Abstract
OBJECTIVE: The purpose of this study was to develop and validate machine learning models that can predict superaverage length of stay in hypercapnic-type respiratory failure and to compare the performance of each model. Furthermore, screen and select the optimal individualized risk assessment model. This model is capable of predicting in advance whether an inpatient's length of stay will exceed the average duration, thereby enhancing its clinical application and utility. METHODS: The study included 568 COPD patients with hypercapnic respiratory failure, 426 inpatients from the Department of Respiratory and Critical Care Medicine of Yancheng First People's Hospital in the modeling group and 142 inpatients from the Department of Respiratory and Critical Care Medicine of Jiangsu Provincial People's Hospital in the external validation group. Ten machine learning algorithms were used to develop and validate a model for predicting superaverage length of stay, and the best model was evaluated and selected. RESULTS: We screened 83 candidate variables using the Boruta algorithm and identified 9 potentially important variables, including: cerebrovascular disease, white blood cell count, hematocrit, D-dimer, activated partial thromboplastin time, fibrin degradation products, partial pressure of carbon dioxide, reduced hemoglobin, and oxyhemoglobin. Cerebrovascular disease, hematocrit, activated partial thromboplastin time, partial pressure of carbon dioxide, reduced hemoglobin and oxyhemoglobin were independent risk factors for superaverage length of stay in COPD patients with hypercapnic respiratory failure. The Catboost model is the optimal model on both the modeling dataset and the external validation set. The interactive web calculator was developed using the Shiny framework, leveraging a predictive model based on Catboost. CONCLUSION: The Catboost model has the most advantages and can be used for clinical evaluation and patient monitoring.