| 摘要: | 摘要: 胸痛是急診病人常見的主訴症狀，這些求診病患都可能是主要心臟不良事件(Major Adverse Cardiac Events, MACE，包括心因性死亡、心肌梗塞及事後接受血管再通術）的潛在危險對象。一般急診常使用的心電圖對於ST波段上升型急性心肌梗塞的檢查快速且有效益，然而對於主要心臟不良事件MACE預測的準確度卻相對不足。
為尋求適合台灣急診醫師使用的MACE預測系統，首先於新北市聯合醫院268位內科急診胸痛病人的先驅性研究中，利用機器學習(Machine Learning)，以人工神經網路(Artificial Neural Network, ANN)演算法篩選出預測非ST波段上升型急性心肌梗塞(NSTEMI)的九個危險因子。為更進一步改善其預測能力，持續收集急診胸痛病人至938位並利用邏輯式迴歸風險評估建立預測模型，新北市聯合醫院另一分院的116位急診胸痛病人則做為外部驗證。
邏輯式迴歸的風險評估結果，由37個變項中篩選出五個預測因子，得到ACE I (Reduced model) 與 ACE II (Full model)二個分數系統。Full model由侵入性檢查變項如血清肌酸酐(Creatinine)與肌鈣蛋白I (Troponin I)及非侵入性檢查變項包括年齡、心臟病危險因子及校正後的QT期間波長(QTc)等五個項目所組成，Reduced model 則僅包含上述三個非侵入性檢查變項。Full model 在低風險族群(low-risk group)定義為2分或以下時，具有高敏感度0.948與特異度0.546，陽性預測率0.228與陰性預測率0.987，且只有1.32% 的MACE漏失診斷率(missed MACE)。Reduced model的低風險族群定義為1分或以下時，雖然特異度較低為0.394，但敏感度仍高達0.966，漏失診斷率僅有1.22%。此結果顯示不論Full model 或Reduced model皆具有高敏感度與低漏失診斷率；Reduced model的非侵入性檢查及容易使用等特色，可提供醫院有效節省醫療資源。
上述預測MACE的五個預測因子在利用多種機器學習演算法(如artificial neural network, random forest, logistic regression, and K nearest Neighbors等)亦獲得同樣良好的預測能力。除此之外，所有變項在經由機器學習的特徵選擇(feature selection)後，所得到的前二名預測因子(QTc與年齡)與上述邏輯式迴歸的結果相同。
本研究所建構的二個危險分級分數系統，Full model 有較好的預測能力，Reduced model則提供快速且有效的預測，機器學習除佐證此二個分數系統，高效能的表現亦提供未來醫院參考之用。
Abstract: Chest pain is one of the common complaints in the emergency department (ED). Of these are potential patients with major adverse cardiac events (MACE), a composite of all-cause mortality associated with cardiovascular-related illnesses. Although ECG is a cost-effective and immediate test for the detection of STEMI, the existence of varying subgroups pose an unmet need for other approaches to accurately identify all MACE patients.
For seeking the potential predictor of MACE in Taiwan, a pilot study of 268 patients with MACE was analyzed by artificial neural network (ANN) method. Nine biomarkers were selected for identifying NSTEMI from common chest pain patients. To refine the predictor, the sample size is increased to 938 patients for model derivation and the patients (n= 162) from another hospital were analyzed as external data for external validation. By using multiple multivariate logistic regression, five biomarkers were chosen from original 37 candidate variables. Two models of stratification criteria, the full and the reduced models, were built.
Full model was based on the characteristics of variables both invasive (i.e., creatinine and troponin I) and non-invasive (i.e., age, CAD risk factors, and QTc) and reduced model was based only on non-invasive characteristics. Full model showed the high sensitivity of 0.948 and specificity of 0.546 when the cutoff was set at 2 points with missed MACE of 1.32, positive predictive value of 0.228, and a negative predictive value of 0.987. Although reduced model showed lower specificity of 0.394, a high sensitivity of 0.966 when the cutoff was set as 1 points with missed MACE of 1.22. This result support that it is possible to apply reduced model as only non-invasive characteristics, because it is easy to collect data without additional lab work which consumes more resources in the hospitals.
The high performance based on the major five biomarkers was also found in the predictor built by machine learning algorithms including artificial neural network, random forest, logistic regression, and K nearest Neighbors. Besides, the top two characteristics from feature selection are the same as the biomarkers of the reduced model by multiple multivariate logistic regression.
Two risk stratification scoring systems were developed specific to Taiwan individuals. Full model had the highest performance and reduced model can be applied as a quick and good performance diagnosis. |
