Machine Learning Approaches for the Early Identification of Malignant Brain Tumors Based on EEG Signal Features

Introduction: Detecting malignant brain tumors at early stages, when lesions are more likely localized, is critical for enhancing surgical resection success and improving patient prognosis. Traditional detection methods primarily rely on costly imaging techniques, which often present issues related to cost-effectiveness and potential complications. This study aims to address these limitations by evaluating the performance of electroencephalography (EEG) features in detecting gliomas across various conditions.

Methods: We recruited patients with motor area (MA) and non-motor area (NMA) gliomas, as well as healthy controls (HC), to undergo EEG recordings under conditions that included resting state, motor execution (ME), and imagery tasks (MI). Multiple machine learning algorithms were employed to classify EEG signals across various groups and conditions. The process involves extracting time-domain and time-frequency features from EEG signals and evaluating the contribution and spatial distribution of these features to classification using SHapley Additive exPlanations (SHAP). Additionally, we developed an online prediction tool for classification of individual EEG signals.

Results: The results demonstrate that the SVM-RBF algorithm outperformed LDA across both subject and trial levels, particularly in distinguishing HC from NMA. Time-domain features, including standard deviation, peak-to-peak, and root mean square (RMS), were identified as the three most significant features, reflecting variations in spatial distribution across groups. Specifically, significant electrode regions included FC1, Cz, and C3 for HC versus MA; Cz, FC1, FC2, CP1, and CP2 for HC versus NMA; F3, Fz, and CP1 for NMA versus MA. Moreover, the online tool developed based on these findings effectively predicts individual EEG signals.

Conclusion : This work underscores the value of EEG as an accessible, cost-effective alternative for measuring biological signals, highlighting its potential for effective glioma screening and contributing to enhanced clinical practices.