Comparative analysis of kidney function prediction: traditional statistical methods vs. deep learning techniques

Chronic kidney disease (CKD) represents a significant public health challenge, with rates consistently on the rise. Enhancing kidney function prediction could contribute to the early detection, prevention, and management of CKD in clinical practice. We aimed to investigate whether deep learning techniques, especially those suitable for processing missing values, can improve the accuracy of predicting future renal function compared to traditional statistical method, using the Japan Chronic Kidney Disease Database (J-CKD-DB), a nationwide multicenter CKD registry.

From the J-CKD-DB-Ex, a prospective longitudinal study within the J-CKD-DB, we selected individuals who had at least two eGFR measurements recorded between 12 and 20 months apart (n = 22,929 CKD patients). We used the multiple linear regression model as a conventional statistical method, and the Feed Forward Neural Network (FFNN) and Gated Recurrent Unit (GRU)-D (decay) models as deep learning techniques. We compared the prediction accuracies of each model for future eGFR based on the existing data using the root mean square error (RMSE).

The RMSE values were 7.5 for multiple regression analysis, 7.9 for FFNN model, and 7.6 mL/min/1.73 m² for GRU-D model. In the subgroup analysis according to CKD stages, lower RMSE values were observed in higher stages for all models.

Our result demonstrate the predictive accuracy of future eGFR based on the existing dataset in the J-CKD-DB-Ex. The accuracy was not improved by applying deep learning techniques compared to conventional statistical methods.