A machine learning-based predictive tool for complete renal response in lupus nephritis leverages the 3-month treatment milestone

Lupus nephritis (LN) remains a major determinant of poor prognosis in systemic lupus erythematosus, yet reliable tools for early identification of unfavorable renal outcomes are lacking. This study aimed to identify early prognostic factors for complete renal response (CRR) at 12 months in active LN and to develop interpretable machine learning (ML) models for early prediction of treatment response. In this prospective cohort of 400 active LN patients followed for 12 months, the primary outcome was 12-month CRR. Candidate predictors were assessed at baseline and 3-month follow-up. Four ML models (Random Forest, Support Vector Machine, Elastic Net, and eXtreme Gradient Boosting) were developed and validated on a test set. Model performance was evaluated using area under the ROC curve (AUC), calibration and decision curve analyses, with feature importance assessed via SHapley Additive exPlanations (SHAP). 296 patients (74%) achieved CRR at 12 months. Baseline factors independently associated with CRR included 24-hour urinary total protein (24hUTP), higher serum albumin, shorter disease duration, and treatment-naïve status. Significant group differences emerged at 3 months across multiple parameters (all p < 0.05). Baseline‑only models showed limited discriminative ability (AUC ≤ 0.75), whereas incorporating 3-month data markedly improved performance; the Random Forest model achieved the best testing-set AUC (0.862, 95% CI: 0.775–0.949). SHAP analysis identified 3-month 24hUTP reduction as the most influential predictor of CRR. We developed an interpretable ML tool that accurately predicts 12-month CRR as early as 3 months post-treatment, highlighting the critical role of early proteinuria reduction as a key indicator for long-term renal outcomes.