Development and validation of an interpretable machine learning model for predicting in-hospital mortality in patients with ventricular fibrillation

Background: Timely and accurate outcome prediction is essential for clinical decision-making in patients with ventricular fibrillation. However, the interpretation of these predictions and the translation of predictive models into clinical practice are equally crucial. This study aims to develop an interpretable machine learning (IML) model that effectively predicts in-hospital mortality for ventricular fibrillation patients. Methods: In this study, 879 patients with ventricular fibrillation from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database were randomly assigned to a training set and a test set at a ratio of 7:3. After least absolute shrinkage and selection operator (LASSO) regression analysis and the Boruta method determined the modeling variables, seven machine learning (ML) algorithms were developed to predict in-hospital mortality for patients with ventricular fibrillation using these data and externally validated in two hospitals. The area under the ROC curve (AUC) of the receiver operating characteristic (ROC) curve was calculated to assess the performance of the seven models. Decision curve analysis (DCA) was conducted to evaluate the clinical utility of the three models by estimating the net benefit at a range of threshold probabilities. Based on performance, The SHapley Additive exPlanation (SHAP) algorithm attributes interpretability to the optimal prediction model. Results: lactate, use of beta-blockers, anion gap, red blood cell count, blood urea nitrogen, hemoglobin, heart rate, respiratory rate, and albumin were selected as the nine most influential variables. The LR model demonstrated the most robust predictive performance, achieving AUROC values of 0.845 and 0.836 in the training set and test set, respectively. Furthermore, it achieved AUROC values of 0.794 and 0.903 in the two external validation sets, respectively. DCA showed that the model had the greatest net benefit rate when the prediction probability threshold is 0.15–0.45. The use of beta-blockers is the most important feature in the prediction process. The SHAP force plot provided a visualization of the direction and degree of influence of each feature on the predicting results of the model. Conclusion: ML is a reliable tool for predicting in-hospital mortality in patients with ventricular fibrillation. SHAP methods were used to explain intrinsic information of the LR model, which may prove clinically useful and help clinicians tailor precise management.