Determining HPV Status in Patients with Oropharyngeal Cancer from 3D CT Images Using Radiomics: Effect of Sampling Methods

Non-invasive detection of human papillomavirus (HPV) status is important for the treatment planning of patients with oropharyngeal cancer (OPC). In this work, three-dimensional (3D) head and neck computed tomography (CT) scans are utilized to identify HPV infection status in patients with OPC by applying radiomics and several resampling methods to handle highly imbalanced data. 1142 radiomic features were obtained from the segmented CT images of 238 patients. The features used were selected through correlation coefficient analysis, feature importance analysis, and backward elimination. The fifty most important features were chosen. Six different sampling methods, which are Synthetic Minority Oversampling Technique (SMOTE), Support Vector Machine Synthetic Minority Oversampling Technique (SVMSMOTE), Adaptive Synthetic Sampling Method (ADASYN), NearMiss, Condensed Nearest Neighbors (CNN), and Tomek’s Link, were performed on the training set for each of the positive and negative HPV classes. Two different machine learning (ML) algorithms, a Light Gradient Boosting Machine (LightGBM) and Extreme Gradient Boosting (XGBoost), were applied as predictive classification models. Model performances were assessed separately on 20% of the data. Oversampling methods displayed better performance than undersampling methods. The best performance was seen in the combination of SMOTE and XGBoost algorithms, which had an area under the curve (AUC) of 0.93 (95% CI: 82–99) and an accuracy of 90% (95% CI: 78–96). Our work demonstrated a reasonable accuracy in the forecast of HPV status using 3D imbalanced and small datasets. Further work is needed to test the algorithms on larger, balanced, and multi-institutional data.