Machine Learning Model Predicts Lupus Flare Risk Using Core Biological Indicators

In a retrospective historical cross-sectional study published in BMC Rheumatology, clinical investigators at the People’s Hospital Affiliated to Fujian University of Traditional Chinese Medicine developed and validated a machine learning model to evaluate flare risk in patients with systemic lupus erythematosus. The research team analyzed multidimensional clinical profiles to address diagnostic challenges in evaluating lupus disease activity, integrating clinical manifestations with routine laboratory indicators and enzymatic markers. By testing seven computational algorithms alongside standard regression models, the investigators demonstrated that six core indicators—joint involvement, renal involvement, fibrinogen, immunoglobulin G, complement 3, and N-acetylglucosamine-1-phosphotransferase—could accurately stratify patients according to disease activity. The optimized framework demonstrated robust predictive performance across an independent validation cohort and was subsequently translated into an interactive online dynamic nomogram to support personalized clinical risk assessment.

The investigators sought to evaluate whether modern machine learning algorithms could effectively capture linear and non-linear interactions among clinical phenotypes and routine hematological indicators. In addition, the study aimed to evaluate the specific predictive role of N-acetylglucosamine-1-phosphotransferase in active systemic disease and determine whether systemic coagulation markers, notably fibrinogen, mediate its observed association with disease flare risk.

Approved by the Ethics Committee of the People’s Hospital Affiliated to Fujian University of Traditional Chinese Medicine under approval identifier 2024KY-67, the study adhered to the Declaration of Helsinki. The authors conducted a retrospective medical chart review of patients treated in the Department of Rheumatology and Immunology between January 2019 and January 2024. All participants met the 2019 classification criteria defined by the European League Against Rheumatism and the American College of Rheumatology. Although 220 patient records were initially retrieved, 154 patients met full inclusion criteria following the exclusion of individuals with malignant tumors or previous major cerebrovascular events. The final analyzed cohort had a median age of 36 years with an interquartile range of 31.25 to 55.00 years, comprising 17 male patients, representing 11.04 percent, and 137 female patients, representing 88.96 percent.

Disease activity was measured using the Systemic Lupus Erythematosus Disease Activity Index, which evaluates manifestations over the preceding ten days with potential scores ranging from zero to 105 points. Patients were categorized into inactive and active disease groups using a standard cutoff score of 4. Under this classification scheme, scores between zero and 4 designate inactive disease, 5 to 9 denote mild activity, 10 to 14 reflect moderate activity, and 15 or higher represent severe disease flares. The activity index applies differential weightings across clinical manifestations, assigning eight points each to severe neurological and vascular events, including seizures, psychosis, organic brain syndrome, visual disturbances, cranial neuropathy, lupus headache, cerebrovascular accident, and vasculitis. In contrast, renal and musculoskeletal manifestations, including arthritis, myositis, urinary casts, hematuria, proteinuria, and leukocyturia, are assigned four points each.

The collected dataset captured an array of routine hematological, coagulation, renal, liver, thyroid, and immunological parameters. Univariate logistic regression screening eliminated 36 variables that exhibited no significant association with lupus flares at an alpha cutoff of 0.05. To select robust predictors, the researchers divided the dataset into training and validation sets using five-fold cross-validation with an 80-to-20 split. Variable screening within the training set combined three independent approaches: univariate screening with multivariate logistic regression, least absolute shrinkage and selection operator regression, and the Boruta algorithm. The shrinkage regression minimized binomial deviance at an optimal penalty parameter of log lambda equal to negative 0.01573788, retaining 18 non-zero candidate variables. The intersection of all three selection methods identified six core predictive factors: joint involvement, renal involvement, fibrinogen, immunoglobulin G, complement 3, and N-acetylglucosamine-1-phosphotransferase.

Seven computational algorithms were developed to predict disease activity: eXtreme Gradient Boosting, Support Vector Machine, Random Forest, Artificial Neural Network, Decision Tree, Logistic Regression, and Light Gradient Boosting Machine. Models were trained using ten-fold cross-validation and hyperparameter grid search. On the independent validation cohort, the Support Vector Machine delivered balanced performance among machine learning methods, achieving an area under the curve of 0.845, an accuracy of 0.818, an F1-score of 0.794, a specificity of 0.800, and a sensitivity of 0.771. Decision curve analysis confirmed clinical utility for this classifier across threshold probabilities. The Artificial Neural Network yielded an area under the curve of 0.731 and an F1-score of 0.771. Meanwhile, the Logistic Regression model exhibited strong discriminative accuracy, achieving an area under the curve of 0.869, an accuracy of 0.800, precision of 0.844, sensitivity of 0.771, specificity of 0.833, and an F1-score of 0.806.

