醫療場域生成式人工智慧之操作導向治理架構

An Operation-Oriented Governance Framework for Generative AI in Healthcare

生成式人工智慧（Generative AI, GenAI）已逐步應用於醫療場域，可支援臨床、醫療文件生成及醫療知識回應等任務。然而，相較於傳統預測式人工智慧，GenAI因內容生成特性而帶來幻覺、提示設計敏感及互動式安全等新興風險。現有AI治理框架多著重於模型效能評估或高層治理原則，對於如何在醫療機構中將治理要求轉化為可持續運作之實務機制仍相對不足。本文透過文獻整合與概念分析，系統性檢視既有AI治理框架、醫療AI導入研究及生成式AI安全相關文獻，並以Donabedian結構－過程－結果（structure-process-outcome）模型作為分析基礎，發展一套具操作導向之醫療生成式AI治理架構。所提出之架構包含三個相互連結之構面：治理領域、生命週期治理及治理結果監測。治理領域涵蓋治理與問責、資料隱私與資訊安全、模型行為與透明性，以及運作保證；生命週期治理包括應用情境納管、資料準備、提示治理、部署前評估、試辦部署、正式上線後審查、變更控制及系統退役等八個決策節點；治理結果監測則聚焦於輸出品質、使用行為及流程效益，以支持系統持續安全性、可靠性及品質改善。整體而言，本架構將原則導向之AI治理轉化為可操作且可持續執行之治理流程，提供醫療機構導入與管理生成式AI之實務參考，並有助於將透明性、安全性、可問責性及持續品質管理整合至醫療GenAI之治理實踐。

Generative artificial intelligence (GenAI) is increasingly being adopted in healthcare to support tasks such as clinical summarization, medical documentation, and knowledge assistance. Compared with traditional predictive AI, however, GenAI introduces novel risks associated with content generation, including hallucinations, prompt sensitivity, and interaction-based security vulnerabilities. Existing AI governance frameworks primarily emphasize model performance or high-level ethical principles, with limited guidance on how governance can be operationalized and sustained in routine healthcare practice. This study developed an operation-oriented governance framework for healthcare GenAI by integrating institutional governance arrangements, lifecycle control mechanisms, and post-deployment monitoring. A conceptual framework development approach was adopted through literature synthesis and concept analysis, drawing upon existing AI governance frameworks, healthcare AI implementation studies, and GenAI safety literature. The Donabedian structure-process-outcome model provided the conceptual foundation for organizing governance across institutional structures, operational processes, and outcome monitoring. The proposed framework comprises three interrelated components. Governance domains define institutional responsibilities across accountability, data privacy and security, model behavior and transparency, and operational assurance. Lifecycle governance establishes eight decision gates covering use-case on boarding, data preparation, prompt governance, pre-deployment evaluation, pilot deployment, post-deployment review, change control, and system retirement. Governance outcome monitoring evaluates output quality, user behavior, and workflow efficiency to support continuous oversight and quality improvement following deployment. Collectively, these components transform principle-based governance into a practical and continuously managed framework for healthcare GenAI. The proposed framework provides healthcare organizations with an operational reference for implementing and managing GenAI applications while embedding transparency, safety, accountability, and continuous quality management throughout the governance lifecycle.