Identifying Clinical Markers of CKD Progression Using Artificial Neural Networks and Shapley Additive Explanations

Background: Chronic Kidney Disease (CKD) is a progressive condition marked by the gradual decline of kidney function. Early stages are often asymptomatic, but progression can lead to severe complications requiring dialysis or kidney transplantation. Although numerous studies have applied machine learning to predict CKD, many suffer from a lack of interpretability due to the black-box nature of these models. Methods: We utilized a dataset comprising 400 patients, each characterized by 25 blood test parameters. Among them, 250 were diagnosed with CKD and 150 classified as non-CKD. Data preprocessing included random value/median imputation for missing values, combined with label encoding to prepare categorical features, followed by the application of a robust scaler to normalize the data. A Random Forest classifier implemented using Scikit-learn achieved a classification accuracy of 98.1%. This was compared to an artificial neural network , which yielded a accuracy of 95.4%. Given the performance of the Random Forest model, Shapley values—an xAI technique—were used to interpret the model’s predictions and identify the most influential features. Results: Shapley value analysis identified hemoglobin as the most critical feature among the 25 blood test parameters for predicting CKD. Red blood cell count and packed cell volume were ranked as the second and third most important features, respectively. Conclusion: The study demonstrates that Random Forest models outperform artificial neural networks in CKD classification. By incorporating xAI techniques such as Shapley values, the model’s decision-making process becomes more transparent. This not only enhances clinical interpretability but also reaffirms the importance of hemoglobin and other anemia-associated indicators in identifying early CKD.SHAP Feature Importance