Predicting Mechanical Ventilation and Mortality in COVID-19 Using Radiomics and Deep Learning on Chest Radiographs: A...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Coronavirus disease 2019 (COVID-19) poses significant challenges in predicting patient outcomes.
• Mechanical ventilation and mortality are critical indicators of severe COVID-19.
• Accurate prognostic tools are needed for effective patient management and resource allocation.

Purpose of the Study:

• To develop and evaluate computational models for predicting mechanical ventilation requirement and mortality in COVID-19 patients using chest radiographs (CXRs).
• To compare the performance of machine learning classifiers with radiologist interpretations.
• To investigate the added value of radiomic features in deep learning models.

Main Methods:

• Retrospective analysis of 530 CXRs from 515 COVID-19 patients across two centers.
• Extraction of radiomic features from CXRs for training machine learning classifiers (LDA, QDA, RF).
• Exploration of deep learning (DL) approaches and a novel radiomic embedding framework.
• Comparison of model performance against expert radiologist grading.

Main Results:

• Radiomic classification models achieved comparable or superior performance to expert radiologist grading for predicting mechanical ventilation and mortality.
• Combined models using both expert scores and radiomic features demonstrated improved predictive accuracy (mAUCs of 0.79 and 0.83).
• Inclusion of radiomic features enhanced deep learning model predictions.

Conclusions:

• Computational models utilizing radiomic features from CXRs show promise in predicting COVID-19 patient outcomes.
• The integration of radiomic features with expert interpretation can enhance prognostic accuracy.
• These AI-driven tools may assist clinicians in decision-making and resource allocation during the pandemic.