Objective <p>To identify poor prognostic factors in Epstein-Barr virus (EBV)-positive systemic lupus erythematosus (SLE) using interpretable machine-learning (ML) models.</p> Method <p>We retrospectively analysed 217 EBV-positive SLE in-patients (2018–2022). Clinical and laboratory variables were compared between favorable and poor prognosis groups. Seven ML algorithms—logistic regression, support vector machine, naïve Bayes, random forest (RF), gradient boosting machine (GBM), artificial neural network, and AdaBoost—were trained with a 70/30 train–test split. Recursive feature elimination with cross-validation selected the optimal predictor set. SHAP (Shapley additive explanations) illustrated feature importance.</p> Results <p>Six clinical features (SLEDAI-2 K, lupus nephritis, splenomegaly, arthritis, rash, fever) and six laboratory indicators (haemoglobin, 24-h urine protein, serum uric acid, EBNA-IgG, AST, CD3 + T-cell percentage) constituted the final model inputs. The RF model performed best for clinical variables (AUC 0.71; F1 0.55); GBM performed best for laboratory variables (AUC 0.56; F1 0.30). SHAP confirmed SLEDAI-2 K, lupus nephritis, and haemoglobin as the most influential predictors.</p> Conclusion <p>Interpretable ML models highlight disease activity, renal involvement, haematological status, and EBV serology as important model-selected features associated with poor prognosis in EBV-positive SLE. These findings provide exploratory insights into potential prognostic factors.<Table Float="No" ID="Taba"> <tgroup cols="2"> <colspec colname="c1" colnum="1" /> <colspec colname="c2" colnum="2" /> <tbody> <row> <entry align="left" nameend="c2" namest="c1"> <p>Key Points</p> <p>• Seven machine learning models were systematically compared for prognostic risk prediction in SLE patients with Epstein–Barr virus infection.</p> <p>• Random Forest showed the best performance for clinical indicators, whereas Gradient Boosting Machine performed best for laboratory indicators.</p> <p>• Both clinical and laboratory variables contributed to exploratory risk stratification, and hemoglobin was identified as a model-selected laboratory feature associated with poor-prognosis predictions.</p> <p>• A prognostic risk detection model integrating clinical and laboratory indicators was established for SLE patients with Epstein–Barr virus infection.</p> </entry> </row> </tbody> </tgroup> </Table></p>

Machine learning models identify prognostic factors in systemic lupus erythematosus patients with epstein-barr virus infection

期刊论文
发表日期：

2026年7月28日

查看全文

Mengyuan Fang,
Mingyu Huang,
Feiping Ge,
Lingzhen Hu,
Xiaowei Chen,
Li Sun,
Jianxin Tu

摘要

Objective

To identify poor prognostic factors in Epstein-Barr virus (EBV)-positive systemic lupus erythematosus (SLE) using interpretable machine-learning (ML) models.

Method

We retrospectively analysed 217 EBV-positive SLE in-patients (2018–2022). Clinical and laboratory variables were compared between favorable and poor prognosis groups. Seven ML algorithms—logistic regression, support vector machine, naïve Bayes, random forest (RF), gradient boosting machine (GBM), artificial neural network, and AdaBoost—were trained with a 70/30 train–test split. Recursive feature elimination with cross-validation selected the optimal predictor set. SHAP (Shapley additive explanations) illustrated feature importance.

Results

Six clinical features (SLEDAI-2 K, lupus nephritis, splenomegaly, arthritis, rash, fever) and six laboratory indicators (haemoglobin, 24-h urine protein, serum uric acid, EBNA-IgG, AST, CD3 + T-cell percentage) constituted the final model inputs. The RF model performed best for clinical variables (AUC 0.71; F1 0.55); GBM performed best for laboratory variables (AUC 0.56; F1 0.30). SHAP confirmed SLEDAI-2 K, lupus nephritis, and haemoglobin as the most influential predictors.

Conclusion

Interpretable ML models highlight disease activity, renal involvement, haematological status, and EBV serology as important model-selected features associated with poor prognosis in EBV-positive SLE. These findings provide exploratory insights into potential prognostic factors.

Key Points

• Seven machine learning models were systematically compared for prognostic risk prediction in SLE patients with Epstein–Barr virus infection.

• Random Forest showed the best performance for clinical indicators, whereas Gradient Boosting Machine performed best for laboratory indicators.

• Both clinical and laboratory variables contributed to exploratory risk stratification, and hemoglobin was identified as a model-selected laboratory feature associated with poor-prognosis predictions.

• A prognostic risk detection model integrating clinical and laboratory indicators was established for SLE patients with Epstein–Barr virus infection.