Parkinson's disease detection based on multi-pattern analysis and multi-scale convolutional neural networks

Abstract

Parkinson's disease (PD) is a complex neurodegenerative disease. At present, the early diagnosis of PD is still extremely challenging, and there is still a lack of consensus on the brain characterization of PD, and a more efficient and robust PD detection method is urgently needed. In order to further explore the features of PD based on brain activity and achieve effective detection of PD patients (including OFF and ON medications), in this study, a multi-pattern analysis based on brain activation and brain functional connectivity was performed on the brain functional activity of PD patients, and a novel PD detection model based on multi-scale convolutional neural network (MCNN) was proposed. Based on the analysis of power spectral density (PSD) and phase-locked value (PLV) features of multiple frequency bands of two independent resting-state electroencephalography (EEG) datasets, we found that there were significant differences in PSD and PLV between HCs and PD patients (including OFF and ON medications), especially in the β and γ bands, which were very effective for PD detection. Moreover, the combined use of brain activation represented by PSD and functional connectivity patterns represented by PLV can effectively improve the performance of PD detection. Furthermore, our proposed MCNN model shows great potential for automatic PD detection, with cross-validation accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve all above 99%. Our study may help to further understand the characteristics of PD and provide new ideas for future PD diagnosis based on spontaneous EEG activity.

Keywords: EEG; Parkinson’s disease; disease detection; multi-pattern analysis; multi-scale convolutional neural networks.

FIGURE 1

Framework diagram of multi-scale convolutional neural networks (MCNN).

FIGURE 2

The contrast maps of group-averaged power spectral density (PSD) in five frequency bands (i.e., δ, θ, α, β, and γ bands) between the Parkinson’s disease (PD) patient group (including PD_OFF and PD_ON) and the HC group for 2 different datasets. The white areas (values less than 0 in the colorbar) indicate that the difference between the two groups is not statistically significant (i.e., P > 0.05). The yellow-red (values greater than 1.0 in the colorbar) area represents the ratio’s numerator with a larger PSD value than the denominator, while the cyan-blue (values greater than 0 and less than 1.0 in the colorbar) are the opposite. The darker the color, the greater the difference in PSD between the two groups.

FIGURE 3

The contrast maps of group-averaged phase-locked value (PLV) in five frequency bands (i.e., δ, θ, α, β, and γ bands) between the PD patient group (including PD_OFF and PD_ON) and the HC group for 2 different datasets. The white areas (values less than 0 in the colorbar) indicate that the difference between the two groups is not statistically significant (i.e., P > 0.05). The yellow-red (values greater than 1.0 in the colorbar) area represents the ratio’s numerator with a larger PLV value than the denominator, while the cyan-blue (values greater than 0 and less than 1.0 in the colorbar) are the opposite. The darker the color, the greater the difference in PLV between the two groups. The numbers on the abscissa and ordinate in each subplot represent the channels of the electroencephalography (EEG).

FIGURE 4

Accuracy, sensitivity, and specificity for the classification based on PSD, PLV, and PSD+PLV features in the β and γ bands by using support vector machine (SVM) for the four comparison groups (i.e., HC vs. PD_OFF, HC vs. PD_ON, PD_OFF vs. PD_ON in UC San Diego dataset, and HC vs. PD in Iowa dataset).

FIGURE 5

Accuracy, sensitivity, and specificity for the classification based on PSD+PLV features in the β and γ bands by using SVM and MCNN for the four comparison groups (i.e., HC vs. PD_OFF, HC vs. PD_ON, PD_OFF vs. PD_ON in UC San Diego dataset, and HC vs. PD in Iowa dataset).

FIGURE 6

Receiver operating characteristic curve (ROC) and their Area Under Curves (AUC) for the SVM and proposed MCNN model used to classify subjects into HC, and PD (PD_OFF and PD_ON) based on PSD+PLV features.

FIGURE 7

(A,B) Loss and accuracy of MCNN models during training and testing process in HC vs. PD (Iowa dataset) classification based on γ-band PSD+PLV features.