Assessment of Machine Learning Approaches to Predict in-Hospital Mortality in Patients Underwent Prosthetic Heart Valve Replacement Surgery

Abstract

Background & Objective: Machine learning and artificial intelligence are useful tools to analyze data with multiple variables. It has been shown that the prediction models obtained by Machine learning have better performance than the conventional statistical methods. This study was aimed to assess the risk factors and determine the best machine learning prediction model/s for in-hospital mortality among patients who underwent prosthetic valve replacement surgery. Materials & Methods: In this retrospective cross-sectional study, patient’s pre-operative, intra-operative and post-operative data underwent univariate analysis. Feature importance determination was carried out using algorithms including principal component analysis (PCA), support vector machine (SVM), random forest (RF) model-based, and recursive feature elimination (RFE). Then, 13 machine learning classifiers were implemented for in-hospital prediction model. Results: The In-hospital mortality rate was 6.36%. Data from 2455 patients underwent final analysis. The machine learning results revealed that among pre-operative features, Adaptive boost (AB) and RF classifiers (AUC: 0.82±0.033; 0.78±0.028, respectively); among intra-operative features, AB and K-nearest neighbors (KNN) classifiers (AUC: 0.68±0.014); among postoperative features, AB and RF classifiers (AUC: 0.9±0.1; 0.88±0.095, respectively); and among all features, AB and LR classifiers (AUC: 0.93±0.049; 0.93±0.055, respectively) had the best performance in prediction of in-hospital mortality. Conclusion: The AB classifier was determined as the best model in prediction of in-hospital mortality in all 4 datasets.