Deep learning of chest X-rays can predict mechanical ventilation outcome in ICU-admitted COVID-19 patients

Abstract

The COVID-19 pandemic repeatedly overwhelms healthcare systems capacity and forced the development and implementation of triage guidelines in ICU for scarce resources (e.g. mechanical ventilation). These guidelines were often based on known risk factors for COVID-19. It is proposed that image data, specifically bedside computed X-ray (CXR), provide additional predictive information on mortality following mechanical ventilation that can be incorporated in the guidelines. Deep transfer learning was used to extract convolutional features from a systematically collected, multi-institutional dataset of COVID-19 ICU patients. A model predicting outcome of mechanical ventilation (remission or mortality) was trained on the extracted features and compared to a model based on known, aggregated risk factors. The model reached a 0.702 area under the curve (95% CI 0.707-0.694) at predicting mechanical ventilation outcome from pre-intubation CXRs, higher than the risk factor model. Combining imaging data and risk factors increased model performance to 0.743 AUC (95% CI 0.746-0.732). Additionally, a post-hoc analysis showed an increase performance on high-quality than low-quality CXRs, suggesting that using only high-quality images would result in an even stronger model.

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

“…The authors' most effective model involved a combination of a risk score associated with clinical variables and two CXR DLF. This model demonstrated an AUC of 0.74 95% CI[0.73, 0.75], a classification accuracy of 0.76, specificity of 0.83, and recall of 0.49 (the authors did not provide a 95% CI for recall), while the logistic regression model only using clinical variables presented an AUC of 0.620 [22]. J.…”

“…Studies focused on the categorization or prediction of ARDS severity and mortality use different angles, from biomarkers to lung physiology and chest imaging [12], [21]. However, the few studies combine clinical data and portable CXR, mostly focus on mortality prediction of COVID-19 ICU patients, not considering the ARDS definition in the selected population [22]- [24]. These studies show improved model performance metrics when using CXR image features, although they rarely provide the probability of this finding being true considering model generalization to different training groups, validation groups and sample sizes.…”

Cascaded Deep Learning Model for Detecting Lung Infections Using Chest X-Rays

“…Alharbi et al [11] explored chest X-ray analysis using deep learning for omicron detection, achieving 99% accuracy with their CNN model, suggesting its potential as a cost-effective and efficient diagnostic method. Gourdeau et al [12] utilized bedside CXR images with deep transfer learning to predict mortality after mechanical ventilation in omicron ICU patients, finding that incorporating CXR images improves predictive accuracy for ICU triage. Elsharkawy et al [13] developed a CAD system using chest X-ray data to assess mortality risk in omicron patients, achieving high accuracy and suggesting its usefulness in evaluating disease severity and informing patient care.…”

Improving Intensive Care Chest X-Ray Classification by Transfer Learning and Automatic Label Generation

“…The potential of deep learning models for the evaluation of ICU chest Xrays was investigated in [6,7], but no automatically annotated labels and their influence on performance were considered. Few studies investigate the impact of transfer learning and automatically generated labels on the analysis of ICU chest X-rays, taking into account not only disease detection but also extrathoracic material location assessment.…”