Machine learning improves 10-year prediction of atherosclerotic plaque progression and cardiovascular events in systemic lupus erythematosus
Nikolaos Papazoglou, Petros P Sfikakis, Maria G TektonidouAbstract
Objectives
To evaluate the performance of generic and disease-adapted tools in predicting 10-year cardiovascular disease (CVD) events and atherosclerotic plaque progression risk in systemic lupus erythematosus (SLE), and develop a machine learning-based model to improve cardiovascular risk stratification.
Methods
In a prospective 10-year follow-up carotid and femoral ultrasound study of patients with SLE without prior CVD events, we assessed the performance of 11 CVD risk scores (7 conventional, 4 SLE-adapted) to predict atherosclerotic plaque progression and incident CVD events. We used discrimination (sensitivity, specificity, area under the curve [AUC], Youden index), calibration (Brier Score [BS]) and classification (Matthews correlation coefficient, F1 score) metrics, and logistic regression and time-to-event analyses. A machine learning-based gradient boosting model was developed and internally validated.
Results
Among 111 patients (91% female, median age: 43.0 years) with 10-year ultrasonography follow-up, 53 (47.7%) developed new plaques, and 8 (7.2%) had incident CVD events. All scores demonstrated suboptimal performance; sensitivity: 0.00–0.44, AUC 0.58–0.79, and BS: 0.44–0.71. In multivariable analysis, QRISK3 was the only tool independently associated with CVD events and/or plaque progression (HR:1.87, p = 0.027). A machine learning-based score that integrates QRISK3 with disease-related risk factors (disease duration, cumulative glucocorticoid dose, DORIS achievement, antiphospholipid antibodies, hydroxychloroquine) and additional traditional CVD risk factors (physical activity, waist circumference), demonstrated significantly better performance compared to all evaluated scores.
Conclusion
Conventional and SLE-adapted tools underestimate CVD events and atherosclerosis progression risk in SLE. A machine learning-based model incorporating disease-specific and traditional cardiovascular risk factors into QRISK3 improves CVD risk stratification.