Machine learning-based multiclass model for autoimmune disease diagnosis and classification through nailfold videocapillaroscopy features

Objective To develop and validate a predictive model for distinguishing controls (Ctr), rheumatoid arthritis (RA) and systemic lupus erythematosus (SLE) based on nailfold videocapillaroscopy (NVC) image features. Methods A total of 600 NVC images from 396 participants (Ctr=117, RA=337 and SLE=146) were collected and divided into training and test sets at a 7:3 ratio. Nine NVC features were extracted, and an eXtreme Gradient Boosting multiclassification model was constructed to distinguish the three groups. SHapley Additive exPlanations (SHAP) analysis was performed to evaluate feature importance and interpret the model. Results Seven NVC features showed significant differences among the groups. The model achieved macro area under the curve values of 0.96 and 0.80 in the training and test sets, respectively. SHAP analysis identified papilla shape, red blood cell aggregation, number of capillary loops, number of crossed capillary loops and subpapillary venous plexus (SVP) as key features among the groups. Each group was characterised by specific NVC patterns. In Ctr, papilla shape emerged as the key feature and showed correlations with neutrophils, white blood cells and monocytes. In patients with RA, the number of crossed capillary loops was the most prominent feature and correlated with erythrocyte sedimentation rate, complement levels (C3 and C4) and inversely with immunoglobulin G. In patients with SLE, the SVP was the dominant feature and effectively distinguished SLE from both Ctr and RA. Conclusions This study developed a robust multiclassification model for differentiating autoimmune diseases using NVC features. The findings enhance our understanding of microvascular alterations and provide a potential tool for clinical diagnosis and disease monitoring.