Bloorview Research Institute, Holland Bloorview Kids Rehabilitation Hospital, Toronto, Ontario, Canada.
J Neural Eng. 2013 Aug;10(4):046018. doi: 10.1088/1741-2560/10/4/046018. Epub 2013 Jul 18.
Near-infrared spectroscopy (NIRS) is an optical imaging technique that has recently been considered for brain-computer interface (BCI) applications. To date, NIRS-BCI studies have primarily made use of temporal features of brain activity, derived from the time-course of optical signals measured from discrete locations, to differentiate mental states. However, functional brain imaging studies have indicated that the spatial distribution of haemodynamic activity is also rich in information. Thus, the progression of a response over both time and space may be valuable to brain state classification. In this paper, we investigate the implication of including spatiotemporal features in the single-trial classification of haemodynamic events for a two-class problem by exploiting this information from dynamic NIR topograms.
The value of spatiotemporal information was explored through a comparative analysis of four different classification schemes performed on multichannel NIRS data collected from the prefrontal cortex during a mental arithmetic activation task and rest. Employing a linear discriminant classifier, data were analysed using spatiotemporal features, temporal features, and a collective pool of spatiotemporal and temporal features. We also considered a majority vote combination of three classifiers; each established using one of the above feature sets. Lastly, two separate task durations (20 and 10 s) were considered for feature extraction.
With features from the longer task interval, the highest overall classification accuracy was achieved using the majority voting classifier (76.1 ± 8.4%), which was greater than the accuracy obtained using temporal features alone (73.5 ± 8.5%) (F3,144 = 7.04, p = 0.0002). While results from the shorter task duration were lower overall, the classifier employing only spatiotemporal features (with an average accuracy of 67.9 ± 9.3%) achieved a higher average accuracy than the rate obtained using only temporal features (64.4 ± 8.4%) (F3,144 = 18.58, p < 10(-4)).
Collectively, these results suggest that spatiotemporal information can be of value in the analysis of functional NIRS data, and improved classification rates may be obtained in future NIRS-BCI applications with the inclusion of this information.
近红外光谱（NIRS）是一种光成像技术，最近已被用于脑机接口（BCI）应用。迄今为止，NIRS-BCI 研究主要利用从离散位置测量的光信号时程得出的脑活动的时间特征来区分心理状态。然而，功能脑成像研究表明，血液动力学活动的空间分布也富含信息。因此，响应在时间和空间上的演变对于脑状态分类可能是有价值的。在本文中，我们通过利用动态近红外光拓扑图中的这种信息，研究了在单次试验分类中包含时空特征对二分类问题的影响。
通过对在前额叶皮层在心理算术激活任务和休息期间采集的多通道 NIRS 数据进行的四种不同分类方案的比较分析，探讨了时空信息的价值。采用线性判别分类器，使用时空特征、时间特征以及时空和时间特征的集合池对数据进行分析。我们还考虑了使用三种分类器的多数投票组合；每个分类器都是使用上述特征集之一建立的。最后，考虑了两个单独的任务持续时间（20 秒和 10 秒）进行特征提取。
在使用较长任务间隔的特征时，使用多数投票分类器获得的总体分类准确性最高（76.1 ± 8.4%），高于仅使用时间特征获得的准确性（73.5 ± 8.5%）（F3,144 = 7.04，p = 0.0002）。虽然较短任务持续时间的结果总体上较低，但仅使用时空特征的分类器（平均准确率为 67.9 ± 9.3%）的平均准确率高于仅使用时间特征获得的准确率（64.4 ± 8.4%）（F3,144 = 18.58，p < 10(-4)）。
总的来说，这些结果表明，时空信息在功能 NIRS 数据分析中可能具有价值，并且通过包含此信息，未来的 NIRS-BCI 应用可能会获得更高的分类率。
