Development and Deployment of a Predictive Model for Ventilator-Associated Pneumonia (VAP) Using Clinical Information System Data and Real-Time Mapping of Infection Variables

Background: Ventilator-associated pneumonia (VAP) is one of the most common infections in intensive care units (ICUs), with an estimated attributable mortality of approximately 10%. It is associated with prolonged duration of mechanical ventilation (MV) and ICU stay, and with increased difficulty in early and adequate antibiotic (AB) treatment. The use of machine learning techniques could enable the early prediction of patients at high risk of VAP and the activation of protocols aimed at confirming the microbiological diagnosis and initiating early appropriate treatment. Our objective is to develop a real-time VAP predic-tion model as a decision support tool for clinicians using data from the clinical information system (CIS). Methods: All pa-tients from 1/1/2014 to 31/12/2024 who required MV for more than two days were included in the study. Day 0 (zero) was de-fined as the clinical diagnosis of VAP, and the clinical and labor-atory variables included in the models were considered with windows of 24, 48 and 72 hours prior to day zero, respectively. These were obtained automatically from the CIS through an ETL (Extract, Transform and Load) in a Python/Jupyter environment. The data was divided into a training set (Train= 80%) and a test set (Test=20%). The imbalance of the positive class (VAP) was corrected in Train by applying a down-sampling adjustment. The models developed using XGBoost were evaluated to inves-tigate their effectiveness in predicting VAP 24, 48 and 72 hours before the date of the event using accuracy, recall, confusion matrix and area under the ROC curve (AUC). Results: 2,714 pa-tients were included in the cohort, of which 314 (11.6%) had VAP. An XGBoost model was developed, achieving accuracies of 0.84, 0.71 and 0.70 for predictions made within the time win-dows of 24, 48 and 72 hours before the onset of VAP, respective-ly. The model also demonstrated high recovery rates of 0.83, 0.76 and 0.73 for these intervals, and showed excellent discriminatory power between the two classes, with AUC values of 0.90, 0.77 and 0.75, highlighting its great predictive capacity in the early detection of VAP. Conclusions: The development of 3 real-time VAP prediction models using machine learning showed ade-quate prediction especially in the 24 hours prior to clinical diag-nosis. This decision support tool can have a favorable impact on the prognosis of patients with VAP. However, these models must be prospectively validated before being implemented in clinical practice.