Table 1 Classification performance and computational complexity of different deep learning models based on voice-related EEG signals, where k represents the convolution kernel size, d represents the time series length, and n represents the number of channels.

From: An interpretable model based on graph learning for diagnosis of Parkinson’s disease with voice-related EEG

Models

 

Evaluation Criterion

ACC

AUC

Sensitivity

1-Specificity

Computational Complexity

Baseline Models

CNN

79.6%

78.1%

80.2%

77.5%

4*O(k·n·d2)+O(n2)

RNN

75.2%

74.2%

80.1%

75.3%

O(n·d2)+O(n2)

2D-CNN-RNN

81.6%

79.5%

78.6%

80.2%

O(k·n·d2)+O(n·d2)+O(n2)

3D-CNN-RNN

82.1%

81.2%

83.1%

78.2%

2*O(k·n·d2)+O(n·d2)+O(n2)

EEGNet

82.3%

81.2%

81.5%

79.6%

3*O(k·n·d2)+O(n2)

Cascade model

80.4%

83.2%

82.1%

77.2%

3*O(k·n·d2)+2*O(n·d2)+2*O(n2)

Parallel model

81.1%

80.2%

82.3%

79.8%

3*O(k·n·d2) + O(n2)

GCNs

PCC+GCNs

84.1%

83.1%

78.2%

86.0%

2*O(nlogn)+O(n2)

PLV+GCNs

84.8%

82.2%

79.1%

84.3%

2*O(nlogn)+O(n2)

PLV + GSP-GCNs

88.9%

87.5%

86.1%

86.2%

2*O(nlogn)+O(n2)

PCC + GSP-GCNs

90.2%

89.1%

84.0%

88.4%

2*O(nlogn)+O(n2)

  1. ACC accuracy, AUC area under curve, CNN convolutional neural network, RNN recurrent neural networks, GSP graph signal processing, GCN graph convolutional network, PCC Pearson correlation coefficient, PLV phase locking value.
Back to article page