Integrating wavelet signal processing and machine learning for EEG-based detection of brain diseases with representative channel selection

This is the authors' abstract. We don't add key points for this paper.

BackgroundBrain diseases pose a significant global health challenge, and their diagnosis often relies on subjective clinical assessments.ObjectiveTo this end, this study proposes an integrated wavelet signal processing and machine learning framework for single-channel Electroencephalography (EEG)-based detection of brain diseases with representative channel selection.MethodsFirst, EEG signals are decomposed into rhythm-specific sub-bands using the Discrete Wavelet Packet Transform (DWPT). Then, for each rhythm, a multi-dimensional feature matrix (MFM) is constructed by integrating Approximate Entropy (AE), Rhythm Energy Ratio (RER), C0-complexity (C0), and Zero-crossing Rate (ZCR), which comprehensively models the EEG characteristics. Six classifiers were compared under subject-level leave-one-subject-out cross-validation, and Decision Tree (DT) was selected as the representative classifier. Each channel was then evaluated independently as a complete single-channel system, and representative channels were identified from the completed results.ResultsValidated on publicly available datasets for SCHIZ (schizophrenia), UNM (Parkinson's disease), and MPHCE (depression), the method achieves impressive accuracies of 84.52% (F4), 86.54% (F4), and 93.10% (Fp1), respectively, using single-channel EEG electrode.ConclusionsSuch results demonstrate that the proposed approach maintains robust detection performance in each of the three disease-specific tasks. The framework is lightweight by design, using only a single EEG channel and a single simple machine-learning model for classification. It needs no multi-channel feature concatenation or decision-level fusion. This single-channel design suits hardware and data sources with limited resources, and may support the development of low-cost portable EEG systems.