Work

Using machine learning models to predict the risk of mild cognitive impairment for COPD patients: A multi-cohort study

Year: 2025

Type: article

Abstract: The evaluation of cognitive function in patients with chronic obstructive pulmonary disease (COPD) is critical. This study aimed to construct a risk prediction model capable of predicting the rsik of mild cognitive impairment (MCI) in COPD patients using machine learning methods This study enrolled 2142 COPD patients from the China Health and Retirement Longitudinal Study (CHARLS), analyzing 16 sociodemographic, health, and psychological indicators. The predictors screened by LASSO regression were integrated with five machine learning models—Logistic Regression, Neural Network, XGBoost, LightGBM, and Random Forest —to identify the optimal model for predicting MCI risk in COPD patients. External validation was conducted using the Health and Retirement Study dataset (HRS). Model performance underwent comprehensive evaluation across three critical dimensions: discrimination capacity (AUC-ROC), calibration accuracy (Brier score and calibration curves), and clinical applicability (decision curve analysis). Additionally, the SHAP method was employed to elucidate the feature contributions to the final model. The analysis included 2,142 COPD patients, of whom 420 were diagnosed with MCI. 11 predictive variables were selected to construct five machine learning algorithms. Among these, the LightGBM model demonstrated superior performance across all evaluation dimensions - including discrimination capacity, calibration accuracy, and clinical utility - achieving AUROC values ranging from 0.722 to 0.782.The eight most significant features for predicting MCI in COPD patients were sex, residence, instrumental activities of daily living (IADL) score, educational level, social activity engagement, marital status, Internet use and depression. The machine learning model accurately predicts MCI risk in COPD patients, enabling timely clinical interventions and improved care through early identification of high-risk individuals. • The model was developed and externally validated using large-scale national cohorts from China (CHARLS) and the USA (HRS), demonstrating robust and stable predictive performance across diverse populations. • Developed an LightGBM model with AUC score ranging from 0.722 to 0.782 for predicting mild cognitive impairment in COPD patients. • Relying solely on easily obtainable clinical and demographic factors makes the model ideal for practical early screening and risk stratification in primary care settings. • The Shapley Additive exPlanations (SHAP) method was employed to elucidate feature contributions to the final model.

Source: Comprehensive Psychoneuroendocrinology

DOI: 10.1016/j.cpnec.2025.100323