Research on the classification of EEG signals for dementia and its interpretability using the GWOCS agorithm

Alzheimer's disease (AD) and frontotemporal dementia (FTD) have insidious, similar and ambiguous clinical symptoms, which make their diagnosis difficult. Currently, in the field of EEG signal analysis, there are relatively few studies on the interpretability analysis of feature selection using intelligent optimization algorithms. To analyze the EEG signals of AD and FTD patients more comprehensively, first, 16 features in three dimensions of entropy, time-frequency domain and SODP were extracted in this paper. Secondly, Pearson correlation analysis, importance ranking and SHAP interpretability analysis methods were adopted to select SE, SW, ZCR, STA, CTM2 and CTM5 as the best discriminative features, and the Relief algorithm was used for fusion and dimension reduction based on weights. Thirdly, GWOCS was used for channel screening to determine the optimal channel combination of Fz, F7, Fp1, Fp2, F3, T3, P4 and C3, achieving the three-classification identification of the two patient groups and the normal control group, with the classification accuracy reaching 89.35[Formula: see text] and 81.12[Formula: see text] in cross-validation and LOSO validation, respectively. Finally, the SHAP method was used to prove that for the diagnosis of dementia, the prefrontal and temporal lobe brain regions play a decisive role, verifying the effectiveness of this framework in rapid channel selection and improving the efficiency of disease detection.