Machine learning re-sampling techniques in imbalanced datasets improve prognostication performance in a multicentre cohort of head and neck cancer patients using a PET-based radiomics model.