VAP: Early identification of ventilator associated pneumonia using machine learning techniques: a prospective cohort study

Funded by Innosuisse - Swiss Innovation Agency

vap

Ventilator-associated pneumonia (VAP) represents the most prevalent hospital-acquired infection in the intensive care unit (ICU). VAP is burdened by increased morbidity and mortality, as well as the prolonged duration of mechanical ventilation and hospital length of stay which consequently increases hospital costs. Unfortunately, there are currently no accurate early diagnostic criteria for VAP because even the most widely used ones are neither sensitive (carrying the risk of antibiotic overuse) nor specific (carrying the risk of delayed antibiotic administration leading to increased morbidity and mortality).

With the raising number of applications of artificial intelligence (AI) and machine learning (ML) techniques in the ICU, several studies have already tried to apply different ML methods in order to predict whether a patient would develop VAP. These studies typically start from an underlying assumption that VAP could be detected based on the measured volatile organic compounds (VOC) found in breath and use electronic nose devices to collect those gases from the exhaled patient breath. However, VOCs in breath can be influenced also by many other diseases such as diabetes, acute renal failure, acute hepatitis. Additional problems with the current literature lie both in the type of the ML models used (so far, only the most traditional ones) as well as in the lack of an internationally recognized gold standard for the definitive diagnosis of VAP (literature refers to different “pseudo-reference standard” such as BAL, CPIS, etc.).

This project intends to advance the state of the art by designing and developing a machine learning model for VAP identification, providing novelty on three different levels:

  • with respect to data collection, data will be directly extracted from the mechanical ventilator during respiratory support of the patients, using a dedicated memory device (Memory Box, Hamilton Medical, Switzerland);

  • with respect to modelling, given that input would mostly consist of sequences of data recorded at specific time intervals, we will employ more sophisticated algorithms that pertain to the type of data at hand instead of just applying traditional classification models;

  • finally, for more reliable definitive diagnosis of VAP, additional staff-physicians with long-standing ICU-experience will be involved having on their disposal complete patient information combined of past medical history, physical examination, laboratory findings, chest radiography, LUS and BAL results.