Construction and validation of a risk prediction model for chronic obstructive pulmonary disease (COPD): a cross-sectional study based on the NHANES database from 2009 to 2018.

BACKGROUND

Chronic obstructive pulmonary disease (COPD) is a major global public health concern, and early screening and identification of high-risk populations are critical for reducing the disease burden. Although several studies have explored the application of machine learning methods in COPD risk prediction, existing models often have limited feature dimensions and insufficient interpretability. Identifying key risk factors and constructing reliable predictive models remain challenges in clinical practice.

OBJECTIVE

This study aims to integrate multidimensional features based on data from the National Health and Nutrition Examination Survey (NHANES) and to compare the performance of different machine learning models in COPD risk prediction. The goal is to identify the optimal model and enhance its clinical applicability through interpretability analysis.

METHODS

This study utilized data from the NHANES collected between 2009 and 2018. After systematic feature selection and preprocessing, three models were developed: multivariate binary logistic regression, XGBoost, and Multilayer Perceptron (MLP). Model training and evaluation were performed using stratified five-fold cross-validation. Model performance was comprehensively assessed based on accuracy, precision, recall, F1 score, and the area under the receiver operating characteristic curve (AUC). To enhance model transparency, the SHapley Additive Explanations (SHAP) method was employed to interpret key features and their influence trends within the MLP model.

RESULTS

The MLP model demonstrated the best performance across all evaluation metrics, achieving an average accuracy of 0.937, precision of 0.6624, recall of 0.6535, and F1 score of 0.657 in stratified five-fold cross-validation. The performance gap between the training and testing sets was minimal, indicating no obvious overfitting. SHAP analysis identified smoking years, asthma, age, dietary health status, total protein, red cell distribution width (RDW), BMI, marital status, secondhand smoke exposure, and total bilirubin as important predictive features. Furthermore, dependence plots revealed critical risk inflection points for key continuous variables.

CONCLUSION

Based on large-scale and multidimensional feature data, this study constructed a COPD risk prediction model with favorable performance and enhanced interpretability. The findings suggest that the MLP model has the potential to effectively identify individuals at high risk for COPD and may offer value in clinical applications. Future studies are warranted to integrate longitudinal follow-up data and multimodal information to further improve predictive accuracy and clinical interpretability, thereby providing a more robust foundation for early screening and personalized interventions in COPD.