Session Information
Date: Tuesday, November 10, 2026
Title: (2129-2151) Vasculitis - ANCA-Associated - Clinical Poster C
Session Type: Poster Session C
Session Time: 10:30AM-12:30PM
Background/Purpose: Objective: To comprehensively identify prognostic risk factors and predict survival outcomes in antineutrophil cytoplasmic antibody-associated vasculitis (AAV) utilizing deep learning techniques.
Methods: Methods: We analyzed 1,140 patients from the prospective Chinese Registry of Systemic Vasculitis (CRSV). After selecting 17 baseline variables via Boruta and LASSO algorithms, we developed a deep learning survival model (CACox) to predict disease outcomes. The predictive performance of CACox was compared against the conventional Cox proportional hazards regression model.
Results: Results: Patients were randomized into training (n=912, median follow-up 1,016 days) and testing (n=228, median follow-up 1,060 days) cohorts. Overall 1-, 3-, and 5-year OS rates were 92.8%, 87.1%, and 83.1%, respectively. In the testing set, CACox achieved a C-index of 0.830 (95% CI: 0.758-0.901, p=0.038) versus the Cox model (0.784, 95% CI: 0.706-0.863). The 1-, 2-, and 3-year AUCs for CACox were 0.850 (95% CI: 0.737-0.948), 0.841 (95% CI: 0.754-0.919), and 0.837 (95% CI: 0.751-0.917). Kaplan-Meier curves demonstrated highly significant risk stratification (p < 0.0001). CACox decoded critical non-linear relationships overlooked by Cox regression. Specifically, age of onset (p=0.878 in Cox) and ALT (p=0.924) emerged as top predictors, revealing a 55-year threshold effect exacerbated by pulmonary fibrosis, alongside masked risk variance within normal ALT dictated by platelet abnormalities. A prognostic web application (http://47.95.220.248:8000/) was developed based on CACox.
Conclusion: Conclusion: The CACox deep learning model outperforms conventional regression in predicting AAV outcomes and achieving balanced risk stratification. This model serves as a robust tool to guide individualized clinical assessments.
SHAP summary plot illustrating the overall impact of the 17 selected features on the model’s predictions. The features on the y-axis are ranked in descending order of their global importance to the model, from highest at the top to lowest at the bottom. The x-axis represents the SHAP value, which quantifies the magnitude and direction of a feature’s effect on an individual prediction. Each dot corresponds to a single patient. The color gradient indicates the actual value of the feature for that patient, ranging from dark purple (low value) to yellow (high value). Points located to the right of the vertical zero line (SHAP value > 0) indicate an increased predicted risk of the event, whereas points to the left (SHAP value < 0) indicate a decreased predicted risk. Abbreviations: ALT, alanine aminotransferase; Cr, creatinine; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate; Hb, hemoglobin; PLT, platelet; SHAP, SHapley Additive exPlanations; WBC, white blood cell.
Screenshot of the interactive web-based risk calculator developed for individualized survival prediction. The user interface consists of three main modules. The left panel (“Patient characteristics”) allows clinicians to input specific demographic, clinical, and laboratory parameters for an individual patient. Upon execution, the top-right panel (“Prediction & survival estimates”) outputs the calculated linear predictor, assigns the patient to a specific risk stratum (e.g., Intermediate Risk), and provides quantitative estimated survival rates at 1 and 2 years. The bottom-right panel (“Overall survival Rate”) visualizes the patient’s predicted risk trajectory alongside the established Kaplan-Meier survival curves for the low-, intermediate-, and high-risk groups. This online application facilitates the clinical translation of the proposed deep learning model by providing real-time, point-of-care decision support.
Predictive performance of the machine learning and traditional models in the training and test sets.
To cite this abstract in AMA style:
Shang X. A deep learning model predicting the survival of Anti-neutrophil cytoplasmic antibody -associated vasculitis [abstract]. Arthritis Rheumatol. 2026; 78 (suppl 9). https://acrabstracts.org/abstract/a-deep-learning-model-predicting-the-survival-of-anti-neutrophil-cytoplasmic-antibody-associated-vasculitis/. Accessed .ACR Meeting Abstracts - https://acrabstracts.org/abstract/a-deep-learning-model-predicting-the-survival-of-anti-neutrophil-cytoplasmic-antibody-associated-vasculitis/
