ADVISE: A Machine Learning Framework for Early Recognition of a Surrogate Marker for Ventilator-Associated Pneumonia Using Routinely Collected Critical Care Data

Ventilator-associated pneumonia (VAP) is the most common nosocomial infection in intensive care units, affecting 20-36% of mechanically ventilated patients. Researchers developed ADVISE, a machine learning model to predict physiological deterioration associated with VAP using electronic health record data. The study included 3,566 patient admissions with 33,208 candidate 48-hour observation blocks. The model was trained on six variables including oxygen concentration, ventilator mode, pressure ratio, and procalcitonin. On test data, the model achieved AUROC of 0.874, representing a 4-fold improvement over baseline. At 80% sensitivity, the model detected 4 of 5 VAP cases with approximately 80 false alarms. The study demonstrated that routinely collected data on oxygenation and ventilatory support can identify patients at risk of VAP-related deterioration.