RESUMO
BACKGROUND: This study was performed to develop and validate machine learning models for early detection of ventilator-associated pneumonia (VAP) 24 h before diagnosis, so that VAP patients can receive early intervention and reduce the occurrence of complications. PATIENTS AND METHODS: This study was based on the MIMIC-III dataset, which was a retrospective cohort. The random forest algorithm was applied to construct a base classifier, and the area under the receiver operating characteristic curve (AUC), sensitivity and specificity of the prediction model were evaluated. Furthermore, We also compare the performance of Clinical Pulmonary Infection Score (CPIS)-based model (threshold value ≥ 3) using the same training and test data sets. RESULTS: In total, 38,515 ventilation sessions occurred in 61,532 ICU admissions. VAP occurred in 212 of these sessions. We incorporated 42 VAP risk factors at admission and routinely measured the vital characteristics and laboratory results. Five-fold cross-validation was performed to evaluate the model performance, and the model achieved an AUC of 84% in the validation, 74% sensitivity and 71% specificity 24 h after intubation. The AUC of our VAP machine learning model is nearly 25% higher than the CPIS model, and the sensitivity and specificity were also improved by almost 14% and 15%, respectively. CONCLUSIONS: We developed and internally validated an automated model for VAP prediction using the MIMIC-III cohort. The VAP prediction model achieved high performance based on its AUC, sensitivity and specificity, and its performance was superior to that of the CPIS model. External validation and prospective interventional or outcome studies using this prediction model are envisioned as future work.