Interpretable Machine Learning Models for Mortality Prediction in Critically Ill Patients with Intra-Aortic Balloon Pump Therapy: Development and Evaluation.
TL;DR
The RF model, based on interpretable machine learning, has proven to be effective in predicting ICU mortality among critically ill patients undergoing IABP therapy, aiding in risk stratification and supporting clinicians in making more informed decisions.
Abstract
Background
This study aims to develop and validate predictive models to assess mortality risk among patients undergoing intra-aortic balloon pump (IABP) therapy in the intensive care unit (ICU).
Method
A retrospective analysis was performed on 764 critically ill patients who received IABP therapy, using data from the MIMIC-IV (Medical Information Mart for Intensive Care IV) database. The data set was chronologically split into training and temporal external validation cohorts. Variable selection was performed using LASSO (least absolute shrinkage and selection operator) regression, and six machine learning models were constructed. Model evaluation was conducted using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis. The optimal model was interpreted using SHAP (SHapley Additive exPlanations), and an online web-based calculator was developed.
Results
LASSO regression identified nine mortality-associated risk factors from 61 variables: lactate, norepinephrine, age, blood urea nitrogen, oliguria, coronary artery bypass grafting, red cell distribution width, renal replacement therapy, and cardiac arrest. Six machine learning models were developed: K-nearest neighbour, support vector machine, random forest (RF), extreme gradient boosting, decision tree, and LASSO logistic regression. The RF model demonstrated superior performance, achieving an area under the ROC curve of 0.965 (95% confidence interval [CI] 0.950-0.980) in the training cohort and 0.900 (95% CI 0.858-0.942) in the temporal external validation cohort. This model achieved the lowest Brier score of 0.104. Decision curve analysis confirmed the clinical utility of the RF model in predicting ICU mortality among patients undergoing IABP.
Conclusions
The RF model, based on interpretable machine learning, has proven to be effective in predicting ICU mortality among critically ill patients undergoing IABP therapy. This advances clinical practice by aiding in risk stratification and supporting clinicians in making more informed decisions.