Development of an interpretable machine learning model for predicting prolonged hospital stay in patients with acute exacerbation of chronic obstructive pulmonary disease: a retrospective cohort study.

objectiveProlonged hospital stay in patients with acute exacerbation of chronic obstructive pulmonary disease (AECOPD) significantly impacts patient outcomes and healthcare resource allocation. This study aimed to develop and validate an interpretable machine learning (ML) model for forecasting prolonged hospital stays of the AECOPD population.

methodsA retrospective analysis was performed utilizing data from the MIMIC-IV database on patients diagnosed with AECOPD. The dataset was split into a training set (80%) and a validation set (20%). Feature selection was executed through LASSO regression and the Boruta algorithm. Logistic regression (LR), random forest (RF), neural network (NN), gradient boosting machine (GBM), naive Bayes (NB), as well as K-nearest neighbors (KNN) models were constructed based on the selected features. Model performance was evaluated via receiver operating characteristic (ROC), calibration and decision curves. The interpretability of the optimal model was enhanced through SHapley Additive exPlanations (SHAP) analysis.

results7,373 AECOPD patients were encompassed. The final model incorporated the following features: hemoglobin, platelet count, anion gap, blood urea nitrogen(BUN), potassium, chloride, antibiotic use, invasive ventilation, vasopressor use, SOFA, body temperature(Tb), heart rate(HR), acute kidney injury (AKI), as well as sepsis. The RF model demonstrated the best predictive performance, with an area of the ROC curve (AUC) of 0.817 in the training set and 0.715 in the validation set, along with good calibration and clinical utility. SHAP analysis further enhanced the interpretability of the model, providing valuable clinical decision support.

conclusionThe RF model developed in our study exhibited excellent performance in predicting prolonged hospital stays in AECOPD patients. The incorporation of interpretability analysis improved the transparency and reliability of its clinical application. CLINICAL TRIAL NUMBER: Not applicable.