CASIA OpenIR  > 毕业生  > 硕士学位论文
基于脑电信号的跨域情绪识别及抑郁评估研究
杨子阳
2024-05-14
Pages100
Subtype硕士
Abstract

情绪是个体对外部刺激的复杂生理和心理反应,其受人格特征、社会环境等多种因素影响,在人类认知活动中发挥着基础作用。由于情绪功能障碍常被视为抑郁症的核心表现,因此实现精准的情绪识别有助于发展评估抑郁症的有效方法。脑电信号的变化反映了大脑活动,通过解码这些信号能够实现情绪识别和抑郁评估。然而,此过程面临一个重大挑战:脑电数据分布在不同个体和时间中存在变化,这种现象被称为“域偏移”。

传统深度学习方法通常需要大量有标签数据才能获得出色的分类模型,但是收集这类数据费时费力。此外,传统深度学习方法对未知数据的泛化能力较弱,微小的数据分布差异也可能导致分类性能显著下降。尽管域适应方法已被用于解决情绪识别中的域偏移问题,但在抑郁评估与分级中应用并优化域适应方法的研究仍然较少。更重要的是,由于算力的限制或隐私保护的需要,医学领域常常无法直接获取源域数据,这给域适应增加了难度。为解决这些问题,本文围绕情绪识别与抑郁评估任务和域适应方法展开研究,主要内容分为以下两个方面:

(1)针对脑电分类任务中的域偏移问题,本文提出了一种分类器驱动的隐式鉴别器无监督域适应方法。该方法有两个关键特性:首先,采用隐式域鉴别器模型设计,减少了参数数量和计算成本;其次,利用联合分类器集成多个优化目标,融合三种类型的损失,能在最小化域间差异的同时最大化预测准确性。借助这些设计,模型可以对齐特征的边缘分布,并强化类别的条件分布。该模型的情绪识别能力在公开的SEED数据集上得到了验证,其在跨个体和跨时间任务中分别达到了90.89%和92.78%的准确率。同时,模型抑郁评估方面的表现在自建的SignBrain DP数据集上得到了检验,其在跨个体抑郁程度四分类、三分类和二分类场景上分别取得了68.57%、82.50%和94.24%的准确率。实验结果显示,所提出的方法在多种任务中展现出优异的性能。此外,对频段和脑区的分析揭示了模型的决策依据,不仅印证了神经生理机制,也证实模型具备良好的可解释性,推动了脑电信号在临床应用中的发展。
(2)针对无监督域适应方法中对源域数据的依赖问题,本文提出了一种基于注意力的分类器差异无源域适应方法。该方法将训练过程分为两个阶段:预训练阶段和域适应阶段。其中预训练阶段仅使用源域数据,目标是得到在源域数据上性能良好的分类模型。域适应阶段只使用目标域数据和源域模型,通过两步训练策略交替优化分类器与特征提取器,实现源域模型到目标域的迁移。实验结果表明,所提出的方法不仅能节省资源和保护隐私,性能还超越了其他无源域适应方法,可与领先的无监督域适应方法相匹敌。在跨个体情绪识别和四分类抑郁程度分级任务中,模型在离线场景下分别取得了87.40%和64.64%的准确率,并在更真实的在线场景下分别取得了78.40%和59.26%的准确率,充分体现了其卓越的分类性能和现实应用潜力。同时,此方法还具有更快的训练速度,从而提升了用户体验。此外,通过注意力机制对不同通道的贡献进行分析,探索了仅用关键通道进行分类的可能性,为模型和设备的轻量化、小型化发展提供了参考。

Other Abstract

Emotion is a complex physiological and psychological response of an individual to external stimuli, which is influenced by a variety of factors such as personality traits and social environment, and plays a fundamental role in human cognitive activities. Since emotional dysfunction is often regarded as the core manifestation of depression, achieving accurate emotion recognition can help develop effective methods for assessing depression. Changes in EEG signals reflect brain activity, and decoding these signals enables emotion recognition and depression assessment. However, this process faces a major challenge: the distribution of EEG data varies across individuals and time, a phenomenon known as "domain shift".

