Development and validation of prediction models for death within 6 months after cardiac arrest

Abstract

BackgroundEven in patients with a successful return of spontaneous circulation (ROSC), outcomes after cardiac arrest (CA) remain poor, with some eventually succumbing after several months of treatment. There is a need for early assessment of outcomes in patients with ROSC after CA. Therefore, we developed three models for predicting death within 6 months after CA using early post-arrest factors, performed external validation, and compared their efficiency.MethodsIn this retrospective cohort study, 199 patients aged 18–80 years who experienced either in-hospital cardiac arrest or out-of-hospital cardiac arrest and achieved ROSC were included as the training set. Patients were divided into an “alive” group (95 cases) and a “dead” group (104 cases) according to their survival status 6 months after CA. Demographic data, medical history, and laboratory results were collected. Univariate and multivariate logistic regression analyses were used to identify risk factors. A risk prediction model was constructed using random forest methods, support vector machine (SVM), and a nomogram based on factors with P < 0.1 in the multivariate logistic analyses. An additional 42 patients aged 18–80 years who experienced CA with ROSC were included as the validation set. Receiver operating characteristic (ROC), decision, and calibration curves were used to assess model performance.ResultsDuration of cardiac arrest, lactate level after ROSC, secondary infections, length of hospital stay, and ventilator support were the top five risk factors for death within 6 months after CA (P < 0.1) in sequence. The random forest model [average area under the ROC curve (AUC), training set = 0.991, validation set = 0.703] performed better than the SVM model (AUC, training set = 0.905, validation set = 0.636) and the nomogram model (AUC, training set = 0.893, validation set = 0.682). Decision curve analysis indicated that the random forest model provided the best net benefit. The calibration curve indicated that the prediction for death within 6 months after CA by the random forest model was consistent with actual outcomes. The AUC of the prediction model constructed using random forest, SVM, and nomogram methods was 0.991, 0.893, and 0.905, respectively.ConclusionsThe prediction model established by early post-arrest factors performed well, which can aid in evaluating prognosis within 6 months after cardiac arrest. The predictive model constructed using random forest methods exhibited better predictive efficacy.