Machine Learning Model to Predict Outcome in Acute Hypoxemic Respiratory Failure

Overview

Acute hypoxemic respiratory failure (AHRF) is the most common cause of admission in the intensive care units (UCIs) worldwide. We will assess the value of machine learning (ML) techniques for early prediction of ICU death in 1,241 patients enrolled in the PANDORA (Prevalence AND Outcome of acute Respiratory fAilure) Study in Spain. The study was registered with ClinicalTrials.gov (NCT03145974). Our aim is to evaluate the minimum number of variables models using logistic regression and four supervised ML algorithms: Random Forest, Extreme Gradient Boosting, Support Vector Machine and Multilayer Perceptron.

Acute hypoxemic respiratory failure (AHRF) is the most common cause of admission in the intensive care units (UCIs) worldwide. We will assess the value of machine learning (ML) techniques for early prediction of ICU death in AHRF patients on mechanical ventilation (MV). Few studies have investigated the prediction of mortality in patients with AHRF. For model development, the investigators will extract data for the first 2 days after diagnosis of AHRF from patients included in the de-identified database of the PANDORA cohort. We had a database with 2,000,000 anonymized and dissociated demographics and clinical, data from 1,241 patients with AHRF enrolled in our PANDORA cohort (Prevalence AND Outcome of acute Respiratory fAilure) from 22 Spanish hospitals and coordinated by the principal investigator (JV). The investigators will follow the Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) guidelines for model prediction. We will screen collected variables employing a genetic algorithm variable selection method to achieve parsimony. We evaluated the minimum number of variables models using logistic regression and 4 supervised ML algorithms: Random Forest, Extreme Gradient Boosting, Support Vector Machine and Multilayer Perceptron. We will use a 5-fold cross-validation in the dataset of 1,000 patients selected randomly in training data (80%) and testing data (20%). For external validation, we will use the remaining 241 patients.

Conditions