Effect of machine learning re-sampling techniques for imbalanced datasets in ${}^{18}$F-FDG PET-based radiomics model on prognostication performance in cohorts of head and neck cancer patients

Biomedical data frequently contain imbalance characteristics which make achieving good predictive performance with data-driven machine learning approaches a challenging task. In this study, we investigated the impact of re-sampling techniques for imbalanced datasets in PET radiomics-based prognostication model in head and neck (HNC) cancer patients. Radiomics analysis was performed in two cohorts of patients, including 166 patients newly diagnosed with nasopharyngeal carcinoma (NPC) in our centre and 182 HNC patients from open database. Conventional PET parameters and robust radiomics features were extracted for correlation analysis of the overall survival (OS) and disease progression-free survival (DFS). We investigated a cross-combination of 10 re-sampling methods (oversampling, undersampling, and hybrid sampling) with 4 machine learning classifiers for survival prediction. Diagnostic performance was assessed in hold-out test sets. Statistical differences were analysed using Monte Carlo cross-validations by post hoc Nemenyi analysis. Oversampling techniques like ADASYN and SMOTE could improve prediction performance in terms of G-mean and F-measures in minority class, without significant loss of F-measures in majority class. We identified optimal PET radiomics-based prediction model of OS (AUC of 0.82, G-mean of 0.77) for our NPC cohort. Similar findings that oversampling techniques improved the prediction performance were seen when this was tested on an external dataset indicating generalisability. Our study showed a significant positive impact on the prediction performance in imbalanced datasets by applying re-sampling techniques. We have created an open-source solution for automated calculations and comparisons of multiple re-sampling techniques and machine learning classifiers for easy replication in future studies.