## Introduction

Ref | Method | Signal | Disease type | Accuracy | Year |
---|---|---|---|---|---|

[6] | Correlation, phase synchrony, and Granger causality measures | EEG | MCI and mild AD | 83% and 88%, respectively | 2012 |

[7] | Hybrid feature selection | EEG | MCI and mild AD | 95% and 100%, respectively | 2015 |

[8] | Complex network theory and TSK fuzzy system | EEG | AD | 97.3% | 2019 |

[9] | Functional connectivity and effective connectivity analysis | MEG | AD | 86% | 2019 |

[10] | Phase locking value, imaginary part, and correlation of the envelope | MEG | MCI | 75% | 2019 |

[11] | High-order FC correlations | fMRI | MCI | 88.14% | 2016 |

[12] | Hierarchical high-order functional connectivity networks | fMRI | MCI | 84.85% | 2017 |

[13] | Strength and similarity guided GSR using LOFC and HOFC | fMRI | MCI | 88.5% | 2019 |

## Method

### Alzheimer’s Datasets

#### The MCI Dataset

#### The Mild AD Dataset

#### Recording Conditions in Both Datasets

### A Decomposition and Recombination System

### Neural Network Classifiers

#### BrainNet CNN

Layer | Output size | Parameter |
---|---|---|

Input layer | [B, 4, C, C] | |

BatchNorm | [B, 4, C, C] | |

ReLU | [B, 4, C, C] | |

E2E | [B, 16, C, C] | (C, 1) |

BatchNorm | [B, 16, C, C] | |

ReLU | [B, 16, C, C] | |

E2E | [B, 32, C, C] | |

ReLU | [B, 32, C, C] | |

E2N | [B, 64, C, 1] | (1, C) |

N2G | [B, 512, 1, 1] | (C, 1) |

Flatten | [B, 512] | |

Linear and softmax | [B, 2] |

#### ResNet

#### EEGNet

Layer | Output size | Parameter |
---|---|---|

Input layer | [B, 1, C, T] | |

ZeroPad2d | [B, 1, C, T+63] | (31, 32, 0, 0) |

Conv2d | [B, 8, C, T] | (1, 64) |

BatchNorm2d | [B, 8, C, T] | |

Conv2d | [B, 16, 1, T] | (C, 1), grouped |

BatchNorm2d | [B, 16, 1, T] | |

ELU | [B, 16, 1, T] | |

AvgPool2d | [B, 16, 1, T//4] | (1, 4) |

Dropout | [B, 16, 1, T//4] | 0.25 |

ZeroPad2d | [B, 16, 1, T//4+15] | (7, 8, 0, 0) |

Conv2d | [B, 16, 1, T//4] | (1, 15), grouped |

Conv2d | [B, 16, 1, T//4] | (1, 1) |

BatchNorm2d | [B, 16, 1, T//4] | |

ELU | [B, 16, 1, T//4] | |

AvgPool2d | [B, 16, 1, T//32] | (1, 8) |

Dropout | [B, 16, 1, T//32] | 0.25 |

Flatten | [B, 16*T//32] | |

Linear | [B, K] | \(bias=False\) |

#### Parameter Setting

## Results

Training set | Testing set | Chance level | ||
---|---|---|---|---|

Data type | Artificial | Original | Original | |

Data type | Artificial | Original | Original | |

Mild AD | 0–500 | 10 | 7 | 0.3333 |

Control | 0–500 | 10 | 14 | |

MCI | 0–500 | 10 | 12 | 0.3000 |

Control | 0–500 | 10 | 28 |

### Feature Distribution

### Performance Analysis

ResNet | Accuracy | Sensitivity | Precision | |
---|---|---|---|---|

Mild AD | Before | 0.7238 | 0.7393 | 0.7243 |

After | 0.7762 | 0.7893 | 0.7628 | |

MCI | Before | 0.6450 | 0.6631 | 0.6450 |

After | 0.6900 | 0.7000 | 0.6775 | |

BrainNet CNN | Accuracy | Sensitivity | Precision | |

Mild AD | Before | 0.7476 | 0.7464 | 0.7338 |

After | 0.7714 | 0.7750 | 0.7850 | |

MCI | Before | 0.6650 | 0.6655 | 0.6469 |

After | 0.6725 | 0.6661 | 0.6517 | |

EEGNet | Accuracy | Sensitivity | Precision | |

Mild AD | Before | 0.6429 | 0.6321 | 0.6336 |

After | 0.5667 | 0.5929 | 0.5887 | |

MCI | Before | 0.4625 | 0.5399 | 0.5399 |

After | 0.4850 | 0.5393 | 0.5337 |