EEG-based classification framework for the detection of Alzheimer’s disease and mild cognitive impairment

Dementia has no cure, but if diagnosed in early stages, its progress can be slowed. For this reason, it is necessary to implement and improve the techniques that aid the diagnosis of this disease. Electroencephalography has been shown to be a potential candidate to support the diagnosis of dementia. Applying correct processing to the EEG signal and a good selection of features will allow a more accurate diagnosis. This paper presents a methodology to discriminate healthy subjects from subjects with Alzheimer’s disease and mild cognitive impairment. The work focuses mainly on the pre-processing stage and feature selection to establish a robust methodology that addresses inter-subject variability. In the pre-processing, a spatial filter is applied conditionally based on a threshold derived from the Signal-to-Noise ratio of the EEG signals; additionally, independent component analysis is used to remove noise present in the signal. For feature extraction, different techniques widely used in the frequency and time domains such as relative power and entropy were employed, obtaining a total of 133 features. Feature selection is performed through particle swarm optimization, where the objective function is based on the distance between the correlation matrix of the two classes considered; out of the 133 extracted features, 26 were selected, with relative power and entropy in the frontal and parietal electrodes being the most relevant in the detection of Alzheimer’s disease. The results obtained demonstrate that the methodology is successfully applied to different datasets achieving an accuracy greater than 95% in most of the tests carried out.