Reliability of Machine Learning in Eliminating Data Redundancy of Radiomics and Reflecting Pathophysiology in COVID-19 Pneumonia: Impact of CT Reconstruction Kernels on Accuracy

Abstract

Background: Radiomical data are redundant but they might serve as a tool for lung quantitative assessment reflecting disease severity and actual physiological status of COVID-19 patients. Objective: Test the effectiveness of machine learning in eliminating data redundancy of radiomics and reflecting pathophysiologic changes in patients with COVID-19 pneumonia. Methods: We analyzed 605 cases admitted to Al Ain Hospital from 24 February to 1 July, 2020. They met the following inclusion criteria: age≥18 years; inpatient admission; PCR positive for SARS-CoV-2; lung CT available at PACS. We categorized cases into 4 classes: mild ≤25% of pulmonary parenchymal involvement, moderate - 25-50%, severe - 50-75%, and critical – over 75%. We used CT scans to build regression models predicting the oxygenation level, respiratory and cardiovascular functioning. Results: Radiomical findings are a reliable source of information to assess the functional status of patients with COVID-19. Machine learning models can predict the oxygenation level, respiratory and cardiovascular functioning from a set of demographics and radiomics data regardless of the settings of reconstruction kernels. The regression models can be used for scoring lung impairment and comparing disease severity in follow up studies. The most accurate prediction we achieved was 6.454±3.715% of mean absolute error/range for all the features and 7.069±4.17% for radiomics. Conclusion: The models may contribute to the proper risk evaluation and disease management especially when the oxygen therapy impacts the actual values of the functional findings. Still, the structural assessment of an acute lung injury reflects the severity of the disease.