Table 2.
Covariates used in the ensemble model.
| Name | Count | Covariates |
|---|---|---|
| Clinical | 7 | Age, HPV status (positive | negative | unknown), Smoking Status (never | former | current), T.category ([T1-T2],[T3-T4]), N.category ([N0-N1],[N2-N3]), Therapeutic Combination (RT alone, Concurrent Chemotherapy (CC), Induction + RT, Induction + CC), AJCC Stage (8th edition) |
| RSF (OS) | Up to 10 | F4.GrayLevelRunLengthMatrix25..90ShortRunLowGrayLevelEmpha,F48.GrayLevelCooccurenceMatrix25180.2ClusterProminence,F48.GrayLevelCooccurenceMatrix25270.1Contrast,F48.GrayLevelCooccurenceMatrix25225.7ClusterShade,F29.IntensityDirectLocalRangeMax,F2.GrayLevelCooccurenceMatrix25270.1Contrast,F2.GrayLevelCooccurenceMatrix25.333.4Correlation,F2.GrayLevelCooccurenceMatrix25180.6MaxProbability,F4.GrayLevelRunLengthMatrix25..90RunLengthNonuniformity,F4.GrayLevelRunLengthMatrix25.333ShortRunEmphasis |
| RSF (RFS) | Up to 10 | F48.GrayLevelCooccurenceMatrix25180.2ClusterProminence,F48.GrayLevelCooccurenceMatrix25315.6ClusterProminence,F8.IntensityDirectKurtosis, F9.IntensityDirectSkewness,F11.IntensityDirectKurtosis, F13.IntensityDirectEnergy,F48.GrayLevelCooccurenceMatrix25180.1InverseDiffNorm,F2.GrayLevelCooccurenceMatrix25180.5ClusterProminence,F2.GrayLevelCooccurenceMatrix25180.5ClusterShade,F14.IntensityDirectEnergy |
| COX (OS) | 5 | F25.ShapeVolume, F29.IntensityDirectLocalRangeMax,F4.GrayLevelRunLengthMatrix25..90RunLengthNonuniformity,F6.IntensityDirectSkewness,F48.GrayLevelCooccurenceMatrix25225.7AutoCorrelation |
| COX (RFS) | 8 | F5.IntensityDirectGlobalMax, F13.IntensityDirectGlobalMax,F14.IntensityDirectGlobalMax, F25.ShapeVolume,F29.IntensityDirectLocalRangeMax,F4.GrayLevelRunLengthMatrix25..90RunLengthNonuniformity,F4.GrayLevelRunLengthMatrix25..90ShortRunLowGrayLevelEmpha,F48.GrayLevelCooccurenceMatrix25225.7AutoCorrelation |
| Cluster | 1 | Cluster label with 2, 3, or 4 values |
The clinical covariates are used independently of the outcome being evaluated. Since Random Survival Forests (RSF) and Coxnet (COX) can be used as supervised feature selection methods, the radiomic features selected depend on the outcome used. The top covariates from RSF are selected for each outcome. For COX, the features selected depend on the number of non-zero weights learned by the regularization coefficient. COX selected 5 and 8 radiomics features for OS and RFS, respectively. Cluster refers to the cluster label extracted using Random Survival Forest Clustering.