A team at Prince of Songkla University demonstrates that a convolutional neural network trained on dynamic EEG connectivity features can classify Alzheimer’s disease, frontotemporal dementia, and healthy controls with 93.5% accuracy. The model transforms EEG recordings into statistical maps—mean, variance, skewness, and entropy across frequency bands—and leverages these patterns to distinguish dementia subtypes, offering a non-invasive, cost-effective diagnostic tool.

Key points

• Dynamic features—mean, variance, skewness, and Shannon entropy—are extracted from EEG connectivity measures (ISPC, wPLI, AEC) across delta to gamma bands.
• Statistical connectivity profiles are encoded as 4×19×19 feature maps and used to train a custom CNN with three convolutional stacks and global average pooling.
• The model achieves 93.5% multiclass accuracy, 97.8% accuracy for Alzheimer’s vs. controls, and 97.4% accuracy for Alzheimer’s vs. frontotemporal dementia classification.

Why it matters: This approach could transform dementia screening by offering rapid, non-invasive, and highly accurate differentiation of Alzheimer’s and frontotemporal subtypes using portable EEG.