A machine-learning-derived online prediction model for risk during the activity period in SLE patients: a retrospective historical cross-sectional study.

objectiveGiven the challenges in predicting systemic lupus erythematosus (SLE) disease activity and the limitations of existing assessment tools, this study aimed to integrate multidimensional clinical and laboratory indicators to develop and validate a machine learning-based predictive model for SLE disease activity. It further sought to explore the association and potential mechanisms linking N-acetylglucosamine-1-phosphotransferase (NAG1) to SLE disease activity, thereby offering clinical risk assessment tools and identifying novel therapeutic targets.

methodsClinical data from 201 SLE patients were retrospectively collected. Patients were stratified according to Systemic Lupus Erythematosus Disease Activity Index (SLEDAI) scores. Core predictive variables were selected using multivariate logistic regression, LASSO regression, and the Boruta algorithm. Seven machine learning algorithms were employed to construct predictive models, with SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) used for model interpretation. The role of NAG1 and the mediating effect of fibrinogen (FIB) were analyzed using multivariate logistic regression, restricted cubic splines, and a quasi-Bayesian approach. Finally, a nomogram and an interactive dynamic risk calculator were developed based on the optimal model.

resultsJoint involvement, renal involvement, fibrinogen (FIB), immunoglobulin G (IgG), complement 3 (C3), and NAG1 were identified as core predictive indicators for SLE disease activity. The support vector machine (SVM) model demonstrated balanced predictive performance, while the logistic regression model was selected for nomogram construction due to its high predictive accuracy and interpretability. NAG1 was the most critical predictive factor; elevated NAG1 levels were significantly associated with an increased risk of SLE disease activity, with a more pronounced correlation in high-risk subgroups, such as patients positive for anti-double-stranded DNA (anti-dsDNA) or with low C3 levels. FIB mediated 9.3% of the pro-flare effect of NAG1. The constructed model and supporting tools showed favorable clinical applicability. Study limitations include its retrospective design and limited sample size.

conclusionA predictive model for SLE disease activity, based on six core indicators, was successfully developed and validated. The logistic regression model demonstrated excellent performance. The nomogram and dynamic risk calculator enable non-invasive, personalized risk assessment for SLE disease activity. As a key predictive factor, NAG1 may act as a pathogenic mediator in SLE, and its mediating relationship with FIB suggests novel crosstalk between metabolic dysregulation and the coagulation-inflammation pathway. This model can facilitate early risk stratification and individualized management of SLE patients. NAG1 represents a promising biomarker for evaluating therapeutic response and predicting relapse in SLE, and may serve as a potential therapeutic target. Future multicenter prospective studies are warranted to validate the model and further investigate the biological functions of NAG1.