| 摘要: | 一、研究目的與背景：
心房纖維顫動（atrial fibrillation, AF）是臨床病人發生缺血性腦中風的重要危險因子。許多臨床研究已證實口服抗凝血劑可有效地預防AF病人中風發生及改善死亡率。從2011年起，非維他命K拮抗劑口服抗凝血劑（Non-vitamin K antagonist Oral Anticoagulants, NOACs ）陸續核准用於非瓣膜性AF的預防。相較於warfarin，由於NOACs有較高的安全性，此類藥物逐漸成為AF病人抗凝治療的首選藥物。一般而言，NOACs造成的嚴重出血風險約為1.6~3.6%，但Chao等人的研究中發現，NOACs在亞洲次族群的出血風險可能較高，顯示評估亞洲族群使用NOACs的安全性為臨床重要議題。本研究將針對NOACs引發重大出血事件之病人進行分析，透過機器學習演算方法，找出出血病人的可能危險因子，並建立適合亞洲族群之出血風險預測模型。
二、研究材料與方法：
本研究分析2010年7月1日至2020年12月31日間於長庚醫療體系就醫的電子病歷資料。研究將先找出於2011年1月1日至2019年12月31日間初次處方NOACs（包含direct thrombin inhibitor dabigatran and direct factor Xa inhibitors apixaban, edoxaban, rivaroxaban等四種藥品）之非瓣膜性AF病人（ICD-9 427.3, ICD-10 I48），並由首次用藥日開始追蹤至用藥後一年、死亡、發生重大出血或失去追蹤。重大出血之定義乃依據國際血栓暨凝血學會（International Society on Thrombosis and Haemostasis ,ISTH）非手術病人重大出血事件所指之致死性出血；體內重要器官或特殊部位出血；以及出血所致hemoglobin下降超過2g/dL或需要輸注2個單位全血或紅血球細胞者等三項。
三、分析與結果：
納入研究對象為14,574人，男性6,138人（42.12%）、女性8,436人（57.88%），平均觀察期為8.4個月；在研究期間內發生致死性出血者，共有109件；體內重要器官或特殊部位出血者，共有527件。研究發現胰島素、抗黴菌劑及貧血是重大出血事件的重要預測因子，以機器學習方法比較現況HAS-BLED score的c-index（RF 0.738, 95% CI 0.703-0.773；XGBoost0.731, 95%CI 0.697-0.766；LR 0.725, 95%CI 0.692-0.759；SVM 0.664, 95%CI 0.623-0.705）及modified HAS-BLED score的c-index（RF 0.81, 95%CI 0.780-0.841；XGBoost0.800, 95%CI 0.767-0.833；LR 0.795, 95%CI 0.763-0.826；SVM 0.737, 95%CI 0.702-0.772）新模型風險預測結果，證實modified HAS-BLED score出血風險預測能力確實優於現有評分工具。
四、討論：
運用真實世界之大數據進行機器學習，確實能找出亞洲族群具重大臨床意義之出血風險因子，與傳統評估工具HES-BLEDscore相比，以胰島素、抗黴菌劑及貧血調整後之新模型確實提供了更準確的風險預測能力，且增加臨床執行可近性。
Background：
Atrial fibrillation（AF）is one of the most important risk factors for embolic stroke. Numerous clinical studies have proved that oral anticoagulants could effectively prevent stroke and improve mortality in patients with AF. Since 2011, non-vitamin k antagonist oral anticoagulants （NOACs）have been approved in the preventions of stroke in patients with non-valvular Atrial fibrillation（NVAF）. Compared to warfarin, NOACs have more acceptable safety profiles, gradually serving as the first-line choice for stroke prevention in NVAF. Overall, the risk of severe bleeding events caused by NOACs is estimated about 1.6~3.6%, but the report from Chao et al. found that NOACs may have higher bleeding risks in Asian population which indicates the critical importance of the safety evaluation regarding the uses of NOACs in Asia countries. This study will include AF patients with major bleeding events caused by NOACs in Taiwan to identify the possible risk factors for bleeding and to establish a risk prediction model for bleeding events by using machine learning algorithms.
Material and Methods：
We conducted a series of retrospective cohort studies by using CGRD. The patients with NVAF（ICD-9 427.3, ICD-10 I48）who were included who had their first prescription of an NOAC including dabigatran, apixaban, edoxaban, or rivaroxaban treatment from January 2011 to December 2019and start tracking from the day of first prescription after one year, death, major bleeding or loss of follow-up. Major bleeding was defined by the International Society on Thrombosis and Hemostasis（ISTH）criteria that including：1.fatal bleeding, 2. symptomatic bleeding in a critical area or organ, and 3.Bleeding causing a fall in hemoglobin level of 2 g/dL （1.24 mmol/L）or more, or leading to transfusion of two or more units of whole blood or red cells. We will analyze patients with major bleeding events caused by NOACs, using machine learning algorithms to identify possible risk factors, and establish a prediction model suitable for Asian ethnic groups.
Results：
We identified a total of 14,574 NVAF patients newly initiating NOACs in this study, most of whom (57.9%) were female. During the mean follow-up of 8.4 months, 2,112patients developed major bleeding events, including 109 for fatal bleeding, 527 for symptomatic bleeding in a critical area or organ. We found the insulin, fungostatics and anemia were the important predictors for major bleeding. Compared toHAS-BLED score（RF 0.738, 95% CI 0.703-0.773；XGBoost0.731, 95%CI 0.697-0.766；LR 0.725, 95%CI 0.692-0.759；SVM 0.664, 95%CI 0.623-0.705）, the modified HAS-BLED score（RF 0.81, 95%CI 0.780-0.841；XGBoost0.800, 95%CI 0.767-0.833；LR 0.795, 95%CI 0.763-0.826；SVM 0.737, 95%CI 0.702-0.772）with the incorporation of three new predictors yielded a higher c-indexfor major bleeding.
Conclusions：
In this real-world cohort from Taiwan, the modified HAS-BLED score by incorporating insulin, fungostatics and anemia was associated with better major bleeding risk predictions, compared with the HES-BLED score. Further external validation of the modified HAS-BLED score should be performed. |
