Prediction of the preclinical stage of coal workers’ pneumoconiosis on non-imaging data integrating prior knowledge and machine learning

Abstract Objective This study aims to establish machine learning models using non-imaging data from health examinations of coal workers, which can screen the preclinical stage of CWP. Methods Non-imaging data from two centers, totaling 34,362 coal miners, were collected. From 84 initial variables, 19 were preliminarily screened, and LASSO selected 8 key features. Six machine learning models were trained to predict the preclinical stage of CWP, evaluated using ROC curve. Results In the internal test set, GB achieved the best discrimination (AUC 88.19%), while DT yielded the highest accuracy (81.09%) and specificity (80.90%). In the external validation set, GB remained the top model by AUC (83.94%) and showed high sensitivity (87.67%). Conclusion Age, FEV1, FEV1%, drinking status, smoking status, FVC, occupational category, and cumulative years of service are significant features for predicting the preclinical stage of CWP.