Prediction models for respiratory outcomes in patients with COVID-19: integration of quantitative computed tomography parameters, demographics, and laboratory features

Abstract

Background: We aimed to develop integrative machine-learning models using quantitative computed tomography (CT) parameters in addition to initial clinical features to predict the respiratory outcomes of coronavirus disease 2019 (COVID-19).

Methods: This was a retrospective study involving 387 patients with COVID-19. Demographic, initial laboratory, and quantitative CT findings were used to develop predictive models of respiratory outcomes. High-attenuation area (HAA) (%) and consolidation (%) were defined as quantified percentages of the area with Hounsfield units between -600 and -250 and between -100 and 0, respectively. Respiratory outcomes were defined as the development of pneumonia, hypoxia, or respiratory failure. Multivariable logistic regression and random forest models were developed for each respiratory outcome. The performance of the logistic regression model was evaluated using the area under the receiver operating characteristic curve (AUC). The accuracy of the developed models was validated by 10-fold cross-validation.

Results: A total of 195 (50.4%), 85 (22.0%), and 19 (4.9%) patients developed pneumonia, hypoxia, and respiratory failure, respectively. The mean patient age was 57.8 years, and 194 (50.1%) were female. In the multivariable analysis, vaccination status and levels of lactate dehydrogenase, C-reactive protein (CRP), and fibrinogen were independent predictors of pneumonia. The presence of hypertension, levels of lactate dehydrogenase and CRP, HAA (%), and consolidation (%) were selected as independent variables to predict hypoxia. For respiratory failure, the presence of diabetes, levels of aspartate aminotransferase, and CRP, and HAA (%) were selected. The AUCs of the prediction models for pneumonia, hypoxia, and respiratory failure were 0.904, 0.890, and 0.969, respectively. Using the feature selection in the random forest model, HAA (%) was ranked as one of the top 10 features predicting pneumonia and hypoxia and was first place for respiratory failure. The accuracies of the cross-validation of the random forest models using the top 10 features for pneumonia, hypoxia, and respiratory failure were 0.872, 0.878, and 0.945, respectively.

Conclusions: Our prediction models that incorporated quantitative CT parameters into clinical and laboratory variables showed good performance with high accuracy.

Keywords: Coronavirus disease 2019 (COVID-19); machine-learning; prediction model; quantitative computed tomography (quantitative CT); respiratory failure.

Figure 1

Density plots of (A-D) HAA (%) and (E-H) LAA (%) according to the severity of COVID-19 pneumonia. HAA showed a tendency to increase in the order of pneumonia, hypoxia, and respiratory failure, whereas LAA did not show the same trend. HAA, high-attenuation area; LAA, low-attenuation area; COVID-19, coronavirus disease 2019; resp, respiratory.

Figure 2

Receiver operating characteristic curve of the logistic regression model for respiratory outcomes. The AUCs of the ROC curves for pneumonia, hypoxia, and respiratory failure were 0.904, 0.890, and 0.969, respectively. AUC, area under the curve; resp, respiratory; ROC, receiver operating characteristic.

Figure 3

Variable importance based on random forest models of (A) pneumonia, (B) hypoxia, and (C) respiratory failure. Variables are shown in the order of importance based on the mean decrease of the Gini index. CRP, C-reactive protein; PLT, platelet; LDH, lactate dehydrogenase; HAA, high attenuation area; WBC, white blood cell; ALT, alanine aminotransferase; AST, aspartate aminotransferase; LAA, low attenuation area; BMI, body mass index; BUN, blood urea nitrogen; Hb, hemoglobin; CVD, cardiovascular disease; CKD, chronic kidney disease.