Machine Learning Models for Prognostic Assessment of Covid-19 Mortality Using Computed Tomography-Based Radiomics

Nima  Yousefi; Vahid  Ghavami; Maryam  Salari; Saeed  Akhlaghi

doi:10.18502/jbe.v10i4.18525

Nima Yousefi Department of Biostatistics, School of Health, Mashhad University of Medical Sciences, Mashhad, Iran.
Vahid Ghavami Department of Biostatistics, School of Health, Mashhad University of Medical Sciences, Mashhad, Iran.
Maryam Salari Department of Biostatistics, School of Health, Mashhad University of Medical Sciences, Mashhad, Iran.
Saeed Akhlaghi Department of Biostatistics, School of Health, Mashhad University of Medical Sciences, Mashhad, Iran.

DOI: https://doi.org/10.18502/jbe.v10i4.18525

Keywords: Machine learning; Radiomics features; Computed tomography scan; Mortality; Covid-19

Abstract

Introduction: This study examines the significance of developing a predictive approach for assessing theprognosis of individuals diagnosed with COVID-19. This method can help physicians make treatment decisionsthat decrease mortality and prevent unnecessary treatments. This study also emphasizes the significance ofradiomics features. Therefore, our objective was to assess the predictive capabilities of Computed Tomography-based radiomics models using a dataset comprising 577 individuals diagnosed with COVID-19.

Methods: The U-net model was applied to automatically perform whole lung segmentations, extracting 107 texture, intensity, and morphological features. We utilized two feature selectors and three classifiers. We assessed the random forest, logistic regression, and support vector machines by implementing a five-fold cross-validation approach. Precision, sensitivity, specificity, accuracy, F1-score, and area under the receiver operating characteristic curve were reported.

Results: The random forest model achieved an area under the receiver operating characteristic curve, precision, sensitivity, specificity, accuracy, and F1-score in the range of 0.85 (CI 95%: 0.76–0.91), 0.75, 0.82, 0.78, 0.68, and 0.71, respectively. Logistic regression attained an area under the receiver operating characteristic curve of 0.80 (CI 95%: 0.72–0.88), corresponding to values of 0.88, 0.62, 0.74, 0.55, and 0.67, respectively. Support Vector Machines computed the above six metrics as an area under the receiver operating characteristic curve, sensitivity, specificity, accuracy, precision, and F1-score in the range of 0.69 (CI 95%: 0.59–0.79), 0.68, 0.64, 0.66, 0.5, and 0.57, respectively.

Conclusion: We are developing a robust radiomics classifier that predicts mortality in COVID-19 patients. Lung Computed Tomography radiomics features may aid in identifying high-risk individuals who need supplementary therapy and decrease the propagation of the virus