Comparative analysis of X-ray and MRI in a nomogram-based approach for rheumatoid arthritis and lupus management

Rheumatoid arthritis (RA) and systemic lupus erythematosus (SLE) exhibit heterogeneous progression and treatment responses. This study develops a machine learning framework integrating clinical, laboratory, and imaging data (X-ray, MRI) to improve risk stratification for disease progression, relapse, and treatment response. A retrospective cohort of 1970 patients (RA: 1,200, SLE: 770) was analyzed using multimodal data from clinical records, laboratory biomarkers (CRP, ESR, autoantibodies), and imaging features extracted from deep learning-based models (ResNet, VGG, DenseNet, EfficientNet) and radiomics. A two-step feature selection approach using Intraclass Correlation Coefficient (ICC) filtering (≥0.75 for X-ray, ≥0.90 for MRI) and LASSO regression was applied. Machine learning models, including logistic regression (LR), random forest (RF), gradient boosting machine (GBM), support vector machine (SVM), decision tree (DT), knearest neighbors (KNN), deep neural networks (DNN), and extreme gradient boosting (XGB), were trained using stratified 5-fold cross-validation. Model interpretability was enhanced with SHapley Additive exPlanations (SHAP) analysis. A nomogram integrating deep learning-based imaging scores, patient demographics, and inflammatory biomarkers was constructed and validated with calibration curves and ROC analysis. A total of 1970 patients (RA: 1,200, SLE: 770) were analyzed, with clinical, laboratory, and imaging data (X-ray, MRI) integrated into predictive models. Deep learning (ResNet, VGG, DenseNet, EfficientNet) and radiomics-based features were optimized using ICC filtering and LASSO regression, reducing dimensionality while enhancing prediction accuracy. XGBoost and GBM achieved the highest AUC (0.98), accuracy (93.79 %), sensitivity (89.77 %), and specificity (95.03 %). SHAP analysis identified imaging biomarkers, CRP, and ESR as key predictors. The nomogram demonstrated excellent calibration, supporting its clinical utility for personalized risk stratification and treatment response prediction in RA and SLE management. This study presents a multimodal AI-driven predictive framework that integrates deep learning, radiomics, and clinical data for precise disease monitoring and treatment optimization in RA and SLE.