Analysis of Gender Disparity in Heart Failure Readmission Prediction with Missing Data

Heart failure is a critical health problem, leading to hospital readmissions that reduce quality of care and increase health expenditure. Readmission prediction models play an important role in reducing readmission rates. On the other hand, gender disparity and missing data imputation methods are overlooked in readmission prediction models. This study aims to investigate gender disparity in predicting heart failure readmissions, focusing specifically on the impact of the missing data imputation method. In our experiments, we use a comprehensive dataset consisting of a total of 160 medical and demographic variables of 2008 heart failure patients obtained from the Physionet repository. Firstly, we impute missing values using K-nearest neighbor (KNN), Soft-Impute, Multivariate Imputation by Chained Equations (MICE), and Mean imputation method. Then, we train Logistic Regression (LR), Support Vector Machine (SVM), and eXtreme gradient boosting (XGBoost) classifiers on the imputed datasets to predict readmission. We evaluate the prediction model performance through the area under the receiver operating characteristic curve (AUC). We achieve the highest AUC score in the literature 0.652 with all classifiers using the Soft-impute method. We analyze the gender disparity of prediction models with AUC, False Negative Rate (FNR) values calculated on a gender basis. According to the gender-based results, the fairest prediction model for all three classifiers are reached with the Soft-impute missing data imputation method.