Abstract
Acute exacerbation of chronic obstructive pulmonary disease (AECOPD) is a leading cause of hospitalization and death in COPD patients. Machine learning (ML) approach is powerful but has a "black box" issue with an undirect interpretation of the ML technique. Herein, we conducted a multicentre, retrospective cohort study in two tertiary hospitals across China, primarily utilizing echocardiographic variables to build and validate an explainable prediction model based on a ML approach to predict the hospitalization death of AECOPD. For model explainability, we utilized a model-agnostic SHapley Additive exPlanations explainer to interpret the output of our final model. Our results showed that the light gradient boosting machine (LightGBM) model achieved the best performance among the 11 ML models. After reducing features according to the feature importance rank, an explainable final LightGBM model was established with 9 features (AUC = 0.956, accuracy = 92.1%, sensitivity = 0.891, specificity = 0.933, PPV = 0.852, NPV = 0.952, F1 score = 0.871). To facilitate its utility for clinicians, this final explainable model had been translated into a convenient application. In addition, the LightGBM model mitigated the concern of the "black-box" via a global and a local explanation of the SHAP method. A publicly accessible web tool was generated for the model. These findings further hold promise for guiding clinical management and improving patient outcomes.