Machine Learning Approaches for Improving the Survival Prediction Rate in Heart Failure Patients

This research work aims to highlight the significance of dimensionality reduction and minority class oversampling techniques for improving the performance metrics result of a machine learning problem involving the survival prediction rate in heart failure patients. Methods: The authors have implemented various existing feature selection and feature extraction methods to find the right subset of features for enhancing the model performance. To tackle the problem of class imbalance present in the data, three different minority oversampling methods are used. A Random Forest model is trained on each new subset of features with a 10-fold cross-validated AUC-ROC score considered as the performance evaluation metric. Results: In our experiments, feature selection methods performed better than feature extraction methods for increasing the area under the ROC curve. Recursive Feature Elimination produced the best results among all the experimented dimensionality reduction techniques. ‘Borderline-SMOTE and Borderline-SMOTE SVM’ oversampling techniques performed better than ADASYN in almost all combinations.