以台北市到院前心肺停止資料建構終止心肺復甦術規則

Abstract

研究目的 到院前心肺停止病人 (out of hospital cardiac arrest, OHCA)的存活率很低，西方國家使用終止心肺復甦術規則（termination of resuscitation, TOR），以減少不必要的運輸並重新分配有限的醫療資源。目前國際上之終止心肺復甦術規則尚未於台灣施行，且其預測台灣到院前心肺停止病人死亡的表現無法達到無效醫療定義。故此藉此回顧性研究建立適合台北市之終止心肺復甦術規則。 研究方法 本研究資料來自於2013/1/1-2017/12/31之台北市到院前心肺停止病例登記冊，目的為建立模型以預測到院前心肺停止患者其預後，主要結果為出院死亡，次要結果為出院之不良神經學預後。資料進行單變量分析，找出與主要結果及次要結果相關且具統計意義之自變項。資料進行一次拆分，隨機抽取2/3的資料作為推導組，1/3資料作為驗證組。以推導組資料進行多變量迴歸分析，建構迴歸模型，並利用ROC曲線下面積（area under the curve, AUC），由多個模型中選取最佳模型，並使用驗證組資料對最佳迴歸模型進行驗證。 研究結果 結果顯示，不論依變項為以出院死亡，或依變項為出院不良神經學預後，其最佳迴歸模型為同一個，最佳迴歸模型之自變項包含目擊倒下、到院前電擊、到院前恢復自發性心跳。對出院死亡之預測其推導組之PPV (positive predict value)為97.65 (95%CI: 97.13-98.17)，AUC為0.73 (95%CI: 0.71-0.74)；於驗證組之PPV為97.37 (95%CI: 96.59-98.14)；其AUC為0.72 (95%CI: 0.69-0.74)，在推導組或驗證組，其PPV皆<99%，對出院死亡之預測無法達無效醫療定義。對出院不良神經學預後之預測其推導組之PPV為99.24 (95%CI: 98.94-99.54)，AUC為0.75 (95%CI: 0.73-0.77)；於驗證組之PPV為99.33 (95%CI: 98.93-99.72)；其AUC為0.75 (95%CI: 0.73-0.77)，在推導組及驗證組，其PPV皆超過99%，對出院不良神經學預後之預測可達無效醫療定義。於全體8,893名患者，若使用此迴歸模型，有3012名患者被預測為不需救治，其中8名CPC1患者及3名CPC2患者，被判定不須救治。 結論 最佳迴歸模型為:目擊倒下、到院前電擊、到院前恢復自發性心跳。此模型對出院不良神經學預測可達無效醫療定義，但對出院死亡之預測無法達無效醫療定義，需進一步對此模型進行外在驗證。

Objective The survival rate of out-of-hospital cardiac arrest (OHCA) is very low. In western countries, the rule of termination of resuscitation (TOR) is used to reduce unnecessary transportation and reallocate limited medical resources. Currently, this international TOR rule is not implemented in Taiwan and fails to reach the definition of futile medicine in predicting the deaths of OHCA patients in Taiwan. A retrospective study in Taipei city was conducted to establish a suitable TOR rule. Method The data in this study are extracted from the Taipei OHCA Register from January 1, 2013 to December 31, 2017. The primary outcome is death from discharge. The secondary outcome is poor neurological outcome of discharge. Univariate analysis was conducted to one-by-one find potential factors related to the primary and secondary outcomes. Our data are randomly split into two subsets, one with 2/3 data is called derivation group and another one with 1/3 data called validation group. Multivariate logistic regression analysis for the data of derivation group was used to construct and find one best regression model with the largest area under the curve (AUC）from several candidate models. Furthermore, this regression model is validated by using the data of validation group. Result The result shows that the best regression model is the same both in death from discharge and poor neurological outcome of discharge. The variables of the best regression model include: witness collapse, shock before hospitalization, ROSC before hospitalization. For the prediction of discharge death, the PPV of the derivation group is 97.65 (95%CI: 97.13-98.17), and the AUC is 0.73 (95%CI: 0.71-0.74); the positive predict value (PPV) of the validation group is 97.37 (95%CI: 96.59-98.14) ); ad AUC is 0.72 (95%CI: 0.69-0.74). In both groups, the PPVs are less than 99% in the best model such that the prediction of death from discharge does not meet the futile medicine definition. For the prediction of poor neurological outcome of discharge, the PPV of the derivation group is 99.24 (95%CI: 98.94-99.54), and the AUC is 0.75 (95%CI: 0.73-0.77); the PPV in the validation group is 99.33 (95%CI: 98.93-99.72); its AUC is 0.75 (95%CI: 0.73-0.77). In both derivation group and validation group, predicting poor neurological outcome of discharge using the best model reaches the definition of futile medicine. Once this regression model is used, of all 8,893 patients, 3012 are predicted to terminate resuscitation including 8 CPC1 patients and 3 CPC2 patients judged not to resuscitate. Conclusion The four factors, witness collapse, shock before hospitalization, return of spontaneous resuscitation (ROSC) before hospitalization, are selected in the model. Under this model, the prediction of poor neurological outcome of discharge, reaches the definition of futile medicine, but the prediction of death from discharge cannot reach futile medicine. A further external validation for this model is needed in the future.