以實作科學觀點探討智慧醫療 AI 系統導入之管理機制與治理成效 — 以臺灣某醫學中心為個案

Abstract

本研究以實作科學（Implementation Science）觀點，運用整合架構 CFIR 2.0（Consolidated Framework for Implementation Research）與 ERIC（Expert Recommendations for Implementing Change）策略分類，探討臺灣某醫院（以下稱「個案醫院」）導入智慧醫療 Artificial Intelligence（AI）系統之管理機制與治理成效。個案醫院為國內具規模之醫學中心，自 2005 年導入 Business Intelligence（BI） 工具奠定資料文化、2014 年完成 Internet of Things（IoT） 醫療儀器串接、2019 年成立智慧醫療中心開啟 AI 1.0 預測型 AI 階段，至 2023 年 5 月以Whisper+ChatGPT 三分鐘會議摘要為觸發點正式推進 AI 2.0 生成式 AI 階段，並於2023 年 11 月起陸續上線 HIS （Hospital Information System）Copilot以及A+ 多職類AI助理（A+ 醫師、A+ 護理師、A+ 藥師、A+ 營養師、A+ 物理治療師等）。 研究採用收斂平行（convergent parallel）混合方法設計。量化資料來源包括：（1）2022 年 1 月至 2024 年 6 月期間 1,525 筆 AI 管路照護臨床紀錄之系統日誌分析，AI 採用率介於 98.4%～99.4% 之間，AI 建議產出量為同期人工評估之 2.8～6.0 倍；（2）導入前（2021 年 7–12 月，n＝ 1,800）與導入後（2023 年 1–6 月，n ＝1,800）共 3,600 份護理交班紀錄之單組前後準實驗比較，交班耗時由 22.7 分鐘（SD＝5.3）降至 4.6 分鐘（SD＝1.2），t ＝138.4，p < .001，Cohen's d＝4.71（95% CI [4.55, 4.87]），並輔以同期未導入科室（兒科）N＝ 420 筆之非等值對照組、Shapiro–Wilk 常態檢定、Mann–Whitney U 敏感度分析及 5,000 次 bootstrap 95% CI 多重檢核；（3）由兩位護理專家獨立評分 30 份交班紀錄之品質評定，ICC(2,k) 介於 0.78 至 0.86 之間。質性資料則透過 15 位關鍵利害關係人之半結構式深度訪談，依 CFIR 2.0 五大構面進行先驗演繹編碼與浮現主題分析；於 20% 樣本之獨立試編碼後 Cohen's κ＝ 0.71，經一致化會議建立編碼決策規則後，於剩餘 80% 樣本之獨立校準編碼達 κ＝ 0.82。 研究發現五大制度化機制：（1）「護理師守門人」角色形塑；（2）跨部門治理委員會之動態調適；（3）數據驅動的持續改善迴圈；（4）臨床工作流程重構；（5）制度正當性建構。本研究進而提出「AI導入制度化模型」（AI Implementation Institutionalization Model, AIIM），將其核心關係命題化為五項可被檢驗之命題（決定因素—策略對應、守門人中介、治理動態調適、三支柱耦合、治理能力資源效率遞增），並透過與非採用（Nonadoption）、棄用（Abandonment）、擴大規模（Scale-up）、擴散（Spread）與永續性（Sustainability）框架 （NASSS）；探索（Exploration,）、準備（Preparation）、實施（Implementation）與維持（Sustainment）框架（EPIS） 對話、Kotter 八個加速因子之鏡照、與個案醫院二十年數位轉型脈絡之並讀，補足既有研究多停留於「導入決定因素診斷」、較少觸及「制度化－治理能力－永續性」轉化路徑之缺口。實務貢獻包含：為醫療機構提供可操作之 AI 治理藍圖、五項制度化機制檢核清單、ERIC 策略對應實踐手冊、給其他醫院之延伸建議，以及自 EiMBA 創業創新管理視角衍生之組織內部創業（intrapreneurship）啟示與商業模式畫布應用。

This study adopts an Implementation Science perspective, integrating the Consolidated Framework for Implementation Research 2.0 (CFIR 2.0) with the Expert Recommendations for Implementing Change (ERIC) strategies, to investigate the management mechanisms and governance outcomes of smart healthcare AI implementation at a South Taiwan hospital (the case hospital). Tracing a twenty-year digital trajectory—Business Intelligence in 2005, Internet-of-Things integration with the Hospital Information System in 2014, the founding of the Smart Healthcare Center and AI 1.0 (predictive models) in 2019, and the AI 2.0 (generative AI) era ignited by a Whisper-plus-GPT three-minute meeting summary in May 2023 and crystalised in the A+ Copilot series for physicians, nurses, pharmacists, dietitians and therapists from November 2023 onward—this study contextualises AI implementation as a long-cycle institutionalisation rather than a one-off project. A convergent parallel mixed-methods design is employed. Quantitative data include: (1) system-log analysis of 1,525 AI tube-care clinical records from January 2022 to June 2024, with AI adoption rates of 98.4%~99.4% and AI recommendation output 2.8~6.0 times that of manual assessment; (2) a single-group before-after quasi-experimental comparison of 3,600 nursing-handover records (n = 1,800 pre and n = 1,800 post), showing handover duration declining from 22.7 min (SD = 5.3) to 4.6 min (SD = 1.2), t = 138.4, p < .001, Cohen's d = 4.71 (95% CI [4.55, 4.87]), corroborated by a non-equivalent control (Pediatrics, N = 420), Shapiro–Wilk normality test, Mann–Whitney U sensitivity analysis, and 5,000-iteration bootstrap 95% CI; and (3) ICC(2,k) inter-rater reliability of 0.78–0.86 from two nurse experts independently scoring 30 handover records. Qualitative data come from semi-structured in-depth interviews with 15 key stakeholders, deductively coded via CFIR 2.0 with emergent thematic analysis (Cohen's κ = 0.71 in pilot, 0.82 in calibrated coding). Five institutionalisation mechanisms are identified: (1) Nurse Gatekeeper Role Formation; (2) Cross-departmental Governance Committee Dynamic Adaptation; (3) Data-driven Continuous Improvement Loop; (4) Clinical Workflow Redesign; (5) Institutional Legitimacy Construction. Drawing these mechanisms together, the study proposes the AI Implementation Institutionalisation Model (AIIM), articulated as five testable propositions linking CFIR 2.0 determinants, ERIC strategies and Scott's three institutional pillars, and triangulated with NASSS, EPIS and Kotter's eight accelerators. Beyond the academic contribution, the study delivers practical artifacts—an AI governance blueprint, a five-mechanism institutionalisation checklist, an ERIC strategy operationalisation handbook, and a Business Model Canvas re-interpretation—oriented to other hospitals seeking to follow the case hospital's path from technology project to organisational capability.