Developing a prediction model for poor prognosis in MPA patients using initial admission examination results: a machine learning study from Southwest China.

Background: Microscopic polyangiitis (MPA) is one of the main types of ANCA-associated vasculitis (AAV), but current admission examination indicators are limited in predicting poor prognosis for MPA. This study aims to develop a prediction model for adverse prognosis in MPA patients using initial admission examination results. Methods: We performed machine learning (ML) algorithms on initial admission examination data to predict adverse outcomes in MPA, such as in-hospital death or self-discharge due to critical condition. We analyzed data from 12,497 patients who underwent ANCA tests in Deyang People's Hospital between November 2017 and October 2025, focusing on 230 hospitalized patients. We used least absolute shrinkage and selection operator (LASSO), logistic regression (LR), and random forest (RF) to select variables, and evaluated the diagnostic efficacy using ML algorithms. SHapley Additive exPlanations (SHAP) was used for model interpretability. A nomogram model was developed to predict adverse outcomes, highlighting variable contributions and including calibration and decision curve analysis (DCA) curves. Results: In this study of 230 MPA patients, 56 had poor prognosis while 174 had good prognosis. Seven indicators were identified using LASSO, LR and RF. They were appropriate use of immunosuppressants, age, serum albumin, infection, BVAS, C-reactive protein (CRP) and anti- myeloperoxidase. Among five ML models, support vector classification (SVC) had a high area under the curve (AUC) (AUC = 0.848, 95% confidence interval [CI]: 0.684-0.965) in the prediction model, and had the highest AUC (AUC = 0.886, 95% CI: 0.741-0.981). The SHAP analysis highlighted elevated CRP levels as the top predictor of poor prognosis. Standardized immunosuppressive therapy was found to mitigate this risk. Nomogram model confirmed these findings, and calibration and DCA curves showed this model was reliable and useful for clinical decisions. Conclusions: We developed a prediction model for adverse outcomes in MPA patients, utilizing clinical and laboratory data collected on the day of admission. SVC algorithm exhibited moderate predictive efficacy. Factors such as elevated serum CRP levels, moderate reductions in serum albumin, infection, advanced age, BVAS larger than 15, and anti-MPO were positively associated with adverse prognoses. Standardized immunosuppressive therapy was shown to mitigate this risk, offering a valuable reference for clinical decision-making.