Multiclass Classification of Alzheimer's Disease and Frontotemporal Dementia from Resting-State EEG Using Spectral Band-Power Features and Cost-Sensitive Gradient Boosting

The two most common forms of early onset dementia are frontotemporal dementia (FTD) and Alzheimer's disease (AD), and their overlapping behavioral and cognitive symptoms can sometimes make clinical distinction difficult. Resting-state electroencephalography (EEG) is a low-cost non-invasive signal that captures the diffuse cortical slowing associated with neurodegeneration and has therefore been proposed as a screening biomarker to subtype dementia.

This work presents a machine-learning pipeline based on relative spectral band-power features and a slowing-ratio score derived from 19-channel resting-state EEG to classify AD, FTD and healthy-control (HC) patients. The recordings were acquired from the publicly available OpenNeuro ds004504 dataset (88 subjects: 36 AD, 23 FTD, 29 HC). The data was segmented into 4-second epochs and band power in delta, theta, alpha and beta bands was extracted per channel using Welch's method. A 95-dimensional feature vector per epoch was used with a cost-sensitive XGBoost classifier, trained with ANOVA-F-selected features and subject-grouped 5-fold cross-validation.

The model achieved 69.32% subject-level accuracy with cross-validation and 66.67% accuracy on a strictly held-out set of 18 subjects never seen during training. Results show interpretable spectral features with gradient boosting can differentiate three overlapping clinical groups from EEG alone, while emphasizing the challenge of AD-FTD confusion.