Balancing Spectral, Temporal and Spatial Information for EEG-based Alzheimer's Disease Classification

TL;DR

The paper addresses cost-effective Alzheimer's disease screening using EEG by asking how spatial information contributes relative to spectral and temporal cues for classification. It introduces a resolution-based feature extraction framework that independently tunes spectral, temporal, and spatial dimensions while keeping the total feature count fixed, with spatial pooling driven by data-defined graph clustering. Using two EEG datasets and a support vector machine classifier, it shows that spatial information is at least as informative as spectral information and more robust than temporal information, including a +1.1% accuracy gain when spatial features replace spectral ones on the larger dataset. The findings highlight the importance of spatial information for EEG-based AD classification and suggest broader applicability to multivariate signal classification.

Abstract

The prospect of future treatment warrants the development of cost-effective screening for Alzheimer's disease (AD). A promising candidate in this regard is electroencephalography (EEG), as it is one of the most economic imaging modalities. Recent efforts in EEG analysis have shifted towards leveraging spatial information, employing novel frameworks such as graph signal processing or graph neural networks. Here, we investigate the importance of spatial information relative to spectral or temporal information by varying the proportion of each dimension for AD classification. To do so, we systematically test various dimension resolution configurations on two routine EEG datasets. Our findings show that spatial information is more important than temporal information and equally valuable as spectral information. On the larger second dataset, substituting spectral with spatial information even led to an increase of 1.1% in accuracy, which emphasises the importance of spatial information for EEG-based AD classification. We argue that our resolution-based feature extraction has the potential to improve AD classification specifically, and multivariate signal classification generally.

Figures (3)

• Figure 1: Illustration of balancing spectral, temporal and spatial features for classification. (A) Firstly, spectral features are extracted as the spectral density. Note that this step already compresses the temporal dimension. In a second step, the features are pooled along the temporal dimension by averaging. The extent of the pooling determines the temporal resolution. Lastly, the features are pooled spatially using graph pooling, which determines the spatial resolution, before being fed into a support vector machine. (B) When the features are extracted, the resolution along any of the three domains can be varied, changing the shape of the feature matrix as illustrated by the cuboids on the triangle. Importantly, the volume of the displayed cuboids, representing the number of total features, remains constant.
• Figure 2: Linearly interpolated accuracy in dependence on the feature resolution configuration for dataset I (A) and II (B). The green crosses mark the experimentally tested configurations. The accuracy along the triangle edges is separately plotted in Figure \ref{['fig:results_edge']}. Both datasets reveal similar levels of accuracy for maximal spectral and graph features, as well as poor accuracy for maximal time features. They further both reveal an accuracy level between the maximal spectral and graph features. In the centre of the available configuration space, corresponding to more balanced resolution configurations, the performance diverges across the two datasets.
• Figure 3: Accuracy along the triangle edges depicted in Figure \ref{['fig:results_triangle']}. The colour of the curve allows to retrieve the feature resolution configuration from Figure \ref{['fig:results_triangle']}. The curve has a large accuracy valley at maximal temporal information (yellow section), but also a smaller accuracy valley between the maximal spatial and spectral configuration (magenta section).