新診斷心房顫動患者之中風風險預測與治療策略評估： 結合可解釋人工智慧與競爭風險分析

Abstract

對於新診斷心房顫動（atrial fibrillation, AF）患者，如何準確預測中風風險並選擇最適治療策略，仍為臨床與研究上的重要挑戰。特別是在AF診斷初期之動態臨床管理階段，治療決策具有高度不確定性，且同時存在死亡競爭風險，使風險評估更為複雜。本研究旨在建立具可解釋性之機器學習模型以進行短期中風風險預測，並結合競爭風險分析架構，評估早期抗凝血治療策略之臨床效果。 本研究使用來自多中心醫療體系之大型電子病歷資料庫，納入人口學特徵、共病狀況與用藥資訊等變數，建構邏輯斯迴歸（logistic regression, LR）與極端梯度提升（extreme gradient boosting, XGBoost）預測模型。模型效能透過內部測試及兩個獨立世代之外部驗證進行評估，並以受試者操作特徵曲線下面積（area under the receiver operating characteristic curve, AUROC）、校準指標、決策曲線分析（decision curve analysis, DCA）及淨重新分類改善指標（net reclassification improvement, NRI）等方法，全面評估模型之鑑別能力、校準度與臨床應用價值。在治療策略評估方面，針對不同早期抗凝血策略（包括直接口服抗凝血藥物、Warfarin及未使用抗凝血治療）之臨床效果比較，本研究採用時間－事件（time-to-event）分析方法，結合landmark時間點設計，並考量死亡為競爭事件，建立Fine–Gray次分布風險模型進行分析。 研究結果顯示，所建立之機器學習模型在1年中風風險預測上顯著優於傳統CHA₂DS₂-VASc評分系統，且透過特徵貢獻分析方法維持模型之可解釋性。競爭風險分析結果則顯示，在調整相關共變數後，各治療策略於中風及重大出血風險上並無顯著差異。 綜上所述，本研究建立之整合性分析框架，結合可解釋人工智慧與嚴謹之流行病學方法，展現出提升個別化風險預測與輔助臨床決策之潛力，並為心房顫動患者之精準醫療提供重要實證基礎。

Accurate stroke risk prediction and optimal treatment strategy selection remain critical challenges in patients with newly diagnosed atrial fibrillation (AF), particularly during the early post-diagnosis period characterized by dynamic clinical management and competing mortality risks. This study aims to develop and validate interpretable machine learning models for short-term stroke prediction, and to evaluate early anticoagulation strategies using a competing risk framework. Using a large-scale electronic health record dataset from a multi-center healthcare system, we constructed prediction models based on logistic regression (LR) and extreme gradient boosting (XGBoost), incorporating demographic characteristics, comorbidities, and medication use. Model performance was assessed through internal testing and external validation across independent cohorts. Discrimination, calibration, and clinical utility were evaluated using area under the receiver operating characteristic curve (AUROC), calibration metrics, decision curve analysis, and net reclassification improvement. To address time-to-event outcomes with competing risks, Fine–Gray sub-distribution hazard models were applied to compare early anticoagulation strategies, including direct oral anticoagulants, warfarin, and no anticoagulation, using a landmark design. The results demonstrated that the machine learning models significantly outperformed the conventional CHA₂DS₂-VASc score in predicting 1-year stroke risk, while maintaining interpretability through feature attribution methods. Competing risk analyses revealed no significant differences in stroke or major bleeding risk among treatment strategies after adjustment. This integrated framework highlights the value of combining explainable artificial intelligence with robust epidemiological methods to enhance individualized risk prediction and inform clinical decision-making in AF patients.