Addressing Class Imbalance Problem in Chronic Kidney Disease

Most classification problems in the healthcare domain perform poorly due to data imbalance. In this paper we did empirical study of literature related to healthcare imbalance data. To perform the experiment we have collected chronic kidney patient’s data with deidentification in collaboration with the hospital in Mumbai, India. Data was collected manually and preprocessing highlighted imbalance in collected data. To understand class distribution and improve data balance we studied and implemented various algorithms. First implemented method is oversampling to increase minority observations. To generate new observations, Synthetic Minority Over-sampling Technique (SMOTE) is performed. To avoid outlier effect Adaptive Synthetic (ADASYN) sampling is performed. Majority class was undersampled with undersampling techniques such as random undersampling and variations of near miss methods. At last all results are compared to generate balance data and results are evaluated with Tomeks links metric. This dataset is further used to build an early detection model in Chronic Kidney Disease (CKD). Early detection model is built using clinical features, Gender, Age, Smoking, Alcohol, Birthplace, Start of Dialysis, Frequency of Dialysis, Diabetes, BP, Lipid Profile, Kidney stone, Creatine, Urea, Uric Acid. As the data was collected manually many features had missing values and data depicted imbalanced nature. Before building a classifier model we preprocessed data to avoid biased results in future. Early detection model is then built using Machine Learning algorithms. We compared predictions by Support Vector Machine, Random Forest, Regression model. With our data Random Forest model outperformed with over 93% success rate in classifying the patients with kidney diseases based on performance metrics.