ISSN 0439-755X
CN 11-1911/B

心理学报 ›› 2019, Vol. 51 ›› Issue (10): 1116-1127.doi: 10.3724/SP.J.1041.2019.01116

• 研究报告 • 上一篇    下一篇

基于事件相关电位(ERPs)和机器学习的考试焦虑诊断 *

章文佩1,2, 沈群伦3, 宋锦涛1, 周仁来1()   

  1. 1 南京大学心理系, 南京 210023
    2 安徽工业大学工商管理系, 马鞍山 243032
    3 中国科学院数学与系统科学研究院, 北京 100190
  • 收稿日期:2018-10-29 发布日期:2019-08-19 出版日期:2019-11-25
  • 通讯作者: 周仁来
  • 基金资助:
    * 中央高校基本科研业务费专项资金()(14370303);江苏省普通高校学术学位研究生科研创新计划项目(KYZZ16_0010);安徽省高校人文科学研究项目资助(SK2017A0084)

Classification of test-anxious individuals using Event-Related Potentials (ERPs): The effectiveness of machine learning algorithms

ZHANG Wenpei1,2, SHEN Qunlun3, SONG Jintao1, ZHOU Renlai1()   

  1. 1 Department of Psychology, Nanjing University, Nanjing, 210023, China
    2 Department of Business Administration, School of Business, Anhui University of Technology, Maanshan, 243032, China
    3 Academy of Mathematics and Systems Science, Chinese Academy of Sciences, 100190, China
  • Received:2018-10-29 Online:2019-08-19 Published:2019-11-25
  • Contact: ZHOU Renlai


考试焦虑对个体的身心具有严重危害。传统诊断考试焦虑的方法容易受到个体主观态度的影响, 从而影响对个体考试焦虑的发现与及早干预。为了克服传统主观问卷对考试焦虑群体诊断的不足, 本研究提出脑电神经数据结合机器学习的客观综合诊断方法评估个体的考试焦虑水平。研究采用情绪Stroop范式, 结合脑电技术测量个体对考试焦虑者的注意抑制功能, 机器学习基于此前提, 提取P1, P2, N2, P3和LPP五种事件相关电位(ERP)成分, 以卷积神经网络(CNN)为主采用7种常见的机器学习算法对个体考试焦虑程度进行进一步的诊断。结果表明CNN对考试焦虑诊断的准确率达86.5%, F1-score为0.911, 显著高于其他6种常见算法。因此采用CNN对脑电信号进行深度学习得出的诊断模型能够有效地对个体的考试焦虑程度进行诊断。

关键词: 机器学习, 考试焦虑, 情绪Stroop, ERPs


Individuals with test anxiety always treat tests/examinations as a potential threat. This cognitive mode impairs these individuals’ cognition, attention and emotions. A traditional method classifying subjects either as high or low on test anxiety (i.e., HTA or LTA, respectively) relies on questionnaire data. Questionnaire data may be unstable due to the subjective nature of participants’ attitudes, implying a reduced classification accuracy. In search for higher levels of (data) stability and classification accuracy a new classification approach is proposed. This new approach overcomes subjective data’s negative impact on classification accuracy by relying on event-related potential (EPR) data (also referred to as ERPs), objective (multivariate, longitudinal) data which adequately capture participants’ reactions to relevant stimuli (over time). However, as ERP data may still be somewhat unstable due to individual differences between participants, (machine) learning algorithms are adopted as their ‘learning’ feature may increase both the stability of ERP data and classification accuracy.

This study recruited 57 HTA participants and 25 LTA participants based on: (a) Test Anxiety Scale (TAS) scores, and (b) (two) specialists’ psychological diagnostic results on a single participant. Reliance on the emotional Stroop (ES) paradigm in combination with ERP technology enabled the assessment of participants’ cognitive mode related to test anxiety. In ES, the information on the ERP components P1, P2, N2, P3 and LPP ERP were selected as input for seven commonly used machine learning algorithms: Convolutional Neural Network (CNN), Logistic Regression (LR), K Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest (RF), Artificial Neural Network (ANN), and Recurrent Neural Network (RNN). To compare the classification accuracy of these algorithms (using the complete sample of HTA and LTA subjects) important indexes (i.e., accuracy and F1-score) were calculated and compared across these algorithms.

The results showed that: (a) the ERPs data collected in ES allow effective differentiation between HTA and LTA (P1: F(1, 80) = 11.68, p < 0.001, η 2 = 0.13; P2: F(1, 80) = 14.10, p < 0.001, η 2 = 0.15; N2: F(1, 80) = 28.55, p < 0.001, η 2 = 0.26; P3: F(1, 80) = 22.41, p < 0.001, η 2 = 0.22; LPP: F(1, 80) = 16.92, p < 0.001, η 2 = 0.18); (b) classification on the basis of ERP data using machine learning algorithms shows high accuracy and stability, that is the classification accuracy of all seven algorithms is found to be high as evidenced by an accuracy index of 71.8% or higher (CNN: 86.5%, LR: 80.3%, KNN: 71.8%, SVM: 79.0%, RF: 73.1%, ANN: 82.7%, and RNN: 79.2%) and an F1-score of 0.814 or higher (CNN: 0.911, LR: 0.868, KNN: 0.817, SVM: 0.865, RF: 0.814, ANN: 0.882, and RNN: 0.870); (c) CNN outperforms the other six common machine learning algorithms showing both the highest accuracy index and F1-score. Moreover, as over and above this (relative) superiority CNN combines the (technical) property known as ‘shift invariance’ and robustness to noise, the algorithm may be considered ideal for effectively classifying test anxious individuals using ERP data.

It is concluded that: (a) as manifested by its ‘discriminatory’ nature and stable classification performance (as evidenced by all machine learning algorithms’ favorable values for all important indices) reliance on the ES paradigm enables machine learning leading up to effective diagnosis of test anxiety; and (b) participants’ classification into HTA and LTA by relying on ERP data which are subsequently analyzed by means of the machine learning algorithm CNN is (most) effective (i.e., as benchmarked against six other commonly used machine learning algorithms). Consequently, using ES in combination with ERP technology and the CNN machine learning algorithm can be conceived as an ideal method for diagnosing test anxiety.

Key words: machine learning, test anxiety, emotional Stroop, ERPs