Models Integrating Disease Heterogeneity and MPA Exposure Predict Systemic and Renal Efficacy of Mycophenolate in Pediatric Lupus Nephritis.

OBJECTIVES

This study aimed to develop and validate integrated prediction models for pediatric lupus nephritis (LN) treated with a mycophenolate mofetil (MMF)-based induction regimen (combined with glucocorticoids and hydroxychloroquine), incorporating both disease heterogeneity and pharmacokinetic variability. The primary model predicts the achievement of low disease activity (Systemic Lupus Erythematosus Disease Activity Index 2000 [SLEDAI-2 K] ≤ 4) at 12 months, while a secondary model predicts renal complete response during the same period.

METHODS

A total of 120 children with LN treated with MMF between 2001 and 2025 were included. The follow-up time from MMF initiation was 12 months. Comprehensive data encompassing clinical phenotypes, organ function, immunological profiles, and pharmacokinetic parameters were collected. We employed a data-driven approach to select key features from a comprehensive set of variables reflecting disease heterogeneity and drug exposure. Seven machine learning (ML) algorithms were evaluated, with their performance assessed using area under the curve (AUC), precision, recall, F1-score, and accuracy. For the secondary outcome of complete renal response, a separate model was constructed using seven similar ML algorithms. For the primary outcome, the incremental value of adding pharmacokinetic data was quantified using the Net Reclassification Improvement (NRI). The optimal models were further validated using calibration curves and decision curve analysis, and interpreted via SHAP analysis.

RESULTS

The final integrated model for predicting low disease activity, based on a logistic regression framework, demonstrated robust performance, achieving an AUC of 0.77 in the test set, with a supporting F1-score of 0.86. The incorporation of the area under the concentration-time curve for mycophenolic acid (MPA-AUC) provided significant incremental predictive value over a clinical-only model (ΔAUC = +0.05), which was further confirmed by an NRI of 0.1309. For predicting complete renal response, a random forest model achieved an AUC of 0.88. Key variables influencing treatment response across models included MPA exposure, corticosteroid dose, immunological, renal, and hepatic parameters. Both models were well-calibrated and provided significant net benefit in decision curve analysis.

CONCLUSION

We present predictive models that incorporate a comprehensive profile of disease and drug-related heterogeneity to stratify pediatric patients with LN by their likelihood of responding to MMF. This integrated approach offers a strategy to optimize initial treatment selection by directly addressing the challenge of clinical variability.