從演算法到醫療器材：應用於醫療數據之醫療器材軟體全生命週期開發與法規科學實踐

Abstract

人工智慧（AI）於醫學影像、訊號與多模態資料上的學術成果，迄今已累積大量演算法創新。然而，將這些演算法轉化為可被臨床醫師實際採用、且通過主管機關審查的「醫療器材軟體」（Software as a Medical Device, SaMD），仍是一條少被系統性論述的「轉譯科學」路徑。本博士論文以此為核心命題，提出一個涵蓋完整 SaMD 生命週期的十三步驟框架，並以四項已取得上市許可之案例評估其適用性與操作一致性：從預期用途定義、風險管理（ISO 14971）、軟體生命週期管理（IEC 62304）、品質管理系統（ISO 13485）、演算法開發、分析性驗證、臨床試驗，一路延伸至法規送審與審查應答，終至學術發表、專利布局與成果推廣；並進一步將框架延伸至上市後監控與預定變更控制計畫（Predetermined Change Control Plan, PCCP）等全生命週期治理層面。 為避免流於純方法論鋪陳，本論文以四項已於本人博士學程期間成功通過 FDA 510(k) 或 TFDA 上市核准之 SaMD 產品為實證案例：（1）EFAI ChestSuite XR Pleural Effusion Assessment System（胸腔 X 光肋膜積水檢測，US FDA 510(k) K222076 & TFDA 衛部醫器製字第 007764 號）——以主動學習（Active Learning）與偽標籤（pseudo-labeling）僅需標註 16% 訓練影像即達成與完整監督相近之訓練準確度（95% vs.\ 97%），並建立跨國多讀者共識決標準作業程序（Standard Operating Procedure, SOP）；（2）DeepMets（核磁共振影像腦轉移瘤分割，US FDA 510(k) K250427）——導入 confidence-score 取聯集搭配裁決者（adjudicator）之新型共識 Ground Truth（GT）流程，並以多中心資料（含美國 16 個 sites）證實泛化能力；（3）PHDS（心電圖（Electrocardiogram, ECG）肺動脈高壓偵測，TFDA 衛部醫器製字第 008156 號）——以預先設定之經胸壁心臟超音波（Transthoracic Echocardiography, TTE）估計收縮期肺動脈壓（sPAP > 50 mmHg）作為研究參考標準，取代主觀讀者判讀，並在 N = 10,085 之大規模跨中心、時序獨立驗證集上，將敏感度提升至 83.1%——相較文獻報告之醫師 ECG 目視判讀敏感度（約 34.1%，非同一 cohort、非配對）約為 2.4 倍；（4）OsteoPredict（胸腔 X 光骨質密度異常篩檢工具，TFDA 衛部醫器製字第 008579 號）——以跨機構聯邦式學習架構在 5,943 組胸腔 X 光（Chest X-ray, CXR）／雙能量 X 光骨密度檢查（Dual-energy X-ray Absorptiometry, DXA）配對上完成訓練，並於第 173 個訓練週期 (epoch) 完成模型鎖定，外部驗證集（N = 3,007）之接收者操作特徵曲線下面積（Area Under the Receiver Operating Characteristic Curve, AUROC）達 0.9421（全世代；預先設定之 40–79 歲主力篩檢目標族群為 0.932）。另規劃涵蓋預定再訓練情境之 PCCP，惟尚未獲核准。 本論文之原創性貢獻體現於五個層面，與第 1.7 節論文貢獻聲明及第 5.5 節總結一致：框架層級——提出並以實證佐證之 13 步驟 SaMD 全生命週期轉譯模型，將醫療 AI 轉譯形式化為可追溯之系統性工程問題，而非孤立之模型優化；方法學層級——針對跨異質性資料模態之不同轉譯瓶頸匹配適當之機器學習策略（主動學習、聯邦式學習、影像與結構化臨床變項之多模態融合——以 OsteoPredict 為代表案例）；驗證策略層級——建立以法規決策為導向之驗證方法學，包含於可行處以量測為基礎錨定參考標準、於直接量測不可得時建立結構化國際多讀者共識流程、預先訂定之允收標準與次群組分析計畫、以及大規模、多中心、跨國、時序獨立之驗證設計；轉譯與法規層級——主導 FDA 重大缺陷清單之科學應答（DeepMets：以資料異質性與匹配方法學差異成功解釋戴斯相似係數（Dice Similarity Coefficient, DSC）落差）、TFDA 共病症失效邊界之主動揭露（PHDS：將混淆效應寫入 Intended Use）、並以聯邦式學習貢獻節點分析方法申請兩項中華民國發明專利，將法規攻防重構為轉譯科學之一部分；生命週期治理層級——透過所規劃之 PCCP 封閉迴路、效能漂移監控設計與聯邦式治理機制，將框架延伸至上市後之持續監督，使醫療 AI 於全產品生命週期（TPLC）接受證據導向之治理。 總結而言，本論文以四項實際取證之 SaMD 產品為實證基礎，提供一個從概念發想至臨床落地之「醫療 AI 轉譯科學實證藍圖」（An Evidence-based Blueprint for Medical AI Translation；其跨團隊、前瞻情境之外部可重複性仍待後續研究驗證），補足過去學術文獻側重演算法創新、卻對法規轉譯過程鮮少深入論述之缺口。

