Abstract: SA-PO0071
Multiclass Machine Learning-Based Diagnostic Prediction of Six Major Kidney Diseases Using Routine Prebiopsy Data
Session Information
- Artificial Intelligence: Kidney Imaging, Pathology, and Precision Phenotyping
October 24, 2026 | Location: Exhibit Hall A, 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
- Noda, Ryunosuke, St. Marianna University School of Medicine, Kawasaki, Japan
- Masaki, Mamoru, St. Marianna University School of Medicine, Kawasaki, Japan
- Kitano, Fumiya, St. Marianna University School of Medicine, Kawasaki, Japan
- Ichikawa, Daisuke, St. Marianna University School of Medicine, Kawasaki, Japan
- Shibagaki, Yugo, St. Marianna University School of Medicine, Kawasaki, Japan
Background
Kidney biopsy is required for definitive diagnosis of kidney disease, but it is invasive, motivating interest in non-invasive diagnostic prediction models. Because kidney diseases are diverse, such models should distinguish among multiple diseases, whereas most studies focus on a single disease versus all others. We therefore developed and internally validated multiclass machine learning models for diagnostic prediction of six major biopsy-confirmed kidney diseases from routine pre-biopsy clinical variables.
Methods
We retrospectively analyzed first native kidney biopsies at a tertiary center from 2006 to 2025. The cohort was restricted to six major diagnoses: IgA nephropathy (IgAN), lupus nephritis (LN), ANCA-associated vasculitis (AAV), minimal change disease (MCD), membranous nephropathy (MN), and diabetic nephropathy (DN). The final cohort included 985 patients: IgAN 400, LN 199, AAV 114, MCD 104, MN 98, and DN 70. Seventeen routine variables were used, spanning demographics, blood pressure, urinalysis, hematology, biochemistry, and immunologic markers. Seven algorithms were compared using nested 5-fold StratifiedGroupKFold cross-validation grouped by patient. Imputation, scaling, and hyperparameter optimization were fitted within training folds only. The primary metric was macro one-vs-rest AUPRC; secondary metrics included macro AUROC, accuracy, balanced accuracy, and top-2 accuracy.
Results
TabPFN achieved the highest macro AUPRC, 0.838 (95% CI, 0.810-0.871), with macro AUROC 0.957 (95% CI, 0.947-0.968), accuracy 0.843, balanced accuracy 0.780, and top-2 accuracy 0.940. TabICL showed similar performance (macro AUPRC 0.834, macro AUROC 0.958). For TabPFN, classwise AUROC/AUPRC were 0.973/0.956 for IgAN, 0.975/0.919 for LN, 0.987/0.939 for AAV, 0.958/0.854 for MCD, 0.892/0.591 for MN, and 0.960/0.768 for DN. MN remained the most challenging class, with lower AUPRC and F1 than other diagnoses.
Conclusion
Multiclass machine learning using routine pre-biopsy data showed strong performance for predicting major biopsy-confirmed kidney diseases. External validation is needed to confirm generalizability.
Acknowledgment
This work was supported by JSPS KAKENHI Grant Number JP25K21327.
Funding
- Government Support – Non-U.S.