Electrical and Electronics Engineering Department, Abdullah Gül University, Kayseri, Turkey.
Biomedical Engineering Department, Erciyes University, Kayseri, Turkey.
Comput Methods Programs Biomed. 2014 Feb;113(2):705-13. doi: 10.1016/j.cmpb.2013.11.010. Epub 2013 Nov 26.
In this study, we have analyzed electroencephalography (EEG) signals to investigate the following issues, (i) which frequencies and EEG channels could be relatively better indicators of preference (like or dislike decisions) of consumer products, (ii) timing characteristic of "like" decisions during such mental processes. For this purpose, we have obtained multichannel EEG recordings from 15 subjects, during total of 16 epochs of 10 s long, while they were presented with some shoe photographs. When they liked a specific shoe, they pressed on a button and marked the time of this activity and the particular epoch was labeled as a LIKE case. No button press meant that the subject did not like the particular shoe that was displayed and corresponding epoch designated as a DISLIKE case. After preprocessing, power spectral density (PSD) of EEG data was estimated at different frequencies (4, 5, …, 40 Hz) using the Burg method, for each epoch corresponding to one shoe presentation. Each subject's data consisted of normalized PSD values (NPVs) from all LIKE and DISLIKE cases/epochs coming from all 19 EEG channels. In order to determine the most discriminative frequencies and channels, we have utilized logistic regression, where LIKE/DISLIKE status was used as a categorical (binary) response variable and corresponding NPVs were the continuously valued input variables or predictors. We observed that when all the NPVs (total of 37) are used as predictors, the regression problem was becoming ill-posed due to large number of predictors (compared to the number of samples) and high correlation among predictors. To circumvent this issue, we have divided the frequency band into low frequency (LF) 4-19 Hz and high frequency (HF) 20-40 Hz bands and analyzed the influence of the NPV in these bands separately. Then, using the p-values that indicate how significantly estimated predictor weights are different than zero, we have determined the NPVs and channels that are more influential in determining the outcome, i.e., like/dislike decision. In the LF band, 4 and 5 Hz were found to be the most discriminative frequencies (MDFs). In the HF band, none of the frequencies seemed offer significant information. When both male and female data was used, in the LF band, a frontal channel on the left (F7-A1) and a temporal channel on the right (T6-A2) were found to be the most discriminative channels (MDCs). In the HF band, MDCs were central (Cz-A1) and occipital on the left (O1-A1) channels. The results of like timings suggest that male and female behavior for this set of stimulant images were similar.
在这项研究中，我们分析了脑电图（EEG）信号，以调查以下问题：（i）哪些频率和 EEG 通道可以作为消费者产品偏好（喜欢或不喜欢）的相对较好的指标；（ii）在这种心理过程中，“喜欢”决策的时间特征。为此，我们从 15 名受试者那里获得了多通道 EEG 记录，在总共 16 个 10 秒长的时段中，当他们看到一些鞋子照片时。当他们喜欢特定的鞋子时，他们按下一个按钮，并标记此活动的时间，相应的时段被标记为 LIKE 案例。如果没有按下按钮，则表示受试者不喜欢显示的特定鞋子，相应的时段被指定为 DISLIKE 案例。在预处理之后，使用 Burg 方法针对每个对应于一只鞋子呈现的时段，在不同频率（4、5、…、40 Hz）下估算 EEG 数据的功率谱密度（PSD）。每个受试者的数据由来自所有 19 个 EEG 通道的所有 LIKE 和 DISLIKE 案例/时段的归一化 PSD 值（NPV）组成。为了确定最具判别力的频率和通道，我们利用逻辑回归，其中 LIKE/DISLIKE 状态用作分类（二进制）响应变量，相应的 NPV 用作连续值输入变量或预测因子。我们观察到，当将所有 NPV（总共 37 个）用作预测因子时，由于预测因子的数量（与样本数量相比）较大且预测因子之间存在高度相关性，回归问题变得不适定。为了解决这个问题，我们将频带分为低频（LF）4-19 Hz 和高频（HF）20-40 Hz 频段，并分别分析这些频段中的 NPV 影响。然后，使用指示估计的预测因子权重与零的差异有多大的 p 值，我们确定了在确定结果（即喜欢/不喜欢决策）方面更具影响力的 NPV 和通道。在 LF 频段中，4 Hz 和 5 Hz 被确定为最具判别力的频率（MDFs）。在 HF 频段中，似乎没有任何频率提供显著的信息。当同时使用男性和女性数据时，在 LF 频段中，左侧的额通道（F7-A1）和右侧的颞通道（T6-A2）被确定为最具判别力的通道（MDCs）。在 HF 频段中，MDC 是中央（Cz-A1）和左侧的枕部（O1-A1）通道。喜欢时间的结果表明，对于这组刺激图像，男性和女性的行为相似。
