[Machine learning-based prediction model for respiratory failure in patients with pneumoconiosis]

Objective: To develop an interpretable machine learning model to predict the risk of respiratory failure in patients with pneumoconiosis. Methods: In February 2025, based on the retrospective West China Occupational Pneumoconiosis Cohort, 1367 hospitalized pneumoconiosis patients in West China Fourth Hospital of Sichuan University from January 1, 2012 to November 30, 2021 were selected as the research objects. Predictive variables were selected using least absolute shrinkage and selection operator (LASSO) -Cox regression. Four machine learning models, including random survival forest model (RSF), gradient boosting survival model (GBM), extreme gradient boosting survival model (XGBoost), and survival support vector machine (SSVM), as well as a traditional Cox proportional hazards regression model (CPH) were constructed. The performance of the model was evaluated using the concordance index (C-index), Brier score, area under the curve (AUC) of the receiver operating characteristic curve, and decision curve analysis (DCA). Shapley additive explanations (SHAP) were used to analyze the interpretability of the model. The stability of the model was tested using an external validation set. Results: The median follow-up time for pneumoconiosis patients was 2.69 years. During the follow-up period, 197 patients were complicated with respiratory failure, and the cumulative incidence rate was 14.41%. Fifteen variables were selected by LASSO-Cox regression and included in the prediction models, mainly including age, pneumoconiosis stage, albumin, and concurrent pulmonary hypertension. The model evaluation results showed that the C-index of the GBM model in the test set was 0.794, the Brier score was 0.044, and the 3-year AUC was 0.788, with better performance than other models. DCA showed that in the probability range of 0% to 50%, the GBM model had a high clinical net benefit in predicting the risk of respiratory failure. In the external validation, the C-index of the GBM model was 0.644. The top three influential factors in the SHAP feature importance ranking were pneumoconiosis stage, age, and body mass index. Conclusion: The prediction model for the concurrent respiratory failure in pneumoconiosis patients constructed based on GBM has high reliability and accuracy.