Affective brain-computer interfaces: Choosing a meaningful performance measuring metric
Affective brain-computer interfaces: Choosing a meaningful performance measuring metric
复制标题
DOI：
10.1016/j.compbiomed.2020.104001
复制
发表时间：
2020-11-01
影响因子：
7.7
通讯作者：
Thompson, David E.
中科院分区：
文献类型：
--
作者：
Mowla, Md Rakibul;Cano, Rachael, I;Thompson, David E.
情感脑机接口是情感计算中一个相对较新的研究领域。情感状态的估计可以改善人机交互，以及改善严重残疾人的护理。为了评估EEG记录识别情感状态的有效性，我们使用了实验室收集的数据以及公开可用的DEAP数据库。我们还回顾了使用DEAP数据库的文章，发现大量文章没有考虑DEAP中存在的类不平衡。不考虑阶级不平衡会产生误导性的结果。此外，忽略类别不平衡使得使用不同数据集的研究之间的结果比较变得不可能，因为不同的数据集将具有不同的类别不平衡。阶级不平衡也会改变机会水平，因此在确定结果是否高于机会时，考虑阶级偏见至关重要。为了适当地考虑类不平衡的影响，我们建议使用平衡的准确性作为性能指标，其后验分布计算可信区间。对于分类，我们使用了文献中的特征以及θ β-1比率。从DEAP和我们的数据的结果表明，β波段功率，θ波段功率，和θ β-1比率是更好的特征集分类效价，唤醒，和优势，分别。
Affective brain-computer interfaces are a relatively new area of research in affective computing. Estimation of affective states can improve human-computer interaction as well as improve the care of people with severe disabilities. To assess the effectiveness of EEG recordings for recognizing affective states, we used data collected in our lab as well as the publicly available DEAP database. We also reviewed the articles that used the DEAP database and found that a significant number of articles did not consider the presence of the class imbalance in the DEAP. Failing to consider class imbalance creates misleading results. Further, ignoring class imbalance makes the comparison of the results between studies using different datasets impossible, since different datasets will have different class imbalances. Class imbalance also shifts the chance level, hence it is vital to consider class bias while determining if the results are above chance. To properly account for the effect of class imbalance, we suggest the use of balanced accuracy as a performance metric, and its posterior distribution for computing credible intervals. For classification, we used features from the literature as well as theta beta-1 ratio. Results from DEAP and our data suggest that the beta band power, theta band power, and theta beta-1 ratio are better feature sets for classifying valence, arousal, and dominance, respectively.