Abstract
This study aimed to explore the independent risk factors for community-acquired pneumonia (CAP) complicated with acute respiratory distress syndrome (ARDS) and to predict and evaluate the risk of ARDS in CAP patients based on artificial neural network models (ANNs). We retrospectively analyzed eligible 989 CAP patients (632 men and 357 women) who met the criteria from the comprehensive intensive care unit (ICU) and the respiratory and critical care medicine department of Changzhou Second People's Hospital, Jiangsu Provincial People's Hospital, Nanjing Military Region General Hospital, and Wuxi Fifth People's Hospital between February 2018 and February 2021. The best predictors to model the ANNs were selected from 51 variables measured within 24 h after admission. By using this model, patients were divided into a training group (n = 701) and a testing group (n = 288 patients). Results showed that in 989 CAP patients, 22 important variables were identified as risk factors. The sensitivity, specificity, and accuracy of the ANNs model training group were 88.9%, 90.1%, and 89.7%, respectively. When ANNs were used in the test group, their sensitivity, specificity, and accuracy were 85.0%, 87.3%, and 86.5%, respectively; when ANNs were used to predict ARDS, the area under the receiver operating characteristic (ROC) curve was 0.943 (95% confidence interval (0.918-0.968)). The nine most important independent variables affecting the ANNs models were lactate dehydrogenase (100%), activated partial thromboplastin time (84.6%), procalcitonin (83.8%), age (77.9%), maximum respiratory rate (76.0%), neutrophil (75.9%), source of admission (68.9%), concentration of total serum kalium (61.3%), and concentration of total serum bilirubin (50.4%) (all important >50%). The ANNs model and the logistic regression models were significantly different in predicting and evaluating ARDS in CAP patients. Thus, the ANNs model has a good predictive value in predicting and evaluating ARDS in CAP patients, and its performance is better than that of the logistic regression model in predicting the incidence of ARDS patients.