Deep Learning-Based Alzheimer's Disease Detection from Multi-Channel EEG Using Fused Time-Frequency Image Grids

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Dementia diagnosis is challenging, requiring accurate and timely detection.
• Electroencephalography (EEG) provides a noninvasive method for assessing neurophysiological changes.
• Automated EEG analysis frameworks are crucial for improving dementia diagnosis.

Purpose of the Study:

• To systematically evaluate the impact of various time-frequency representations (TFRs) on dementia classification accuracy.
• To assess TFR performance within a unified multi-channel EEG image fusion framework.
• To compare different convolutional neural network (CNN) architectures for EEG-based dementia classification.

Main Methods:

• EEG data from 88 subjects (Alzheimer's disease, frontotemporal dementia, controls) were analyzed.
• Channel-wise EEG signals were transformed into time-frequency images using STFT, CWT, HHT, WVD, and CQT.
• Fused image data from 19 EEG channels were classified using MobileNetV2, ResNet-50, and InceptionV3.

Main Results:

• Classification performance varied significantly based on the TFR used.
• The STFT representation with InceptionV3 achieved the highest accuracy (98.8% random split, 84.3% subject-wise split).
• Constant-Q Transform (CQT) showed competitive results, while HHT and WVD were less effective.

Conclusions:

• The choice of TFR significantly impacts EEG-based dementia classification accuracy.
• Structured multi-channel fusion and systematic TFR evaluation are vital for robust diagnostic frameworks.
• The findings provide a foundation for developing interpretable and reliable EEG diagnostic tools for dementia.