In-Hospital Mortality Prediction for Heart Failure Patients Using Electronic Health Records and an Improved Bagging Algorithm

An improved bagging algorithm, combined with a resample strategy, a neural network, and a support vector machine (SVM), is proposed for in-hospital mortality prediction using imbalanced data with very uneven ratio of positive and negative samples. This approach was compared with other machine learning algorithms such as SVM, neural network and GBDT to evaluate its effectiveness. Permutation importance algorithm was employed to assess risk factors for heart failure patients and experimental validation was conducted using medical data from the Chinese PLA General Hospital which consisted of 207 positive and 5975 negative samples, achieving area under curve (AUC), sensitivity, and specificity values of 0.850, 0.800, and 0.752, respectively. The top 5 risk factors extracted are creatinine, serum albumin, lactate dehydrogenase, platelet count, and lymphocytes. These results suggest that the proposed method has the potential to be a valuable new tool for in-hospital mortality prediction using electronic health record data.