Scalable Neural Network with Multistage Feature Selection for CKD Diagnosis

Introduction: This study aims to develop a scalable and accurate CKD prediction model using a hybrid Multi Stage Feature Pipeline (MSFS) strategy integrated with Artificial Neural Networks (ANN), optimized for diverse and real-time datasets. Methods: The real-time data utilized in this research comprises approximately 7596 instances having around 21 parameters. Selection of features is performed using a novel multistage pipeline that includes Boruta (BOR), Mutual Information (MI), Shapley Additive Explanations (SHAP), and Recursive Feature Elimination (RFE), which emphasizes Urea, Uric acid, Creatinine, Bun, ane, and freq as significant contributors. The methodology includes basic preprocessing, Synthetic Minority Oversampling Technique (SMOTE) with Tomek Links (TL) for handling class imbalance, hyperparameter optimization, etc. for better performance. Results: The proposed framework achieves a benchmark accuracy of 98.6% on the real-time dataset, surpassing most of the research works. Additionally, analysis of the parallel computing environment suggests that the execution time decreases proportionally as the worker nodes increase, demonstrating the framework’s computational efficiency and scalability. Discussions: The superior performance of our hybrid approach stems from the synergistic combination of multiple feature selection techniques, which collectively identify clinically relevant predictors while eliminating redundant features. The model’s ability to maintain high accuracy across diverse datasets demonstrates its generalizability, a critical requirement for real-world clinical deployment. Conclusion: The model exhibits substantial improvements in terms of performance compared to conventional techniques. Additionally, the essential features identified during the hybrid feature selection process align closely with clinically established biomarkers, underscoring the model’s relevance.