Real-time prediction of ventilator-associated pneumonia onset in ICU: development of a dynamic machine learning model

BACKGROUND: Ventilator-associated pneumonia (VAP) is a preventable complication of invasive mechanical ventilation (IMV) with a significant healthcare impact. Early risk prediction is crucial, but current models lack real-time adaptability. This study develops a real-time VAP prediction model using machine learning and high-resolution EHR data from the MIMIC-III database. METHODS: We analyzed 3523 ICU stays (3204 patients) from MIMIC-III, including adults who received IMV for at least 48 h. VAP was labeled based on microbiological cultures and antibiotic initiation. A real-time ensemble model of XGBoost regressors was developed to predict time to VAP onset, incorporating vital signs, ventilator data, and lab results. Two static classifiers (24 h and 48 h) were also compared. RESULTS: VAP occurred in 595 ICU stays (16.89%), with an incidence rate of 23.77 per 1000 IMV-days. Median VAP onset was 113.5 h post-IMV. The real-time model outperformed static models with a C-index of 0.68, AUROC of 0.71, and AUPRC of 0.36. It provided a median lead time of 53 h before VAP onset, with key predictors including temperature, respiratory rate, and minute ventilation. CONCLUSION: We present a real-time VAP prediction model that outperforms static classifiers, providing actionable lead time for proactive microbiological surveillance. The model enables risk stratification for enhanced monitoring and, when clinically indicated, timely targeted antimicrobial therapy. Future work will focus on multicenter prospective validation and integration into ICU workflows to assess clinical utility and impact on patient outcomes.