Abstract: FR-PO0002
Prognostic Prediction Model for CKD Stages by Machine Learning in Patients with ADPKD: A Nationwide Cohort Study in Japan
Session Information
- Artificial Intelligence and Digital Health at the Bedside
November 07, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Shimada, Yosuke, Intelligent Systems Laboratory, SECOM CO., LTD., Tokyo, Japan
- Kataoka, Hiroshi, Department of Nephrology, Tokyo Women's Medical University, Tokyo, Japan
- Nishio, Saori, Department of Hemodialysis and Apheresis, Hokkaido University Hospital, Hokkaido, Japan
- Hoshino, Junichi, Department of Nephrology, Tokyo Women's Medical University, Tokyo, Japan
- Hiromura, Keiju, Department of Nephrology and Rheumatology, Gunma University Graduate School of Medicine, Gunma, Japan
- Isaka, Yoshitaka, Department of Nephrology, Graduate School of Medicine, The Univeristy of Osaka, Osaka, Japan
- Muto, Satoru, Department of Urology, Juntendo University Nerima Hospital, Tokyo, Japan
Background
Autosomal dominant polycystic kidney disease (ADPKD) is a common cause of chronic kidney disease (CKD), and frequently progresses to end-stage renal disease. Accurately predicting CKD progression in ADPKD patients is essential for personalized treatment strategies. Several prognostic prediction models have been proposed, however, the utilization of machine learning (ML) in this context has not been sufficiently explored. Therefore, this study aims to develop a ML model to predict the progression of CKD stage in ADPKD patients.
Methods
This study analyzed data from 2,737 ADPKD patients enrolled in the Japanese Nationwide Cohort. Using this dataset, we developed three ML models to predict CKD stages; random forest, support vector machine, and naïve Bayes. These models were evaluated for their predictive accuracy. Feature importance analysis was performed to identify key predictive variables.
Results
Random forest exhibited the highest predictive accuracy among the three models. Feature importance analysis identified the estimated glomerular filtration rate (eGFR), serum creatinine, the CKD heat map, urinary protein, and total kidney volume as the most significant predictors of CKD stage. As a nonlinear model, random forest effectively captured the complex interactions between the variables, outperforming the linear support vector machine.
Conclusion
These findings suggest that considering complex interactions among explanatory variables may be important for predicting accurate CKD stage, as demonstrated by the superior performance of the nonlinear random forest model. These findings emphasize ML’s potential in personalized CKD management and highlight the need for individualized treatment approaches.
Comparison of accuracy and ROC-AUC among three methods
Funding
- Commercial Support – Otsuka Pharmaceutical Co., Ltd., Chugai Pharmaceutical Co., Ltd., Kyowa Hakko Kirin Co., Ltd., and JMS Co. Ltd.