Machine Learning in Lupus Nephritis: Bridging Prediction Models and Clinical Decision-Making Towards Personalized Nephrology

Diego Fernando Garcia Bañol¹Adrianny Mahelis Arias-Choles¹Silvia Aldana-Peréz^1,2Gustavo J. Aroca-Martínez^1,3Carlos Guido Musso⁴Roberto Carlos Navarro Quiroz⁵Alex Dominguez-Vargas²Henry Joseth Gonzalez-Torres^1,2*

• ¹Facultad de Ciencias de la Salud, Centro de Investigaciones en Ciencias de la Vida, Simón Bolívar University, Barranquilla, Colombia
• ²Data Analysis and Mining Department, Data & Proyect Consulting Service SAS, Barranquilla, Colombia
• ³Departamento de Medicina Interna, Clinica de la Costa Ltda, Barranquilla, Colombia
• ⁴Departamento de Fisiología Renal, Hospital Italiano de Buenos Aires, Buenos Aires, Argentina
• ⁵Instituto de Química, Universidade Estadual Paulista Julio de Mesquita Filho, São Paulo, Brazil

Background: Lupus nephritis (LN) is one of the most severe manifestations of systemic lupus erythematosus (SLE), affecting up to 65% of patients and contributing significantly to morbidity and mortality. The heterogeneous clinical course of LN—characterized by alternating flares and remissions—stems from complex immunological, genetic, endocrine, and environmental factors. Current management strategies rely on immunosuppressants and corticosteroids, yet predicting disease progression, treatment response, and relapse risk remains challenging. Objective: This review synthesizes current evidence on the use of machine learning (ML) models for predicting, diagnosing, and monitoring LN, emphasizing their translational potential to improve clinical decision-making and enable personalized nephrology. Methods: A narrative synthesis was conducted of studies published between 2015 and April 2024, identified through PubMed using the terms ("lupus nephritis" OR "LN") AND ("machine learning" OR "artificial intelligence" OR "deep learning"). Eligible studies included those applying ML models to LN for diagnosis, histological classification, flare prediction, treatment response, or prognosis. Results: We identified diverse ML approaches— including logistic regression, decision trees, random forests, support vector machines, neural Página 2 de 17 networks, gradient boosting, and clustering—applied to multimodal data sources (clinical, laboratory, imaging, histopathology, and omics). These models demonstrated high performance in tasks such as non-invasive histology classification (AUC up to 0.98), flare prediction, and individualized risk stratification. Integration with big data frameworks enhanced the identification of molecular drivers, improved prognostic accuracy, and facilitated remote patient monitoring. However, model development in LN remains limited by small datasets, lack of external validation, and heterogeneous outcome definitions. Conclusions: ML models have the potential to transform LN management by enabling earlier flare detection, personalized treatment strategies, and non-invasive disease monitoring. To achieve clinical integration, future research must prioritize robust validation, interoperability with electronic health records, and transparent model interpretability. Bridging the gap between computational performance and real-world application could substantially improve outcomes and quality of life for LN patients.