Developing and externally validating machine learning models to forecast short-term risk of ventilator-associated...

Area of Science:

• Critical Care Medicine
• Health Informatics
• Machine Learning in Healthcare

Background:

• Ventilator-associated pneumonia (VAP) is a severe hospital-acquired infection in ICUs, associated with high morbidity and mortality.
• Current methods for early VAP detection are insufficient, limiting timely interventions.
• Identifying patients at risk for VAP requires improved tools for early physiologic signal detection.

Purpose of the Study:

• To develop supervised machine learning models for predicting the short-term onset of VAP.
• To leverage routinely collected electronic health record (EHR) data for VAP prediction.
• To assess the generalizability of predictive models across different healthcare settings.

Main Methods:

• Analysis of EHR data from a prospective observational ICU cohort with physician-adjudicated VAP diagnoses.
• Extraction of clinical features including vital signs, ventilator settings, and laboratory values.
• Development and validation of machine learning models using various prediction windows (3, 5, 7 days prior to VAP) and external datasets (MIMIC-IV, AMIKINHAL trial).

Main Results:

• The best model achieved an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.866 for predicting VAP up to seven days before diagnosis in the internal cohort.
• External validation in MIMIC-IV showed an AUROC of 0.817 for predicting VAP within five days.
• Key predictive features included platelet count, positive end-expiratory pressure (PEEP), ventilator duration, and inflammatory markers.

Conclusions:

• Machine learning models can effectively predict VAP onset up to a week in advance using standard ICU data.
• Model performance demonstrated generalizability to a different hospital system, though it was limited by feature overlap.
• Future research should focus on real-time prospective evaluation of these predictive models.