    "performance_testing": [
        "Standalone performance: sensitivity, specificity, AUC",
        "Sub-group analysis: age, sex, race, disease severity",
        "Comparison to predicate device",
        "Clinical study: reader study with radiologists",
    ],
    "cybersecurity": [
        "SBOM (Software Bill of Materials)",
        "Threat modeling",
        "Security testing results",
        "Patch management plan",
    ],
    "ai_ml_specific": [
        "Predetermined change control plan (PCCP)",  # FDA 2023
        "Algorithmic fairness analysis",
        "Training/validation/test data split methodology",
        "Distributional shift analysis",
    ]
}

2.1。讀者研究設計（對於 510k 至關重要）

from scipy import stats
import numpy as np

def reader_study_analysis(
    ai_sensitivities: list[float],
    ai_specificities: list[float],
    reader_sensitivities: list[float],
    reader_specificities: list[float]
):
    """
    FDA yêu cầu Non-Inferiority Study:
    AI performance >= Radiologist performance - non-inferiority margin
    
    Typical margin: 5% for sensitivity (không thể miss nhiều hơn 5% so với radiologist)
    
    Thiết kế: Multi-reader Multi-case (MRMC)
    - N cases (200-500+ cho rare disease)
    - M readers (5-10 radiologists)
    - AI vs readers (paired comparison)
    """
    # Non-inferiority test cho sensitivity
    ni_margin = 0.05  # 5% non-inferiority margin
    
    mean_ai_sens = np.mean(ai_sensitivities)
    mean_reader_sens = np.mean(reader_sensitivities)
    
    # Compare distributions
    t_stat, p_value = stats.ttest_ind(
        ai_sensitivities, reader_sensitivities,
        alternative='greater'  # H1: AI >= Reader - margin
    )
    
    # 95% confidence interval
    diff = mean_ai_sens - mean_reader_sens
    se = np.std(np.array(ai_sensitivities) - np.array(reader_sensitivities)) / np.sqrt(len(ai_sensitivities))
    ci_lower = diff - 1.96 * se
    ci_upper = diff + 1.96 * se
    
    non_inferior = ci_lower > -ni_margin
    
    return {
        "ai_mean_sensitivity": round(mean_ai_sens, 4),
        "reader_mean_sensitivity": round(mean_reader_sens, 4),
        "difference": round(diff, 4),
        "ci_95_lower": round(ci_lower, 4),
        "ci_95_upper": round(ci_upper, 4),
        "non_inferior": non_inferior,
        "p_value": round(p_value, 4),
        "conclusion": (
            "AI is non-inferior to radiologists" if non_inferior
            else "AI did not demonstrate non-inferiority"
        )
    }

---

3. 預定變更控制計畫 (PCCP)

FDA 2023 年基於 AI/ML 的 SaMD 指南：允許模型更新而無需完全重新提交，如果：

PCCP_REQUIREMENTS = {
    "description_of_modifications": {
        "algorithm_changes": "Retraining trên new data, same architecture",
        "performance_goals": {
            "sensitivity": ">= 85%",
            "specificity": ">= 90%",
            "auc": ">= 0.90",
        },
        "out_of_scope_changes": [
            "New intended use (cần submission mới)",
            "New patient population",
            "Architecture changes (same model family OK)",
        ]
    },
    "monitoring_protocol": {
        "real_world_performance_tracking": True,
        "drift_detection_metrics": ["PSI", "KS test", "performance drift"],
        "retrigger_training_conditions": [
            "AUC drops below 0.85 for 30 consecutive days",
            "Distribution shift PSI > 0.25",
        ],
        "hold_out_reference_dataset": "Lock 500-sample reference set for comparison"
    },
    "verification_and_validation": {
        "testing_on_new_data": True,
        "comparison_to_locked_reference_performance": True,
        "bias_analysis": ["sex", "age", "race", "disease_severity"],
    }
}

---

4. 歐盟人工智慧法案 (2024) 和越南

歐盟人工智慧法案—高風險人工智慧系統

Medical AI = HIGH RISK → nghĩa vụ nghiêm ngặt:

1. Risk Management System
2. Data Governance (quality, representativeness, documentation)
3. Technical Documentation
4. Transparency (người dùng biết đang dùng AI)
5. Human Oversight (bác sĩ vẫn có final decision)
6. Accuracy, Robustness, Cybersecurity
7. Conformity Assessment (CE marking)

越南 — 法律架構 2024

VN_REGULATORY_FRAMEWORK = {
    "primary_laws": [
        "Luật Khám chữa bệnh 2023 (Luật 15/2023/QH15)",
        "Nghị định hướng dẫn thiết bị y tế (2022)",
        "Thông tư về phần mềm y tế (Bộ Y tế)",
    ],
    "key_requirements": {
        "registration": "Đăng ký thiết bị y tế tại Cục Quản lý Dược",
        "clinical_approval": "Thử nghiệm lâm sàng tại Bệnh viện hạng I hoặc Trung ương",
        "data_protection": "Luật An ninh mạng 2018, Nghị định BVLCC",
        "ai_specific": "Chưa có luật riêng — áp dụng khung thiết bị y tế chung",
    },
    "timeline_estimate": "12-24 tháng cho sản phẩm AI y tế lần đầu",
}

---

5. 文檔框架

def generate_510k_summary_template(product_info: dict) -> str:
    """Generate FDA 510(k) Summary template."""
    return f"""
FDA 510(k) SUMMARY
==================

Device Name: {product_info.get('device_name')}
Applicant: {product_info.get('company')}
Date of Summary: {product_info.get('date')}

INDICATIONS FOR USE:
{product_info.get('indications_for_use')}

DEVICE DESCRIPTION:
Type: AI/ML-based Software as a Medical Device (SaMD)
Input: {product_info.get('input_description')}
Output: {product_info.get('output_description')}
Intended User: {product_info.get('intended_user')}

PREDICATE DEVICE(S):
{product_info.get('predicate_devices')}

SUMMARY OF PERFORMANCE TESTING:
Primary Analysis Population: {product_info.get('test_population')}
Sensitivity: {product_info.get('sensitivity')} (95% CI: {product_info.get('sens_ci')})
Specificity: {product_info.get('specificity')} (95% CI: {product_info.get('spec_ci')})
AUC: {product_info.get('auc')} (95% CI: {product_info.get('auc_ci')})

CONCLUSION:
{product_info.get('device_name')} is substantially equivalent to predicate device(s)
and is safe and effective for its intended use.
"""

---

6. 練習

1. 根據 SaMD 類別對以下 5 種醫療 AI 軟體進行分類，並確定適當的 FDA 途徑：(a) AI 檢測糖尿病視網膜病變，(b) AI 分診 COVID-19 CT，(c) AI 化療劑量建議，(d) AI 健康應用計數步驟，(e) AI 預測 ICU 中的敗血症。

2.設計AI胸部X光讀者研究：所需病例數、非劣效標準、統計分析計畫。

3. 為您在第 4 課中建立的 AI 模型起草 PCCP 大綱。

第 14 課：部署醫療 AI — 符合 HIPAA，可投入生產。