| 摘要: | 冠狀動脈心臟病(Cardiac artery disease, CAD)，有時被稱為冠心病(Coronary heart disease, CHD)或是缺血性心臟病(Ischemic heart disease, IHD) 是一種知名度相當高的危險病症。對於因冠狀動脈心臟病而急診或住院的病患而言，增加其「存活率」及避免其因為冠狀動脈心臟病而「死亡」為主要醫療和照護的目標。如何事先了解影響這些結果之因素，以及提早預測這些結果事件的發生機率，並適時介入適當之改善措施，乃是臨床上疾病治療的最重要議題之一。
本研究目的有兩項：1.建立首次因冠狀動脈心臟病急診或住院病患30天內的存活預測模型。2.分析首次因冠狀動脈心臟病急診或住院病患30天內存活的重要預測因子。
本研究是以臺北醫學大學臨床研究資料庫(Taipei Medical University Clinical Research Database, TMUCRD)所收集的臺北醫學大學附設醫院、臺北市立萬芳醫院與衛生福利部雙和醫院等三家醫院臨床數據為資料來源共計33,446人，透過預測目標定義篩選出14,492人與選取50項研究特徵作為模型訓練的基礎，再選定八種機器學習演算法：LR、LDA、LGBM、GB、XGB、RF、AdaBoost、Voting進行模型訓練。結果表明Voting Classifier所呈現出來的預測能力為最佳，其測試集AUC值為0.899，準確性(=0816)較低於RF(=0.824)與AdaBoost(=0.851)在正確率上仍舊在八種演算法當中為優秀，敏感度(=0.817)與特異度(=0.816)兩方兼具優異，這點也在F1-score(=0.458)上全面評估模型性能為最佳。
在重要預測因子排名第一為前一年做過Troponin的心肌檢驗，第二則是病患年齡，第三名為肝功能檢驗的GOT，第四為腎功能檢驗的血尿素氮，第五名為心肌檢驗中的肌酸磷酸?(MB同功?)，且前五名有明顯正向貢獻度對存活的影響。
由於本研究限縮在CAD首次急診或住院的患者，因此不適用重複住院而死亡的患者，且目標為住院30天內的存活預測相較於其他文獻預測期間較短，CAD為長期心血管疾病，又加上近年就醫環境的改變，使得CAD病患在短期內住院的存活率相當高。未來若要再進一步研究，可以拉長預測期間。
Coronary heart disease (CAD), sometimes called Coronary heart disease (CHD) or ischemic heart disease (IHD), is a well-known and dangerous condition. For patients who are emergency or hospitalized due to coronary heart disease, increasing their "survival rate" and avoiding their "death" due to coronary heart disease are the main goals of medical treatment and care. How to understand the factors that affect these results in advance, predict the probability of these results in advance, and intervene in appropriate improvement measures in a timely manner is one of the most important issues in clinical disease treatment.
This study’s Target: 1. To establish a survival prediction model within 30 days of emergency or hospitalized patients due to coronary heart disease for the first time. 2. To analyze the important predictors of survival within 30 days of emergency department or hospitalization for the first time due to coronary heart disease.
This study’s dataset collected by Taipei Medical University Clinical Research Database (TMUCRD) , including Taipei Medical University Hospital, Taipei Municipal Wanfang Hospital and Ministry of Health and Welfare Shuang-Ho Hospital. All of 33,446 people to 14,492 people of target definition and 50 research features were selected as the basis for model training, and then eight machine learning algorithms were selected: LR, LDA, LGBM, GB, XGB, RF, AdaBoost, Voting for model training . The results show that the prediction ability presented by Voting Classifier is the best. Its AUC value is 0.899, and accuracy (=0816) is lower than that of RF (=0.824) and AdaBoost (=0.851). Among the algorithms, it’s sensitivity (=0.817) and specificity (=0.816), which is also the best overall evaluation model performance on F1-score (=0.458).
The important predictors, the first is Troponin, the second is age, the third is GOT, the fourth is blood urea nitrogen, and the fifth is creatine phosphatase (MB isoenzyme), and the top five have obvious positive contribution to survival.
This study is limited to patients who were first emergency or hospitalized for CAD, it is not applicable to patients who died after repeated hospitalization, and the goal is that the survival prediction period within 30 days of hospitalization is shorter than that predicted by other literatures. CAD is a long-term cardiovascular disease , coupled with the changes in the medical environment in recent years, the short-term hospital survival rate of CAD patients is quite high. For further research in the future, the forecast period can be extended. |
