Predicting nosocomial pneumonia of patients with acute brain injury in intensive care unit using machine-learning models

Junchen Pan ¹ Liqing Bi ² Zhen Yue ² Jing Ji ² Yongping You ^3* Yun Liu ¹ Xinglin Xiong ¹ Genying Gu ¹ Ming Chen ¹ Shen Zhang ⁴

• ¹ Nanjing BenQ Hospital, Nanjing, China
• ² First Affiliated Hospital, Nanjing Medical University, Nanjing, Jiangsu Province, China
• ³ Department of Neurosurgery, First Affiliated Hospital, Nanjing Medical University, Nanjing, China
• ⁴ Engineering Research Center of Health Service System Based on Ubiquitous Wireless Networks, Ministry of Education, Nanjing, China

The aim of this study is to construct and validate new machine learning models to predict pneumonia events in intensive care unit (ICU) patients with acute brain injury.: Acute brain injury patients in ICU of hospitals from January 1, 2020, to December 31, 2021 were retrospective reviewed. Patients were divided into training, and validation sets. The primary outcome was nosocomial pneumonia infection during ICU stay. Machine learning models including XGBoost, DecisionTree, Random Forest, Light GBM, Adaptive Boost, BP, and TabNet were used for model derivation. The predictive value of each model was evaluated using accuracy, precision, recall, F1-score, and area under the curve (AUC), and internal and external validation was performed. Results: A total of 280 ICU patients with acute brain injury were included. Five independent variables for nosocomial pneumonia infection were identified and selected for machine learning model derivations and validations, including tracheotomy time, antibiotic use days, blood glucose, ventilator-assisted ventilation time, and C-reactive protein. The training set revealed the superior and robust performance of the XGBoost with the highest AUC value (0.956), while the Random Forest and Adaptive Boost had the highest AUC value (0.883) in validation set.Machine learning models can effectively predict the risk of nosocomial pneumonia infection in patients with acute brain injury in the ICU. Despite differences in populations and algorithms, the models we constructed demonstrated reliable predictive performance.