Feature selection strongly affected radiomics prediction of xerostomia after head-and-neck RT

KEY POINTS

• The multicentre cohort included 250 head-and-neck cancer patients from three institutions. Centers 1 and 2 contributed 178 patients for model development, while 72 patients from Center 3 were completely withheld for independent institutional testing.
• The models combined 321 radiomic features from CT, T1-weighted MRI and T2-weighted MRI with 20 clinical and dosimetric variables. Nine feature-reduction strategies and five machine-learning classifiers were evaluated.
• Toxicity was patient reported using EORTC QLQ-H&N35. Moderate-to-severe early xerostomia occurred in 71%, late xerostomia in 44%, early sticky saliva in 55% and late sticky saliva in 16%.
• Feature-selection strategy had a large effect on discrimination. For early xerostomia, R-MRmr with T1 MRI achieved AUC 0.96, with accuracy 0.90, sensitivity 0.87 and specificity 0.93 in the independent-center evaluation.
• Early sticky saliva was also predicted best using R-MRmr with T1 MRI, reaching AUC 0.85. Relief performed best for late xerostomia, with AUC 0.88 using T1 MRI.
• Late sticky saliva was considerably harder to predict. The best configuration, Relief using combined CT, T1 and T2 MRI, achieved only AUC 0.65, despite class-balancing methods.
• More imaging was not necessarily better. T1-weighted MRI repeatedly outperformed CT and T2 MRI, while combining CT + T1 + T2 did not consistently improve prediction, potentially because the larger feature space increased redundancy and model complexity.
• External testing was limited to a single 72-patient institution, calibration and clinical net benefit were not assessed, feature stability was not formally quantified and numerous exploratory comparisons were performed without multiplicity correction.

CLINICAL TAKEAWAY

SOURCE

Cancers