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.

Abstract

Purpose: Achieving good performance on real biomerkcal data that frequently contains imbalance characteristics remains a challenging task. This study aimed to investigate the impact of re-sampling techniques in imbalanced datasets for PET radiomics-based prognostication in head and neck (HNC) cancer patients. Methods and materials: PET-based radiomics analysis was performed in 166 patients (median age 49.3 years. 77.1% male) diagnosed with nasopharyngeal carcinoma (NPC) in our centre and 182 HNC patients from the TCIA database. Conventional PET and 15 robust texture features were extracted for the correlation analysis of overall survival (OS) and disease progression-free survival (DES). We investigated cross-combination of 10 re-sampling methods and 4 classifiers for radiomics-based survival prediction. Results: Re-samplng techniques including oversampling and hybrid sampling achieved significant improvement on the area under the receiver operating characteristic curve (AUC) and prediction of the minority dass in terms of precisions, recalls, and F-measures in both cohorts, as compared to no oversampling (Wilcoxon signed rank-sum test, p < 0.05). We observed that the combination method ADASYN oversampling*XGboost classifier (AUC of 0.70. accuracy of 0.74. sensitivity of 0.75. and specificity of 0.70) presented the highest performance for DFS prediction. ADASYN+SVM performed best (AUC of 0.80, accuracy of 0.76. sensitivity of 0.74. and specificity of 0.86) for OS prediction in our NPC cohort. Conclusion: We identified optxnal machine learning methods for the prediction of prognostication in NPC. which enhanced the applications of radiomics in precision oncology and clinical practice. Of note, re-sampling techniques showed a significant positive impact on prediction performance in imbalanced datasets. Limitations: Our study is limited by retrospective datasets in terms of relatively small numbers of instances.