兰州大学机构库 >信息科学与工程学院
基于脑电及卷积神经网络的抑郁症实时监测研究
Alternative TitleResearch on Depression's Real-time Monitoring Based on Electroencephalograph and Convolutional Neutral Network
赵盛杰
Subtype硕士
Thesis Advisor姚志军
2018-04-03
Degree Grantor兰州大学
Place of Conferral兰州
Degree Name硕士
Keyword普适化EEG 机器学习 CNN 实时EEG采集 抑郁风险预测
Abstract

近年来,越来越多的研究开始着眼于以脑电(electroencephalograph,EEG)为代表的生理信号的量化分析,以实现对抑郁症等心理障碍疾病更为客观和有效的诊断。然而,抑郁症作为一种非常复杂且随时间发展动态变化的心理障碍疾病,其有效诊断往往依赖于对患者心理状态的持续评估,而传统的多导脑电采集仪因价格昂贵且需要专人操作很难实现对患者EEG数据的有效跟踪,急需可以对患者EEG进行长期监测和跟踪采集的解决方案。同时,由于EEG信号具有非线性、非平稳等特点,且存在个体生理差异性,如何提取EEG信号中更能反映患者心理状态的特征,构建泛化能力更强的预测模型就成为实现抑郁症有效诊断和监测的重要前提。

针对上述问题,本文基于便携式脑电传感器构建了普适化EEG信号的跟踪采集和实时量化评估框架,利用卷积神经网络(Convolutional Neural Network,CNN)对普适化EEG构建抑郁症分类模型,借鉴其自动特征提取优势、优良空间特性以及面对大量数据时优秀的新分类模式发现能力,保证模型的快速更迭与泛化能力。主要工作和贡献包括以下三个方面:

1) 结合EEG实时采集模块与抑郁风险预测模块设计两阶段实时抑郁症监测方法:EEG实时采集阶段主要实现实时滤波、实时眼电伪迹剔除、数据可视化,通过普适化方法将传统的单次EEG采集上升到长期性的抑郁症监测上;抑郁风险预测阶段主要实现多个分类模型决策层融合给出被试的抑郁风险评估。

2) 对静息态脑电进行CNN网络结构探索,从浅层网络入手,逐步分析网络参数不同选取对最终分类效果的影响。针对浅层CNN中面临的问题,提出更为鲁棒的抑郁症分类模型:基于特征图的神经网络以及基于残差连接的深度残差网络。相比传统的分类算法,这两种基于CNN的方法可以充分利用导联之间的地理关联,使得分类性能有显著提升,其中特征图神经网络达到了76.33%的平均准确率,深度残差网络达到了80.33%的平均准确率。

3) 对引入音频实验范式的EEG构建CNN与长短期记忆网络(Long Short-term Memory, LSTM)相结合的分类模型。与静息态脑电中采用原始信号不同,时序标记的EEG数据需转为更易于分析的图片模式转为类似视频分类的任务。卷积操作可预先利用空间信息自动提取特征,而LSTM则对卷积输出结果在时序上建模。本文对三种主流时空信息提取方法进行对比实验,包含三种特征提取器:时域卷积、CNN +LSTM、ConvLSTM。其中,ConvLSTM达到了最优的82.35%分类准确率。

Other Abstract

In recent years, more and more studies have begun to focus on the quantitative analysis of physiological signals represented by electroencephalogram (EEG) in order to achieve a more objective and effective diagnosis of mental disorders such as depression. However, depression is a kind of mental disorder with very complex and dynamic changes over time. Its effective diagnosis often depends on the continuous evaluation of the patient's psychological status. However, the traditional multi-channel EEG acquisition device is difficult to achieve effective tracking of the patient's EEG data due to its high price and need of special operations, and it is urgent to provide a solution for long-term monitoring and tracking acquisition of the patient EEG. At the same time, because the EEG signal has characteristics such as non-linearity, non-stationarity, and individual physiological differences, how to extract the features of the EEG signals that can better reflect the patient's psychological state and build a prediction model with stronger generalization ability become an important prerequisite for effective diagnosis and monitoring of depression.

In order to solve the above problems, this paper builds a framework for tracking acquisition and real-time quantitative evaluation of pervasive EEG signals based on wearable EEG sensors, and uses the Convolutional Neural Network (CNN) to build a classification model of depression for the pervasive EEG, which draws on the advantages of automatic feature extraction, excellent spatial characteristics, and new classification pattern discovery capabilities in large amounts of data to ensure rapid model update and generalization capabilities. The main work and contributions include the following three aspects:

1)Combined with the EEG real-time acquisition module and the depression risk prediction module, a two-stage real-time depression monitoring method is designed: the real-time EEG acquisition stage mainly realizes real-time filtering, real-time eye artifact elimination, and data visualization, which make the traditional EEG acquisition rise to long-term monitoring of depression; the depressive risk prediction stage mainly implements multiple classification model decision fusion to give the participants' depression risk assessment.

2)The exploration of the CNN network structure in the resting state EEG starts with the shallow network and gradually analyzes the influence of different selection of network parameters on the final classification effect. For the problems faced in shallow CNN, a more robust classification model of depression is proposed: neural network based on feature maps and deep residual network based on residual connections. Compared with traditional classification algorithms, these two methods based on CNN can fully utilize the geographic association between the electrodes, which makes the classification performance significantly improved. The feature map neural network achieves an average accuracy of 76.33%, and the deep residual network reaches an average accuracy of 80.33%.

 

3) For the EEG that lead into the audio experiment paradigm, a classification model combining CNN and Long Short-term Memory (LSTM) is constructed. Different from the original signal used in the resting state EEG, the time-stamped EEG data needs to be converted into a more easily-analyzed picture mode and turned into a task similar to video classification. The convolution operation can use the spatial information to automatically extract features in advance, while the LSTM models the convolution output results in time series. This article compares three main temporal and spatial information extraction methods and includes three feature extractors: time domain convolution, CNN + LSTM, and ConvLSTM. Among them, ConvLSTM achieved the best 82.35% classification accuracy.

URL查看原文
Language中文
Document Type学位论文
Identifierhttps://ir.lzu.edu.cn/handle/262010/232636
Collection信息科学与工程学院
Recommended Citation
GB/T 7714
赵盛杰. 基于脑电及卷积神经网络的抑郁症实时监测研究[D]. 兰州. 兰州大学,2018.
Files in This Item:
There are no files associated with this item.
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Altmetrics Score
Google Scholar
Similar articles in Google Scholar
[赵盛杰]'s Articles
Baidu academic
Similar articles in Baidu academic
[赵盛杰]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[赵盛杰]'s Articles
Terms of Use
No data!
Social Bookmark/Share
No comment.
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.