Related Experiment Video
Updated: Jun 2, 2025
10:37
Comparative Proteomic Analysis of Whole Kidney, Medulla, and Cortical Tubules in Diabetic Pathogenesis of Kidney Injury in Mice
Published on: May 2, 2025
94
Comparative analysis of kidney function prediction: traditional statistical methods vs. deep learning techniques
Mizuki Ohashi1, Yuya Ishikawa2, Satoshi Arai3
1Shiga University of Medical Science, NCD Epidemiology Research Center, Shiga, Japan.
Clinical and Experimental Nephrology
|January 15, 2025
Summary
Deep learning models did not improve kidney function prediction accuracy compared to traditional methods in chronic kidney disease (CKD) patients. This study highlights the predictive capabilities of existing statistical approaches for future estimated glomerular filtration rate (eGFR).
Area of Science:
- Nephrology
- Artificial Intelligence in Medicine
- Biostatistics
Background:
- Chronic kidney disease (CKD) is a growing public health concern, necessitating improved methods for predicting renal function.
- Accurate prediction of future kidney function is crucial for early detection, prevention, and management of CKD.
- The Japan Chronic Kidney Disease Database (J-CKD-DB) provides a valuable resource for studying CKD progression.
Purpose of the Study:
- To evaluate the effectiveness of deep learning techniques in predicting future estimated glomerular filtration rate (eGFR) in CKD patients.
- To compare the predictive accuracy of deep learning models against traditional statistical methods using real-world CKD data.
- To assess the utility of models handling missing data for improved eGFR prediction.
Main Methods:
- Utilized data from the J-CKD-DB-Ex prospective longitudinal study, including 22,929 CKD patients with at least two eGFR measurements.
- Employed multiple linear regression as the conventional statistical method.
- Applied deep learning models: Feed Forward Neural Network (FFNN) and Gated Recurrent Unit (GRU)-D, to predict future eGFR.
- Compared model performance using Root Mean Square Error (RMSE) to quantify prediction accuracy.
Main Results:
- The multiple linear regression model achieved an RMSE of 7.5 mL/min/1.73 m².
- The FFNN model resulted in an RMSE of 7.9 mL/min/1.73 m².
- The GRU-D model yielded an RMSE of 7.6 mL/min/1.73 m².
- All models showed improved performance in higher stages of CKD, with lower RMSE values.
Conclusions:
- The study demonstrates the predictive accuracy of existing datasets for future eGFR in CKD patients.
- Deep learning techniques, including FFNN and GRU-D, did not significantly improve eGFR prediction accuracy compared to conventional multiple linear regression.
- Traditional statistical methods remain effective for predicting future renal function in the context of the J-CKD-DB-Ex dataset.