Traditional deep learning methods usually require a large amount of labeled data to obtain excellent classification models, but collecting such data is time-consuming and laborious. In addition, traditional deep learning methods have a weak ability to generalize to unknown data, and small data distribution differences may also lead to a significant degradation of classification performance. Although domain adaptation methods have been used to solve the domain shift problem in emotion recognition, there are still fewer studies applying and optimizing domain adaptation methods in depression assessment and grading. What's more, due to the limitation of computing power or the need of privacy protection, the medical field often cannot directly access the source domain data, which increases the difficulty of domain adaptation. To address these issues, this paper centers on emotion recognition and depression assessment tasks as well as domain adaptation methods, with the main contents divided into the following two aspects:
(1) Aiming at the domain shift problem in EEG classification tasks, this thesis proposes a classifier-driven implicit discriminator unsupervised domain adaptation method. The method has two key properties: firstly, the employment of an implicit discriminator model design, which reduces both the number of parameters and computational costs, and secondly, the integration of multiple optimization objectives through a joint classifier and the fusion of three types of losses, minimizing the differences between domains and maximizing prediction accuracy. With these designs, the model can align the marginal distribution of features and strengthen the conditional distribution of categories. The model's emotion recognition ability is validated on the publicly available SEED dataset, where it achieves accuracies of 90.89% and 92.78% in the cross-subject and cross-session tasks, respectively. Meanwhile, the model's performance in depression assessment was examined on the self-made SignBrain DP dataset, which achieved accuracies of 68.57%, 82.50%, and 94.24% in cross-subject four-class, three-class, and two-class depression severity grading scenarios, respectively. Experimental results show that the proposed method exhibits excellent performance in various tasks. In addition, the analysis of frequency bands and brain regions reveals the decision-making basis of the model, which not only corroborates the neurophysiological mechanisms, but also confirms that the model possesses good interpretability, promoting the development of EEG signals in clinical applications.
(2) Aiming at the problem of dependence on source domain data in unsupervised domain adaptation methods, this thesis proposes an attention-based classifier discrepancy source-free domain adaptation method. The method divides the training process into two phases: a pre-training phase and a domain adaptation phase. Where the pre-training phase uses only the source domain data, and the goal is to obtain a classification model that performs well on the source domain data. The domain adaptation phase uses only the target domain data and the source domain model, and achieves the transfer of the source domain model to the target domain by alternating the optimization of the classifier and the feature extractor through a two-step training strategy. Experimental results show that the proposed method not only saves resources and protects privacy, but also outperforms other source-free domain adaptation methods and can match the leading unsupervised domain adaptation methods. In the cross-subject emotion recognition and four-class depression severity grading tasks, the model achieves accuracies of 87.40% and 64.64% under the offline setting and 78.40% and 59.26% under the more realistic online setting, which fully demonstrates its excellent classification performance and real-world application potential. At the same time, this method also has a faster training speed, which enhances the user experience. In addition, by analyzing the contribution of different channels through the attention mechanism, the possibility of using only key channels for classification is explored, which provides a reference for the development of lightweight and miniaturization of models and devices.

Keyword脑电 (EEG) 情绪识别 抑郁评估 深度学习 域适应
Subject Area模式识别 ; 计算机应用
MOST Discipline Catalogue工学::计算机科学与技术(可授工学、理学学位)
Language中文
IS Representative Paper
Sub direction classification脑机接口
planning direction of the national heavy laboratoryAI For Science
Paper associated data
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/56629
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
杨子阳. 基于脑电信号的跨域情绪识别及抑郁评估研究[D],2024.
Files in This Item:
File Name/Size DocType Version Access License
毕业论文.pdf(5686KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[杨子阳]'s Articles
Baidu academic
Similar articles in Baidu academic
[杨子阳]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[杨子阳]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.