Background: Cardiovascular events (CVE) are one of the major cause of mortality in Systemic lupus erythematosus (SLE). Traditional cardiovascular risk (CVR) assessment tools are insufficient for SLE patients, prompting the need for innovative predictive models [1-5].

Objectives: This study aims at using using machine learning (ML) modelst to identify key predictors of CVE encompassing both traditional CVR factors and SLE-specific parameters. Additionally, the study compares the predictive abilities of the SLEDAI-2K and SLE-DAS indices for CVE.

Methods: Data from 2000 to 2023 from 176 patients at the Lupus clinic of University hospital) were analysed retrospectively. Patient demographics, medications, and key CVR factors were recorded. Disease activity was assessed using SLEDAI-2K and SLE-DAS, with remission following the Definitions Of Remission In SLE (DORIS) criteria, and cumulative damage assessed via the SLE Damage Index (SDI). The primary outcome was new CVE occurrences. To idenfy the predictors oft he occurrence of CVE three ML methods were considered: Support Vector Machine (SVM), Random Forest (RF), and k-Nearest Neighbor (KNN), with RF selected for detailed analysis. Model interpretation used SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDP). The logistic regression model (LRM) was used as a comparator despite the limited efficacy in predicting rare events.

Results: In our study, 18 patients experienced a CVE, with specific events including 5 myocardial infarctions, 1 pulmonary thrombo-embolism, 3 strokes, 5 transient ischemic attacks, and 4 cases of peripheral artery disease.

The RF model emerged as the most effective when evaluating ML techniques. It exhibited the highest area under the curve (AUC) in the Receiver Operating Characteristic (ROC) analysis in differentiating between patients with and without CVE. Specifically, the RF model achieved an accuracy of 88% and a ROC AUC score of 71.2%. In contrast, LRM, a more traditional statistical method, showed an accuracy of 97.22% but an ROC AUC of only 0.5, indicating its limited effectiveness due to the small sample size and rare events.

The RF model’s predictive power was further explored through Recursive Feature Elimination with Cross-Validation (RFECV). This process identifies the most influential features in a dataset for the model. Seven features were found to be pivotal: age, disease duration, SDI, SLEDAI-2K, SLE-DAS) prednisone dosage, and DORIS Remission status.

We then compared the predictive capabilities of two key disease activity indices: mean SLEDAI-2K and mean SLE-DAS. The ROC AUC scores were 0.6742 for mean SLE-DAS and 0.5303 for mean SLEDAI-2K, indicating a comparable predictive value with a slight edge for SLE-DAS. However, SHAP analysis, which explains the impact of each variable on the model’s predictions, revealed that mean SLEDAI-2K had a higher impact on predicting CVEs.

PDPs were used to understand how changes in SLEDAI-2K and SLE-DAS scores influence the likelihood of experiencing a CVE and suggest a non-linear relationship for SLEDAI-2K but a more linear dependency for SLE-DAS.

Conclusion: Machine learning demonstrates significant potential in addressing scientific challenges, particularly in identifying predictors of rare events in small sample sizes. Our study confirms that characteristics associated with SLE, such as steroid usage and disease activity, are crucial determinants in the occurrence of cardiovascular events. This technology paves the way for developing novel predictive tools for cardiovascular events and, potentially, disease flares, especially when applied to larger cohorts. The integration of machine learning in this field may lead to more personalized and effective management strategies for patients with SLE.

REFERENCES: [1] Restivo V. et al. Autoimmun Rev 2022; 21:102925.

[2] Lee YH et al. Lupus 2016; 25:727–734.

[3] Drosos GC et al. Ann Rheum Dis 2022; 81:768–779.

[4] Urowitz MB et al. J Rheumatol 2016; 43:875–879.

[5] Navarini L et al. Rheumatol Ther 2020; 7:867–882.

Acknowledgements: NIL.

Disclosure of Interests: None declared.