Prediction of Prolonged Length of Hospital Stay in Patients with Acute Exacerbations of Chronic Obstructive Pulmonary Disease: An Interpretable Machine Learning Tool.

Objective: Reducing the length of hospital stay (LOS) is a core objective in the management of acute exacerbations of chronic obstructive pulmonary disease (AECOPD). This study aimed to develop an interpretable and clinically applicable machine learning tool for predicting prolonged LOS in this population. Methods: This retrospective three-center study enrolled 1342 patients, who were randomly allocated to a training set (70%) and a test set (30%). Candidate predictors were screened using least absolute shrinkage and selection operator (LASSO) regression, and six machine learning models were developed and compared: logistic regression, random forest, gradient boosting machine, CatBoost, support vector machine, and neural network. Model discrimination was evaluated using the area under the receiver operating characteristic curve (AUC), calibration plots, and decision curve analysis. Model interpretability was achieved through SHapley Additive exPlanations (SHAP) and subgroup analyses. Results: Prolonged LOS, defined as >10 days (the median LOS in the training set), occurred in 42.0% of patients. Nine predictors were identified, including daily inhaled medication use, sputum microbiological examination, antibiotic administration, corticosteroid therapy, diuretic use, ICU admission, oxygenation index, platelet count, and neutrophil count. The random forest model demonstrated superior and consistent discriminative performance, achieving an AUC of 0.778 (95% CI: 0.748-0.807) in the training set and 0.721 (95% CI: 0.672-0.771) in the test set. SHAP analysis ranked diuretic use, daily inhaled medication use, corticosteroid therapy, sputum microbiological examination, and antibiotic administration as the five most influential features, revealing treatment-related variables associated with prolonged LOS. Subgroup analyses indicated better predictive performance in lower-risk patients (age ≤ 65 years with eGFR 1-2 stage). Conclusions: Random Forest may enable to promptly identify prolonged LOS high-risk patients, enhancing clinical vigilance and optimizing healthcare resource allocation in AECOPD patients.