Hybrid Approaches in Chronic Kidney Disease Prediction: From Dimensionality Reduction to Model Explainability

Early screening of Chronic Kidney Disease represents a rising international threat since failure to undertake screening leads to manifestations including kidney failure. Researchers applied classifier models and machine learning together with deep learning algorithms to forecast the risk of CKD development within a wide demographic population. During data preprocessing all continuous variables received mean-imputation treatment while all categorical variables got mode-imputation treatment plus the data was encoded for categorical and one-hot encoded for continuous features. Principal Component Analysis (PCA) was applied to the dataset which produced four variables with an average variance of 95% in the new extracted dataset. Five different predictive models were developed through the combination of Logistic Regression and Random Forest and LightGBM and SVM and Deep Learning deep learning network. The logistic regression model and LightGBM achieved 100% accuracy but the random forest and SVM variants correctly predicted 98.75% of the classifications while making one error out of 100 instances. A single false negative-like performance could be achieved through the deep learning model implementation. The iterative process of hyper parameter tuning with GridSearchCV enabled optimization of Random Forest model by determining optimal parameters such as n_estimators = 200, max_depth = 8 along with others. The model interpretation improved with SHAP analysis where the first two principal components emerged as most important factors. Most of the research results received support through visualization tools which included confusion matrix and ROC and SHAP analysis. These evaluation metrics with statistical tests proved the robustness and reliability of the models to be introduced to clinical settings for improving CKD diagnosis and treatment methods.