Benefits of Deep Learning Chest X-ray Features in the Mortality Prediction of COVID-19-associated ARDS Patients

Abstract

Acute respiratory distress syndrome associated with COVID-19 (ARDS-COV19) is a severe pulmonary syndrome leading to acute respiratory failure. ARDS is complex and heterogeneous, with patients frequently needing invasive mechanical ventilation (IMV) in intensive care units (ICUs). The identification of risk groups is crucial for precision medicine, although the lack of diagnostic methods can be limiting. Chest radiography is a qualitative and accessible imaging examination routinely used in ICU settings. The development of a multivariate and quantitative classifier based on radiomics is essential for predicting the mortality of patients under IMV and to evaluate the possible benefits of imaging features in the risk stratifications of this patients. For this purpose, 87 ARDS-COV19 patients from an ICU, with an average age of 64 ±11,6 years, of whom 62.07% were male, were included. The mortality rate was 48.28%. Chest X-rays from the 1st and 3rd days of IMV were collected, pre-processed, and concatenated. Deep learning features were then extracted using a pre-trained convolutional neural network (CheXnet). These features were combined with clinical variables (CV) to build four machine learning models based on logistic regression (LogReg) and multilayer perceptron (MLP) classifiers. Age, the PaO 2 /FiO 2 ratio on the 3rd day of IMV, and a deep learning image feature (DLF) were selected by filter techniques to be used in two of the models. The remaining models were created only using the CV variables (Age, PaO2/FiO2). Using 10-fold cross validation the models which included the DLF_258 showed 89% (LogReg) and 82% (MLP) probability of having better accuracy than CV models with a 77% (MLP) and 86% (LogReg) chance of improved F1-Score. In the LogReg based models, AUC had a 74% chance of being improved with the added image features while on the MLP models only a 55% chance was noted. In cross-validation the combined LogReg models provided a mean AUC of 0.77, 95% CI [0.67, 0.85], an accuracy of 0.75, 95% CI [0.67, 0.85], and an F1 score of 0.73, 95% CI [0.62, 0.82], while the MLP models provided an AUC of 0.78, 95% CI [0.68, 0.86], an accuracy of 0.77, 95% CI [0.67, 0.85], and an F1 score of 0.76, 95% CI [0.67, 0.84]. Despite promising results, the sample size was limited, and external testing is lacking. Therefore, data collection and subsequent validation are essential.