To develop and verify machine learning(ML)models for the early warning of ventilator-associated pneumonia(VAP)within 24 hours after invasive mechanical ventilation, so as to provide more evidence and ideas for the clinical management of VAP in elderly patients.

In this study, clinical data of elderly patients with acute respiratory failure and invasive mechanical ventilation in intensive care unit were extracted from MIMIC Ⅳ 2.2 database.Using VAP as the outcome index, patients were divided into training set and testing set in a ratio of 7∶3.Four ML algorithms were used to build a model in the training set, and the performance of the model was verified by the test set.The model was compared with SOFA, systemic inflammatory response syndrome(SIRS) and acute physiology score(APS)Ⅲ scores in the same dataset.

A total of 1 859 elderly patients were included, 336 of whom were diagnosed with VAP.The area under the curve(AUC)of the receiver operator characteristic curve of ML models were higher than the clinical risk scores(SOFA score: 0.44, SIRS score: 0.49, APS Ⅲ score: 0.46), and the LightGBM model and XGBoost model had better predictive performance, with AUC of 0.85(95%CI: 0.82, 0.88)and 0.84(95%CI: 0.81, 0.87). SHAP was used to further explain the model.The results showed that SOFA neurological score, maximum white blood cell count, maximum respiratory rate, maximum alkali residual and age were important factors for early prediction of elderly VAP.

In this study, ML algorithms were used to build an early warning model of VAP in elderly patients, which has important guiding significance for clinical timely initiation and adjustment of treatment plan.In the future, external verification of the model should be further carried out.

Pneumonia, ventilator-associated; Artificial intelligence; Warning models