Development and multicentric external validation of a prognostic COVID-19 severity model based on thoracic CT

Table 4 Mean receiver operating characteristic (ROC) area under the curve (AUC) ± standard deviation (SD) from the four-fold cross-validation for the most relevant models that were tested

Model	Mean AUC ± SD
Mean Teacher ConvNeXt	0.79 ± 0.033
Logistic regression with features added to the baseline model:
None	0.65 ± 0.031
Number of lesions	0.66 ± 0.037
Volume fractions	0.74 ± 0.045
Volume fraction per lung lobe	0.71 ± 0.039
Mean intensity, kurtosis and skewness	0.72 ± 0.033
Volume fractions, mean intensity, kurtosis and skewness	0.74 ± 0.033
Radiomic features from Chen et al. [38]	0.69 ± 0.025
Radiomic features from Huang et al. [39]	0.72 ± 0.037
Volume fractions, radiomic features from Huang et al. [39]	0.73 ± 0.040
Best three radiomic features from univariate selection	0.74 ± 0.037
Volume fractions, best three radiomic features from univariate selection	0.74 ± 0.038
Best three radiomic features from multivariate selection	0.71 ± 0.032
Volume fractions, best three radiomic features from multivariate selection	0.74 ± 0.029

For the logistic regression, the baseline model considers the patient age and sex. The other regression models build on the baseline through the addition of different features. The number of lesions and the volume fractions include the respective values for ground glass opacity (GGO) and consolidation separately. Mean intensity, kurtosis and skewness include the values for healthy lung parenchyma, GGO and consolidation separately. Best three radiomic features from univariate feature selection: lbp-3D-k_glszm_ZoneVariance, original_shape_Maximum2DDiameterColumn, lbp-3D-m1_glrlm_LongRunLowGrayLevelEmphasis. Best three radiomic features from multivariate feature selection: wavelet-HLL_glcm_MaximumProbability, wavelet-LLL_glrlm_HighGrayLevelRunEmphasis, wavelet-LHL_glcm_Correlation

ISSN: 1472-6947