Kidney Week

Abstract: SA-PO0057

Machine Learning-Based Diagnostic Prediction of Lupus Nephritis from Routine Prebiopsy Clinical 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

• Masaki, Mamoru, Sei Marianna Ika Daigaku, Kawasaki, Kanagawa Prefecture, Japan

Mamoru Masaki, MD

• Kitano, Fumiya, Sei Marianna Ika Daigaku, Kawasaki, Kanagawa Prefecture, Japan

Fumiya Kitano, MD

• Ichikawa, Daisuke, Sei Marianna Ika Daigaku, Kawasaki, Kanagawa Prefecture, Japan

Daisuke Ichikawa, MD

• Shibagaki, Yugo, Sei Marianna Ika Daigaku, Kawasaki, Kanagawa Prefecture, Japan

Yugo Shibagaki, MD

• Noda, Ryunosuke, Sei Marianna Ika Daigaku, Kawasaki, Kanagawa Prefecture, Japan

Ryunosuke Noda, MD, DrMed

Background

Lupus nephritis (LN) requires kidney biopsy for definitive diagnosis, but biopsy is invasive. We therefore evaluated whether non-invasive prediction models using routine pre-biopsy clinical data could estimate the likelihood of biopsy-confirmed LN. We developed and internally validated machine learning models and compared them with a simple antinuclear antibody (ANA)-only baseline (ANA >=80), reflecting the ANA threshold used as an entry criterion in the current EULAR/ACR SLE classification criteria.

Methods

We retrospectively analyzed 1,420 first native kidney biopsies performed at a tertiary center from 2006 to September 2025, excluding inconclusive and overlapping diagnoses; 201 patients (14.2%) had LN. Seventeen routine variables were used: demographics, mean arterial pressure, urinalysis, hematology, biochemistry, ANA, MPO-ANCA, IgA, IgG, and C3. Seven algorithms (logistic regression, random forest, XGBoost, LightGBM, multilayer perceptron, TabPFN, and TabICL) were evaluated using nested 5-fold StratifiedGroupKFold cross-validation grouped by patient. Imputation, standardization, and hyperparameter optimization were performed within training folds only, and performance was estimated from pooled out-of-fold predictions with patient-level bootstrap confidence intervals.

Results

TabICL achieved the highest discrimination, with AUROC 0.971 (95% CI, 0.960-0.981) and AUPRC 0.869 (95% CI, 0.823-0.920). All seven models achieved AUROC >0.94. For context, an ANA >=80 baseline had AUROC 0.854 and AUPRC 0.484. At a probability threshold of 0.25, TabICL achieved sensitivity 0.856, specificity 0.966, PPV 0.808, and NPV 0.976, compared with sensitivity 0.811, specificity 0.897, PPV 0.564, and NPV 0.966 for the ANA baseline. In single-predictor analyses, ANA showed the highest discrimination, while models integrating multiple routine variables outperformed the ANA-only baseline.

Conclusion

Using routinely available pre-biopsy data, machine learning models achieved high internally validated performance for diagnostic prediction of biopsy-confirmed LN and substantially outperformed an ANA-only baseline. External multicenter validation is needed before clinical deployment.