Glioma grade discrimination based on class imbalance learning with terahertz spectral data

Grading gliomas is crucial for prognosis and survival prediction. Clinically, gliomas are categorized into low-grade gliomas (LGG, WHO II) and high-grade gliomas (HGG, WHO III and IV). Compared to traditional diagnostic methods, terahertz spectroscopy offers advantages such as time efficiency and label-free detection. In this study, we propose a machine learning model for discriminating glioma grades based on terahertz spectral data, incorporating, for the first time, the class imbalance of terahertz spectral data from different glioma grades. A total of 420 sample data from 21 cases were included. Three data processing methods—ROS, RUS, and SMOTE—were utilized, and three classifiers were employed to construct the model. We compared the effects of the three data processing methods and validated the proposed method on a test set. The best performance of the proposed algorithm, evaluated by the area under the curve (AUC), achieved a maximum value of 88.8%. This represents a significant advancement in the development of a rapid glioma grade diagnostic tool and provides more effective guidance for developing surgical protocols for gliomas.