Background/Objectives: Dementia is a progressive neurodegenerative disorder for which accurate and timely diagnosis remains a major clinical challenge. Electroencephalography (EEG) offers a noninvasive and cost-effective means of capturing neurophysiological alterations, motivating the development of reliable EEG-based automated diagnostic frameworks. This study aims to systematically examine how different time–frequency representations (TFRs) affect dementia classification performance within a unified multi-channel EEG image fusion framework. Methods: Resting-state, eyes-closed EEG recordings from 88 subjects, including Alzheimer’s disease, frontotemporal dementia, and cognitively normal controls, were preprocessed and segmented. Channel-wise signals were converted into two-dimensional time–frequency images using Short-Time Fourier Transform (STFT), Continuous Wavelet Transform (CWT), Hilbert–Huang Transform (HHT), Wigner–Ville Distribution (WVD), or Constant-Q Transform (CQT). Images from 19 EEG channels were fused into a structured grid and classified using pretrained convolutional neural networks, including MobileNetV2, ResNet-50, and InceptionV3. Results: Results indicate that classification performance is highly dependent on the chosen TFR. The STFT-based representation combined with InceptionV3 achieved the highest accuracy, reaching 98.8% with random splitting and 84.3% with subject-wise splitting, outperforming previous studies. CQT also showed competitive performance, whereas HHT and WVD were less effective. Gradient-weighted class activation mapping provided interpretable visualization of physiologically relevant EEG channel contributions. Conclusions: The proposed framework demonstrates the importance of structured multi-channel fusion and systematic TFR evaluation for robust and interpretable EEG-based dementia classification and serves as a foundation for future cross-dataset validation.
Yildiz et al. (Mon,) studied this question.