Artificial intelligence (AI) applied to medical imaging, signals, and multimodal data has produced extensive algorithmic innovation, yet translating these algorithms into clinically adopted, regulatorily authorized Software as a Medical Device (SaMD) is rarely articulated in a systematic way. This dissertation proposes a thirteen-step framework spanning the full SaMD lifecycle — from Intended Use definition, risk management (ISO 14971), software lifecycle management (IEC 62304) and quality management system (ISO 13485), through algorithm development, analytical validation and clinical trial, to regulatory submission and academic and patent translation — and assesses it across four authorized cases, extending it to post-market surveillance and Predetermined Change Control Plans (PCCP) as a lifecycle-governance layer. Four SaMD products authorized during the author's doctoral program ground the framework. EFAI ChestSuite XR Pleural Effusion Assessment System (chest X-ray pleural effusion detection; US FDA 510(k) K222076; TFDA 衛部醫器製字第 007764 號) used Active Learning with pseudo-labeling to reach training accuracy comparable to full supervision (95% vs.\ 97%) while annotating only 16% of the training images, and established a formal standard operating procedure for international multi-reader consensus ground truth. DeepMets (magnetic resonance imaging brain-metastasis segmentation; US FDA 510(k) K250427) introduced a confidence-score union with adjudicator consensus ground truth and generalized across 16 US sites. PHDS (electrocardiogram (ECG)-based pulmonary hypertension detection; TFDA 衛部醫器製字第 008156 號) adopted a pre-specified transthoracic echocardiography-derived estimated systolic pulmonary artery pressure (> 50 mmHg) as the study reference standard in place of subjective reader adjudication, achieving 83.1% sensitivity on a cross-site, temporally independent cohort (N = 10,085). OsteoPredict (chest X-ray bone-density screening; TFDA 衛部醫器製字第 008579 號) applied inter-institutional Federated Learning to 5,943 paired chest X-ray / dual-energy X-ray absorptiometry cases, reaching an area under the receiver operating characteristic curve of 0.9421 externally (N = 3,007). A PCCP covering predefined retraining scenarios is planned but has not yet been authorized. The original contributions span five dimensions, consistent with Sections 1.7 and 5.5: framework-level — a 13-step full-lifecycle model formalizing SaMD translation as a traceable systems-engineering problem rather than isolated model optimization; methodological — matching machine-learning strategies (Active Learning, Federated Learning, multimodal fusion) to specific translational bottlenecks; validation-strategy — a regulatory decision-oriented methodology using measurement-based reference standards where available and structured multi-reader consensus where they are not, with pre-specified acceptance criteria, subgroup plans, and cross-site, temporally independent designs; translational and regulatory — leading the scientific response to an FDA Major Deficiency Letter, disclosing comorbidity failure boundaries to TFDA, and filing two Taiwanese invention patent applications on Federated Learning contribution-node analysis; and lifecycle governance — extending the framework into post-market oversight via PCCP-driven closed loops, drift monitoring, and federated governance across the Total Product Life Cycle (TPLC). In summary, this dissertation offers an evidence-based blueprint for medical AI translation from concept to clinical deployment, grounded in four authorized SaMD products — its cross-team, prospective external reproducibility remaining to be validated — thereby addressing a literature gap that emphasizes algorithmic innovation while under-documenting regulatory translation.