Transforming Raw CKD Data into a Fixed Dataset: Evaluating Preprocessing Methods

Chronic Kidney Disease (CKD) is a advanced condition that requires early detection for effective treatment and management. Predicting its severity is challenging due to inconsistencies in data, missing values, and class imbalances. In this study, real-time CKD data consisting of 12,201 patient records were collected from Ponjesley Super speciality Hospital. After preprocessing, 11,446 records were retained by removing missing values using row-wise deletion, handling redundant entries through data standardization, and applying label encoding for categorical attributes. The distribution of the dataset was analyzed using a heat map and histograms, while outliers were detected through box plots. Z-score normalization was applied to ensure uniform scaling, and outliers were addressed through removal and replacement techniques. Class imbalance was handled using Tomek Links and NearMiss methods. Three versions of the dataset, including the cleaned real-time dataset, the outlier-removed dataset, and the outlier-replaced dataset, were analyzed using ensemble learning models. Bagging, Voting, Stacking, and Boosting techniques were applied for predictive analysis. Performance evaluation was conducted based on accuracy, error rate, precision, recall, and F1-score. Results indicate that Stacking achieved the highest accuracy of 99.9% on the outlier-removed dataset, followed by Boosting (99.8%), Voting (97.9%), and Bagging (97.4%). The findings highlight the impact of preprocessing techniques on model performance and demonstrate that handling outliers improves the predictive capability of Machine Learning (ML) models for CKD severity classification.