Validity and applicability of machine learning models for systemic lupus erythematosus diagnosis

The diagnosis of systemic lupus erythematosus (SLE) is clinically complex, and early identification is essential for timely intervention and reducing disease burden. Machine learning offers a promising approach to distinguish early-stage SLE patients from healthy individuals.

A total of 2672 SLE patients and 154 798 healthy controls from the Luzhou (discovery) cohort, along with 2532 SLE patients and 38 597 healthy controls from the Enshi (validation) cohort, were enrolled in this study. A complete machine learning pipeline-including data preprocessing, feature selection, model training and postanalysis, was developed in the Luzhou cohort and subsequently validated in the Enshi cohort. Optimal features and the best-performing model were identified in the Luzhou cohort, then scaled and validated in the Enshi cohort. Model performance was evaluated using 13 binary classification metrics. The optimal feature set and model were integrated to construct an Artificial Intelligence Prediction tool for SLE (AI-PSLE).

Fifty candidate features were initially selected in the Luzhou cohort, among which the light gradient boosting (LGB) model demonstrated the best performance following data preprocessing. After scaling in the Enshi cohort, 35 reproducible features were retained. The LGB model based on these 35 features maintained superior performance in the Luzhou cohort and was further successfully validated in both the Enshi and combined Luzhou+Enshi cohorts.

We developed an open-access, clinically user-friendly tool-AI-PSLE-based on 35 routine features, aimed at facilitating the early identification of SLE patients from healthy populations.