人工智慧強化之下肢義肢數據註冊互動平台的開發：先導性研究

Abstract

背景：義肢照護面臨多項挑戰，包括資料收集分散、文件紀錄不一致，以及臨床決策支援有限。現有的義肢登錄系統缺乏標準化、完整的成果衡量，以及先進的分析能力。本研究旨在開發並評估一個人工智慧（AI）強化的互動平台，以解決義肢文件記錄中的這些問題。 方法：本研究採用混合方法設計，分為兩個階段。第一階段採用內容效度指數（CVI）方法，由三位義肢師專家針對六個領域的40個資料項目進行內容驗證。內容資料來自ISPO LEAD和COMPASS指引，以確保標準化。經由兩輪嚴謹的專家驗證，資料集依循系統性評估內容相關性與完整性的標準加以修訂。 第二階段開發AI推薦系統，使用隨機森林演算法。透過模擬多種截肢情境產生500筆合成病歷資料，訓練模型預測五項關鍵義肢組件：接受腔設計、足部型式、膝關節型式、襯墊型式及懸吊系統。AI模型設計用於提供每項義肢組件的循證臨床決策建議。 本平台作為三層式網頁應用程式開發，整合經驗證的資料項目及AI建議。並邀請五位合格義肢師進行可用性測試，評估平台功能、用戶體驗與決策支援能力，包括任務操作與滿意度問卷。 結果：內容效度驗證最終保留39個資料項目，平均量表內容效度指數（S-CVI/Ave）達0.991，顯示專家高度一致認可內容之相關性。平台實現臨床工作流程一條龍服務，涵蓋病患註冊與資料管理，所有重要項目皆與國際最佳實務對齊。 AI推薦系統於五項義肢組件之預測正確率皆達97.0%至100.0%。特徵重要性分析顯示，截肢層級為接受腔選擇最強預測因子（38.9%），行動能力評分為足部型式預測最強因子（52.2%），殘肢問題則為襯墊選擇最重要（83.5%）。即使在樣本數最少的情境下，模型表現亦超過98%正確率。 可用性測試顯示，SUS對應分數平均為84/100，屬於「優異」等級。所有用戶皆能完成所有操作，雖然初次註冊所需時間較長（平均477秒），但全程未發生重大錯誤。總體滿意度高（平均4.2/5），其中對預測功能的滿意度（平均4.80/5）顯著高於聊天機器人（平均4.07/5，p = 0.034）。預測功能滿意度與平台總體滿意度（rs = 0.73, p = 0.031）及專業價值認知（rs = 0.82, p = 0.013）呈現高度正相關。 討論：本研究解決義肢文件記錄重大不足，藉由標準化及AI強化的方法推動資料收集與臨床決策支援。極高的內容效度及國際指引對齊，提升跨機構互通及資料整合的潛力。 AI治療建議系統的強大表現，證明其於義肢臨床決策支援的有效性。特徵重要性分析確認模型捕捉到真實的臨床決策邏輯，支持功能性評估作為義肢處方核心依據的觀點。 可用性評估顯示平台在技術表現與預測功能方面具優勢，同時指出工作流程尚有優化空間。 結論：本研究提出一套完整的下肢義肢AI強化資料登錄平台，顯著提升義肢照護之標準化及決策支援。整合經驗證資料、智慧推薦系統及以用戶為中心的設計，為提升義肢復健成果奠定基礎。儘管結果令人振奮，仍有樣本數、合成資料、單一場域測試及缺乏長期追蹤等限制。 未來建議進行長期追蹤真實資料的驗證研究，拓展AI應用、強化行動裝置、全球多語言支援、患者端入口及利用新興科技強化資料安全。

Background: Prosthetic care faces several challenges including fragmented data collection, inconsistent documentation, and limited clinical decision support. Existing prosthetic registries lack standardization, comprehensive outcome measurement, and advanced analytical capabilities. This research aims to develop and evaluate an AI-enhanced interactive platform to address these gaps in prosthetic documentation. Methods: This research employed a mixed-methods approach, dividing the study into two phases. In Phase 1, content validation was conducted using the Content Validity Index (CVI) methodology with three expert prosthetists evaluating 40 data elements across six domains. The content was extracted from ISPO LEAD and COMPASS guidelines to ensure standardization. Through a rigorous two-round validation process, the dataset was refined following established guidelines for systematically assessing content relevance and comprehensiveness. Phase 2 involved developing an AI-powered recommendation system using Random Forest algorithms. A synthetic dataset of 500 patient records simulating diverse amputation scenarios was generated to train models predicting five key prosthetic components: socket design, foot type, knee type, liner type, and suspension system. The AI-based prediction model was designed to support clinical decision-making by providing evidence-based recommendations for each prosthetic component. The platform was developed as a three-tiered web application integrating validated data elements and AI recommendations. Usability testing with five licensed prosthetists evaluated platform functionality, user experience, and decision support capabilities through task-based assessments and satisfaction questionnaires. Results: Content validation resulted in a set of 39 data elements, with a scale-level CVI average (S-CVI/Ave) of 0.991, reflecting strong expert agreement on content relevance. The platform enabled end-to-end clinical workflow, including patient registration and data management, with all critical elements aligned to international best practices. The AI recommendation system demonstrated high predictive accuracy (97.0%–100.0%) across all five prosthetic components. Feature importance analysis showed that amputation level was the strongest predictor for socket selection (38.9%), mobility score for foot type (52.2%), and residual limb problems for liner selection (83.5%). Model performance was consistently high across all amputation levels, with even the least frequent scenarios exceeding 98% accuracy. Usability testing yielded a mean SUS-equivalent score of 84/100, categorized as "Excellent." All users successfully completed required tasks, though initial patient registration took more time (mean 477 seconds). No critical errors occurred during use. Overall satisfaction was high (mean 4.2/5), with prediction features (mean 4.80/5) receiving higher satisfaction than the chatbot (mean 4.07/5, p = 0.034). There were strong positive correlations between satisfaction with the prediction feature and both total platform satisfaction (rs = 0.73, p = 0.031) and perceived professional value (rs = 0.82, p = 0.013). Discussion: This study addresses significant limitations in prosthetic care documentation and bridging these gap by providing a standardized, AI-enhanced approach to data collection and clinical decision support data registry. The exceptionally high content validity and alignment with international frameworks enhance potential interoperability and cross-institutional data aggregation. The robust performance of the AI treatment suggestion validates its effectiveness for clinical decision support in prosthetics. Feature importance analysis confirmed that the model captured authentic clinical decision-making patterns, supporting the position that functional assessment should constitute a central determinant in prosthetic prescription. Usability evaluation identified strengths in technical performance and prediction capabilities while highlighting opportunities for workflow optimization. Conclusion: This study presents a comprehensive AI-enhanced platform for lower limb prosthetics data registry that significantly advances standardization and decision support in prosthetic care. By integrating validated data elements, intelligent recommendation systems, and user-centered design, the platform provides a foundation for improving prosthetic rehabilitation outcomes. Despite promising results, limitations include sample size constraints, reliance on synthetic data, single-site evaluation, and limited longitudinal assessment. Future directions should focus on longitudinal validation studies with real patient data, expanded AI applications, mobile-first enhancements, multilingual adaptations for global implementation, patient portal integration, and enhanced data security through emerging technologies.