Development of a LASSO-based machine learning model for relapse prediction in lupus nephritis
Hao Liu, Yan Zheng, Jin Ding, Ying Li, Jinghua Wang, Wenjuan Zhang, Zhaohui ZhengObjective
To identify the predictors of renal relapse in patients with lupus nephritis (LN) and develop a predictive model.
Methods
The patients with LN who received renal biopsies from 2005 to 2024 in Xijing Hospital were enrolled. This study was designed as a cohort study. Clinical and serological characteristics at the time of renal biopsy, nephritis histology, the duration to remission and treatment regimen were evaluated. Variables were screened using least absolute shrinkage and selection operator (LASSO)-Cox analyses, followed by the application of multivariate Cox regression to construct a relapse risk prediction model. The performance of the model was evaluated using bootstrap resampling. A landmark analysis was performed at 12 months, excluding patients who relapsed within this period. The model is intended for application at the 12-month postbiopsy time point.
Results
A total of 831 patients were enrolled in the study, 225/831 patients (27.08%) experienced relapse. Compared with the patients without relapse, patients with relapse demonstrated younger age of onset (24.0 vs 27.0 years, p=0.001), higher SLE Disease Activity Index (SLEDAI) (12.0 vs 7.0, p=0.001) and proteinuria (1.2 vs 0.6 g/day, p=0.001). However, the nephritis histology showed no significant difference between the two groups. The proportion of patients receiving intravenous methylprednisolone pulse therapy or belimumab in the non-relapse group was significantly higher. Using LASSO-Cox regression, four variables associated with LN relapse were identified, including level of 24-hour proteinuria, SLEDAI 2000 (SLEDAI-2K), low complement C3 level and remission within 12 months. The predictive model exhibited good predictive performance, with a bootstrap-corrected C-index of 0.75 (95% CI 0.72 to 0.79).
Conclusions
Higher level of 24-hour proteinuria, higher SLEDAI-2K score and low complement C3 level were risk factors for LN relapse, while remission within 12 months was a protective factor. The LASSO-Cox-based predictive model exhibited good performance and may serve as a useful tool for the clinical management of LN.