Predicting the Use of Invasive Mechanical Ventilation in ICU COVID-19 Patients

Abstract

Critically ill patients often need Invasive Mechanical Ventilation (IMV) when treated at intensive care units (ICU). However, it is a complex treatment that most medical doctors avoid when possible. This technique demands appropriate equipment such as ventilators and specialized personal to operate it. Patients with Coronavirus Disease (COVID-19) may need IMV, usually for an extensive period. Due to the pandemic, IMV resources became scarce, and the decision to institute mechanical ventilation based on medical judgement should be avoided unless it is absolutely necessary. This study proposes the use of clinical and laboratory data from the 24 h preceding and succeeding the ICU admission and Machine Learning classifiers such as Random Forest (RF) to predict the probability of a patient requiring IMV. The proposed methodology is split into pre-processing, modelling, and feature selection. A wide range of different classifiers with a diverse set of variables were tested. The final model is an RF model with sixteen features and a $91.88\%$ out of sample accuracy. It can predict if a patient needs IMV, and produce an explanation for the model using Local Interpretable Model-agnostic Explanation in seven seconds. We believe this to be an advantageous tool for supporting clinical decisions, minimize ventilator-associated complications and optimize resources allocation.