Predictive model of cognitive dysfunction among the patients with systemic lupus erythematosus

Cognitive dysfunction (CD) is a frequent but often underrecognized clinical feature in systemic lupus erythematosus (SLE) patients. It markedly impairs their health-related quality of life. Investigating the risk associated with the onset of CD and establishing a prediction model are crucial for early detection of CD. This allows for timely intervention, potentially delaying or reversing the progression of CD.

This study aimed to establish a predictive model for CD in SLE patients and evaluate its predictive efficacy.

This study enrolled adult patients with SLE who underwent inpatient management at the Nanjing Drum Tower Hospital Clinical College of Nanjing Medical University. Variables for analysis included demographic characteristics, physiological and psychological status, medication details, and laboratory examination. Key predictors were selected using the least absolute shrinkage and selection operator (LASSO) approach. A range of machine learning (ML) classification models were developed and evaluated to determine the best-performing model. The prediction accuracy of the best-performing model was evaluated using calibration curve analysis, and the clinical applicability of the model was further evaluated by decision curve analysis. Additionally, Shapley Additive exPlanations (SHAP) were applied to realize personalized risk assessment and enhance model interpretability.

The study included 283 eligible patients, divided into a training set of 199 and a test set of 84. The eXtreme Gradient Boosting (XGBoost) model emerged as the optimal model. In the training set, the area under the curve (AUC) (95% confidence interval, CI) was 1.000 (0.953, 1.000), and in the test set, it was 0.774 (0.763, 0.831). The robustness of the XGBoost model was further substantiated by repeated cross-validation on the training set, with a mean AUC of 0.750. Regarding the F1 value in the test set, Logistic Regression had the highest value (F1 = 71.58%), followed by XGBoost (F1 = 65.85%) and Decision Tree and Random Forest (F1 = 65.82%, 60%, respectively). The congruence of the XGBoost model's predictions with actual findings was corroborated by the calibration curve. Furthermore, decision curve analysis affirmed the clinical value of the model in predicting CD.

The prediction model of the present study provides clinicians with a tool to identify SLE patients at high risk of cognitive dysfunction. Its application could support the early initiation of preventive interventions in this vulnerable population. Key Points • The study integrated LASSO regression for feature selection and multiple ML algorithms, with XGBoost emerging as the optimal predictive model for CD in SLE. • SHAP interpretation was applied to provide individualized risk predictions, enhancing clinical utility for patient-specific management. • The developed ML-based tool aids clinicians in early identification of high-risk SLE patients, enabling timely interventions to potentially mitigate or reverse CD progression.