Owing to its predictive performance and high interpretability, the logistic regression framework was selected for nomogram construction. To clarify feature contributions within the non-linear models, the researchers conducted SHapley Additive exPlanations and Local Interpretable Model-agnostic Explanations on the Support Vector Machine. Global feature importance rankings revealed that N-acetylglucosamine-1-phosphotransferase was the most influential predictor with an average absolute attribution value of 0.13, outpacing immunoglobulin G at 0.08, joint involvement at 0.06, renal involvement at 0.05, fibrinogen at 0.04, and complement 3 at 0.03. Dependence analysis indicated that joint and renal involvement produced sharp risk increases when transitioning from absence to presence. Serum fibrinogen and immunoglobulin G exhibited positive linear relationships with flare risk, complement 3 showed a negative linear relationship, and N-acetylglucosamine-1-phosphotransferase demonstrated a non-linear trajectory that rose rapidly before peaking around 40 to 50 units.

Multivariate regression confirmed the strong association between N-acetylglucosamine-1-phosphotransferase and active disease across models adjusted for demographic and clinical covariates. Evaluated continuously, each one-unit increase in the marker was associated with an unadjusted odds ratio of 1.07 with a 95 percent confidence interval of 1.03 to 1.11, indicating a 7 percent increase in flare risk. When examined across quartiles, patients in the highest quartile displayed an unadjusted odds ratio of 4.72 with a 95 percent confidence interval of 2.31 to 9.66 compared to the lowest quartile. Trend tests confirmed a statistically significant dose-response relationship across all adjusted specifications. Restricted cubic spline analysis with knots at the third quantile further characterized this pattern, demonstrating minimal risk variation at levels between zero and 20 units, followed by an upward inflection above 20 units where outcome risk increased markedly in an approximately linear manner.

Subgroup analyses showed that the predictive significance of N-acetylglucosamine-1-phosphotransferase was concentrated within high-risk patient subsets. In 112 patients positive for anti-double-stranded DNA antibodies, continuous biomarker levels significantly correlated with flare risk in fully adjusted models, yielding an odds ratio of 1.10 and a highest-quartile odds ratio of 3.85, whereas 48 antibody-negative patients showed no significant association. Among 87 patients with low complement 3 levels, the marker correlated positively with disease activity with an odds ratio of 1.09, but showed no association among 73 patients with normal complement levels. In 63 patients presenting with thrombocytopenia, the biomarker yielded an adjusted odds ratio of 1.12, whereas 97 individuals with normal platelet counts exhibited no correlation. Furthermore, among 56 patients with 24-hour urine protein levels of 3.5 grams or greater, the marker demonstrated an odds ratio of 1.11, while 104 patients below this threshold showed no association.

Causal mediation analysis was conducted using a quasi-Bayesian approach with 1,000 simulations to evaluate whether circulating fibrinogen mediated the link between N-acetylglucosamine-1-phosphotransferase and disease flares. The total effect of the enzyme on disease activity was significant with a regression coefficient of 0.01385, while the direct effect coefficient was 0.01243 and the indirect effect mediated by fibrinogen had a coefficient of 0.00142. Fibrinogen accounted for an estimated 9.3 percent of the total effect, with a 95 percent confidence interval spanning 0.0 to 28.4 percent. The authors noted that this statistical mediation provides an exploratory clue suggesting crosstalk between metabolic regulation and coagulation-inflammation pathways. However, because fibrinogen functions as an acute-phase reactant and coagulation protein, the investigators emphasized that the marker represents an indicator of disease activity rather than a proven causal driver or therapeutic target in the absence of functional experimental validation.

To facilitate bedside clinical utility, the investigators incorporated the six identified variables into a clinical nomogram and deployed an interactive dynamic web calculator hosted on the ShinyApps platform. Clinicians can enter values for joint involvement, renal impairment, fibrinogen, immunoglobulin G, complement 3, and enzyme concentration to calculate personalized flare probabilities. The authors acknowledged key limitations in their study design, including its retrospective cross-sectional framework and single-center setting, which introduce potential selection and information biases. With 154 patients enrolled, the sample size remained underpowered to evaluate uncommon manifestations such as neuropsychiatric lupus. Furthermore, the analysis relied on routine serum and clinical indicators without incorporating genomic or transcriptomic profiles. The researchers concluded that prospective multi-center studies and basic laboratory investigations in animal models are needed to validate the prediction model and clarify the biological functions of N-acetylglucosamine-1-phosphotransferase across diverse lupus populations.

Subject of Research: Machine learning risk prediction for systemic lupus erythematosus disease activity

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

Article References: Yu, X., Liu, S., Yuan, S., Lyu, X., Lin, H., Zeng, S., Zhang, W., Zhu, A., & Huang, H. (2026). A machine-learning-derived online prediction model for risk during the activity period in SLE patients: a retrospective historical cross-sectional study. BMC Rheumatology, 10(1), Article 77. https://doi.org/10.1186/s41927-026-00673-5

Image Credits: AI Generated

DOI: 10.1186/s41927-026-00673-5

Keywords: Systemic Lupus Erythematosus, Machine Learning, Disease Activity, Biomarkers, NAG1, Nomogram, Clinical Prediction, Immunology, machine-learning-derived, online, prediction, model