4.6 Article

EEG-based multi-level stress classification with and without smoothing filter

期刊

出版社

ELSEVIER SCI LTD
DOI: 10.1016/j.bspc.2021.102881

关键词

Stress; Classification; EEG; Signal smoothing; Data leakage

资金

  1. Spanish Ministry of Science, Innovation and Universities [PGC2018-098813-B-C31]
  2. European Regional Development Funds
  3. Nicolo Association

向作者/读者索取更多资源

The study examined the impact of EEG-PSD smoothing on three-level stress classification and found that smoothing can lead to data leakage and affect classification performance. Two-level stress classification without smoothing met the criteria for practical applicability, suggesting individual processing of each epoch is necessary for realistic stress classifiers.
Recently, multi-level stress assessment has become an active research subject. In this context, researchers typically develop models based on machine learning classifiers and features extracted from biosignals like electrocardiogram (ECG) or electroencephalogram (EEG). For that purpose, EEG power spectral density (PSD) is a recurrent feature owing to its high responsiveness and remarkable performance. However, PSD is usually smoothed to cope with its bursty nature, what may cause data leakage and hence call into question classification performance. In this study, our aim was twofold: first, to examine the effect of EEG-PSD smoothing in three-level stress classification, and second, to evaluate the practical viability of a two-level stress detector without smoothing. To this end, we conducted participants through a stress-relax session while recording their EEG. Then, we estimated the EEG-PSD and used the stress reported by the participants as labels for classification. Initially, we developed a three-level stress classifier and examined the effect of smoothing on its performance. We found that classification performance was directly proportional to smoothing intensity (F1-score 0.61-0.94), and also that when smoothing was not applied to features, classification performance was insufficient for practical applicability (AUC < 0.7). We link this behavior to train-test contamination due to smoothing. Subsequently, we attempted two-level stress classification without smoothing. In this case, performance met the criteria for practical applicability (AUC = 0.76). This suggest that performance enhancement in three-level stress classification was caused by data leakage produced by smoothing, and hence, to render realistic stress classifiers each epoch should be processed individually.

作者

我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。

评论

主要评分

4.6
评分不足

次要评分

新颖性
-
重要性
-
科学严谨性
-
评价这篇论文

推荐

暂无数据
暂无数据