Development and validation of an interpretable machine learning model for pulmonary heart disease in patients with pneumoconiosis.

Background: Pneumoconiosis patients with pulmonary heart disease (PHD) face an elevated risk of mortality. Early identification of individuals at higher risk of PHD is important for timely risk stratification and clinical management. However, few risk-screening models have been specifically developed for pneumoconiosis in primary care settings, where access to advanced diagnostic resources may be limited. Therefore, this study developed and validated a machine learning model based on routinely available data to support early risk stratification and clinical decision-making in resource-limited environments. Method: A dual-design occupational pneumoconiosis cohort, comprising retrospective and prospective components, was established using occupational, epidemiological, and laboratory data obtained from electronic medical records and questionnaires. Key variables were selected using least absolute shrinkage and selection operator (LASSO) regression; five machine learning models were developed using five-fold cross-validation with grid search for hyperparameter tuning. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), F1 score, accuracy, precision-recall area under the curve (PR AUC), calibration metrics, and threshold analysis. The prospective cohort was used to evaluate the generalizability of the models. Model interpretability was assessed by Shapley additive explanations (SHAP). Results: Following strict application of the inclusion and exclusion criteria, the retrospective cohort included 984 hospitalized patients (2012-2021), whereas the prospective cohort comprised 829 patients enrolled since 2021. LASSO identified 10 variables for model construction. All models demonstrated satisfactory discrimination, with logistic regression (LR) achieving an AUC of 0.816 in the external validation cohort. Lower BMI, albumin, and platelet large cell ratio (P-LCR), as well as higher red blood cell distribution width coefficient of variation (RDW-CV), older age, and stage III pneumoconiosis, were associated with higher predicted risk of PHD. The SHAP interaction network further indicated prominent statistical interaction patterns among hematological abnormalities, nutritional status, and disease severity, based on their joint contributions to model predictions. Conclusion: This study indicates that machine learning models, particularly LR, may be useful for PHD risk screening and stratification among patients with pneumoconiosis. SHAP analysis identified clinically accessible variables that were important for model predictions, indicating a multidimensional pattern of associations across hematological, nutritional, and disease-severity domains.