Background: Organ damage is a key determinant of unfavorable long-term prognosis and increased mortality, thus being reflective of disease severity in SLE patients.

Objectives: To develop a clinical, machine learning based model for the prediction of early organ damage in SLE towards disease severity stratification.

Methods: Using a cohort of 914 adults with SLE [1], panels of deconvoluted classification criteria [ACR 1997 (ACR), SLICC 2012 (SLICC) and EULAR/ACR 2019 (EULAR)] and non-criteria features present at any timepoint throughout the first five years since SLE diagnosis were assessed. Permanent organ damage was evaluated using the SLICC/ACR Damage Index (SDI). We randomly divided the patient cohort into a training (70%) and a test (30%) set. Employing feature selection algorithms, the smallest set of clinical features that most accurately predicted early organ damage accrual (defined as SDI increase within the first 5 years since SLE diagnosis) was selected. Five different prediction models (random forest (RF), logistic regression (glm), linear discriminant analysis (LDA), k-nearest neighbors (KNN), extreme gradient boosting (XGBoost)) were adopted. The best model in 10-fold cross-validation was tested in the test set. Accuracy, sensitivity, specificity, and area under (AUC) the receiver operating curve (ROC) were determined in the test set.

Results: The LDA model demonstrated the highest performance in predicting early organ damage with an AUC of 0.831 (95% CI: [0.7817, 0.8739]) with sensitivity of 0.955 and specificity of 0.463. The leading predictors included synovitis, non-scarring alopecia, acute cutaneous lupus, SLICC 2012-based neurologic disorder, leukopenia, and the age at the time of SLE diagnosis. The XGBoost model exhibited the highest specificity (0.841) with an accuracy of 0.805 (95% CI: [0.7619 - 0.8241]) and sensitivity of 0.653. Age at the time of diagnosis, the presence of ACR 1997-based neurologic disorder, non-scarring alopecia, and low complement (2012 criteria) emerged as the strongest predictors in this model.

Conclusion: Machine learning methods using standard disease features may identify SLE patients at risk for early damage accrual. Further validation in external cohorts is warranted.

REFERENCES: [1] Adamichou C, Genitsaridi I, Nikolopoulos D, Nikoloudaki M, Repa A, Bortoluzzi A, Fanouriakis A, Sidiropoulos P, Boumpas DT, Bertsias GK. Lupus or not? SLE Risk Probability Index (SLERPI): a simple, clinician-friendly machine learning-based model to assist the diagnosis of systemic lupus erythematosus. Ann Rheum Dis. 2021 Jun;80(6):758-766. doi: 10.1136/annrheumdis-2020-219069. Epub 2021 Feb 10. PMID: 33568388; PMCID: PMC8142436.

Acknowledgements: NIL.

Disclosure of Interests: None declared.