Independent component analysis of oddball EEG recordings to detect Parkinson's disease

Abstract

Parkinson's Disease (PD) is one of the most common diseases affecting the human brain, thus approaches are needed to help diagnose it. Since the changes caused by PD are visible in electroencephalograms (EEG), analysis of EEG represents one such approach. In this study, we used 25 EEG recordings of PD patients and 25 of healthy controls, subjected to auditory tasks, available in the Parkinson's Oddball database. The mean age of the PD patients was 69.7 years (std. 8.7) and 69.3 years (std. 9.6) of the control subjects. We employed the Independent Component Analysis (ICA) method to characterize the PD and control EEG recordings, to represent the changes in habituation as a response to different auditory events via the ICA components in the form of topological distributions, and to classify the EEG recordings of the two groups. Characterization of the frontal and central electrodes of the topological distribution showed high separation power to differentiate EEG recordings of the PD patients and healthy subjects. The average classification results using 5-fold cross-validation over 50 trials and the first four features ranked according to the variance of the ICA components, while the features were logarithm of the variance of the ICA components, yielded the following performances: classification accuracy of 88.56%, sensitivity of 89.36%, and specificity of 87.76%. The use of the ICA method appears to be a promising approach for characterizing and classifying auditory EEG recordings.

Keywords: Classification; Electroencephalogram; Independent component analysis; Parkinson’s disease.

Fig. 1

Topological distributions of the first 12 (top four rows) and the last 6 (bottom two rows) ICA components for the Target task for the PD (the first three columns of the topological plots) and the control groups (the last three columns). The value range is the same for all distributions and is shown on the right side. Legend: In the top left corner of each topological plot PX or CX indicate the component number (X) for the PD or control group. In the top right corner, V indicates the variance for the given component (i.e., the actual value of the variance is multiplied by 100), while the bracketed value below indicates the component rank according to the variance. In the bottom right corner, the bracketed value indicates the rank of the component according to the MAD value, while M indicates the MAD value between corresponding electrodes (i.e., the actual MAD value is multiplied by 10). In the bottom left corner, the d is the mean effect size for the ICA component for the two groups, while the bracketed value indicates the rank of the component according to the mean effect size. Figure was created using Matlab 2024a.

Fig. 2

Representation of the values used to rank the components for classification. In subplot (a), the variances for ICA components are shown, where components are ranked according to the diminishing value of variance, while in subplot (b), the MAD values for components are presented, where components are again ranked according to variance value. Figure was created using Matlab 2024a.

Fig. 3

Representation of EEG electrode placement used in the study. The region of the head with the most notable differences between the groups is shown as the dark gray area. The electrodes from which the signals used for feature extraction were obtained are shown with the magenta position markers. Figure was created using Matlab 2024a.

Fig. 4

Classification accuracies for the three auditory tasks using the increasing number of features. The order of the features was determined according to the decreasing value of the variance of the components (left column) and the decreasing value of the MAD (right column). SVM P3 support vector machine classifier with the polynomial kernel, SVM RBF support vector machine classifier with the radial basis function kernel, Tree Decision Tree classifier, Bayes Naive Bayes classifier, LDA linear discriminant analysis classifier, QDA quadratic discriminant analysis classifier. Figure was created using Matlab 2024a.