文献速递:深度学习–深度学习方法用于帕金森病的脑电图诊断

01

文献速递介绍

人类大脑在出生时含有最多的神经细胞,也称为神经元。这些神经细胞无法像我们身体的其他细胞那样自我修复。随着年龄的增长,神经元逐渐死亡,因此变得无法替代。PD(帕金森病)通常随着神经元的死亡而发生。神经元产生一种称为多巴胺的化学物质,其主要功能是控制身体的运动。因此,随着神经元的死亡,大脑中产生的多巴胺量减少。结果,这种神经系统状况开始非常缓慢地发生,并影响大脑中的各种通信方式。已观察到大约50岁或更老的人被诊断出患有PD。这种疾病的主要症状包括不稳定的姿势、肌肉僵硬、动作缓慢、震颤、平衡失调和精细运动技能受损。根据世界卫生组织提供的统计数据,这种疾病已经影响了近1000万人。在未观察到明显的运动或非运动症状时,诊断这种疾病存在困难。因此,计算机辅助诊断(CAD)系统可能有助于早期检测任何异常。CAD系统是一种自动化检测系统,可以使用脑电图(EEG)信号客观地诊断PD。借助EEG,可以轻松识别大脑皮层和皮层下部分的功能。神经系统疾病如癫痫、精神分裂症、阿尔茨海默症也可以使用EEG信号确定。因此,在这项研究中,我们使用EEG信号开发了用于检测PD的CAD系统。

根据先前的研究,EEG信号是复杂和非线性的,因此许多线性特征提取方法无法准确描述这些信号。当EEG信号显示复杂性时,观察到PD的加重。这是因为EEG信号中存在非线性成分。因此,可以注意到,使用非线性特征提取技术在正常和PD EEG信号的区分中将是有用的。

然而,近年来在模式识别和自然语言处理的多个领域成功实施了机器学习的一个分支——深度学习。卷积神经网络(CNN)是研究者采用的最流行的深度学习形式之一。它允许通过数据训练,无需人工干预即可学习高级特征,不同于大多数传统的机器学习算法。据我们所知,这是第一篇利用深度CNN实施PD CAD系统的论文。我们实现了一个新颖的十三层深CNN来表征两个类别(PD和正常)。图1展示了所提出网络的架构。网络及每一层的详细信息在后续章节中介绍。

Title

题目

A deep learning approach for Parkinson’s disease diagnosis from EEG signals

深度学习方法用于帕金森病的脑电图诊断

Abstract

摘要

An automated detection system for Parkinson’s disease (PD) employing the convolutional neural network (CNN) is proposed in this study. PD is characterized by the gradual degradation of motor function in the brain. Since it is related to the brain abnormality, electroencephalogram (EEG) signals are usually considered for the early diagnosis. In this work, we have used the EEG signals of twenty PD and twenty normal subjects in this study. A thirteen-layer CNN architecture which can overcome the need for the conventional feature representation stages is implemented. The developed model has achieved a promising performance of 88.25% accuracy, 84.71% sensitivity, and 91.77% specificity. The developed

classification model is ready to be used on large population before installation of clinical usage.

本研究提出了一种使用卷积神经网络(CNN)的帕金森病(PD)自动检测系统。PD的特点是大脑运动功能逐渐退化。由于它与大脑异常有关,因此通常考虑使用脑电图(EEG)信号进行早期诊断。在这项工作中,我们使用了二十名PD患者和二十名正常受试者的EEG信号进行研究。实现了一种十三层的CNN架构,它可以克服传统特征表示阶段的需求。开发的模型达到了88.25%的准确率、84.71%的敏感性和91.77%的特异性的有希望的性能。开发的分类模型已经准备好在临床使用前在大型人群中使用。

**Results
**
结果

All the EEG signals were subjected to the proposed CNNmodel. The CNN network was designed in Python lan guage using Keras and was executed on a computer with a system configuration of two Intel Xeon 2.40 GHz (E5620)processors with a 24 GB random access memory.

The evaluation parameters, namely the accuracy, sen**sitivity, and specificity, were used. The best diagnostic performance is achieved wit the learning rate of 0.0001. The proposed CNN model yielded an accuracy of 88.25%, sensitivity, and specificity of 84.71% and 91.77%, respectively. Figures 3 and 4 show the performance of the model with and without dropout layer, respectively. It can be noted that without the dropout layer, there is a possi bility of overfitting of data. In Fig. 3, the accuracy of the training set does not differ much from the accuracy of the validation set, whereas, in Fig. 4, the accuracy of the val idation set performs a lot worse as compared to the training data.

Figure 5 shows the confusion matrix of our results. It can be observed that 11.34% of normal subjects are mis classified as PD and 11.51% of the PD EEG signals are wrongly categorized into the normal class.

所有EEG信号都被应用到了所提出的CNN模型中。该CNN网络是用Python语言通过Keras设计的,并在一台配置有两个Intel Xeon 2.40 GHz(E5620)处理器和24 GB随机访问内存的计算机上执行。

评估参数,即准确率、敏感性和特异性被使用。最佳的诊断性能是在学习率为0.0001时达到的。所提出的CNN模型取得了88.25%的准确率,以及84.71%的敏感性和91.77%的特异性。图3和图4分别展示了模型带有和不带有dropout层的性能。可以注意到,没有dropout层时,数据过拟合的可能性存在。在图3中,训练集的准确率与验证集的准确率相差不大,而在图4中,验证集的准确率与训练数据相比表现得更差。

图5展示了我们结果的混淆矩阵。可以观察到,11.34%的正常受试者被误分类为PD,而11.51%的PD EEG信号被错误地归类为正常类。

Conclusion

结论

An automated thirteen-layer CNN model to diagnose PD using EEG signals is proposed. Furthermore, this is the first study which implemented the deep learning concept to diagnose the PD using EEG signals. We have obtained an accuracy of 88.25%, sensitivity of 84.71%, and specificity of 91.77% despite the limited number of subjects. Based on the positive performances achieved, the presented model may be able to serve as a trusted and long-term tool to assist clinicians in PD diagnoses. In the future, authors propose to test the developed model with a huge number of subjects and also aim to detect the early stage of PD.

提出了一种自动化的十三层CNN模型,用于利用EEG信号诊断PD。此外,这是第一项将深度学习概念应用于使用EEG信号诊断PD的研究。尽管受试者数量有限,我们仍获得了88.25%的准确率、84.71%的敏感性和91.77%的特异性。基于所取得的积极表现,所展示的模型可能能够作为一个可信赖的和长期的工具,以协助临床医生诊断PD。未来,作者提议使用大量受试者测试开发的模型,并且还旨在检测PD的早期阶段。

Figure

Fig. 1 The proposed CNN architecture

图 1 所提出的CNN架构

Fig. 2 A sample of a normal and b PD EEG signal

图 2 a 正常和 b PD EEG信号的样本

Fig. 3 Accuracy versus different epoch plot

图 3 准确率与不同轮次的关系图

Fig. 4 Accuracy versus different epoch without dropout layer plot

图 4 没有dropout层时准确率与不同轮次的关系图

Fig. 5 Confusion matrix of the proposed method

图 5 所提出方法的混淆矩阵

Fig. 6 Web-based CAD system to diagnose PD

图 6 基于网络的CAD系统用于诊断PD

Table

Table 1 Details of parameters belonging to different layers of the developed CNN model

表 1 开发的CNN模型不同层的参数详情

Table 2 The summary of CADsystem developed using EEG signals to diagnose PD

表 2 使用EEG信号开发的CAD系统诊断PD的总结