Value of Nomogram in Early Prediction of Death in Patients Post Extracorporeal Membrane Oxygenation
Liming Wang,Lin Chen,Huabin Wang,Kun Chen
DOI: https://doi.org/10.3760/cma.j.cn121430-20220211-00122
2022-01-01
Abstract:OBJECTIVE:To construct an early predictive model for the death of patients after extracorporeal membrane oxygenation (ECMO) based on the baseline information of patients and laboratory indicators.METHODS:The clinical data of 139 patients who underwent ECMO in Affiliated Jinhua Hospital, Zhejiang University School of Medicine from January 2015 to December 2021 were collected, including age, gender, primary disease, ECMO model, other clinical characteristics, and laboratory indicators 2 hours after establishment of ECMO. The patients were divided into training cohort (n = 111) and validation cohort (n = 28) according to a ratio of 4:1. The least absolute shrinkage and selection operator (Lasso) regression and multivariate Logistic regression were used to select predictive factors, and a nomogram was used to establish the predictive model. The calibration and discrimination of the model were assessed using the receiver operator characteristic curve (ROC curve), the calibration curve and Kaplan-Meier analysis.RESULTS:Four predictive variables including anion gap (AG), lactic acid (Lac), arterial partial pressure of oxygen (PaO2), and serum amylase (AMY) 2 hours after ECMO were selected from 34 laboratory indicators by Lasso regression. These variables and three other clinically important factors [primary diseases, ECMO model, and acute kidney injury (AKI)] were analyzed using multivariate Logistic regression (forward: LR method), Finally, four strong predictive factors (Lac-2 h, PaO2-2 h, AMY-2 h, and primary diseases) for mortality were used to construct a nomogram [the area under ROC curve (AUC) was 0.85,95% confidence interval (95%CI) was 0.78-0.92]. The optimal cut-off value of the ROC curve was 0.398 based on the maximum principle of the Youden index, corresponding to 80.2% accuracy, 89.1% sensitivity, and 68.1% specificity. The validation cohort with an in-hospital mortality of 64.3% was used to verify the performance of the model by ROC curve and Kaplan-Meier analysis. According to the optimal cut-off value of the ROC curve (0.398), the validation cohort was divided into low- and high-risk groups. Based on this model, the survival probability of the low-risk group was significantly higher than that of the high-risk group (P = 0.018), indicating this model had good discriminative ability in the validation cohort. Based on this model, the AUC of the validation cohort was 0.76, 95%CI was 0.58-0.94, and the accuracy rate was 71.43%, which indicated this model showed good calibration consistency.CONCLUSIONS:The predictive model incorporating Lac-2 h, PaO2-2 h, AMY-2 h, and primary diseases may be significant for predicting the in-hospital mortality of patients undergoing ECMO.