High-resolution whole brain recordings have the potential to uncover unknown functionality but also present the challenge of how to find such associations between brain and behavior when presented with a large number of regions and spectral frequencies. In this paper, we propose an exploratory data analysis method that sorts through a massive quantity of multivariate neural recordings to quickly extract a subset of brain regions and frequencies that encode behavior. This approach combines existing tools and exploits low-rank approximation of matrices without a priori selection of regions and frequency bands for analysis. In detail, the spectral content of neural activity across all frequencies of each recording contact is computed and represented as a matrix. Then, the rank-1 approximation of the matrix is computed using singular value decomposition and the associated singular vectors are extracted. The temporal singular vector, which captures the salient features of the spectrogram, is then correlated to the trial-varying behavioral signal. The distribution of correlations for each brain region is efficiently computed and used to find a subset of regions and frequency bands of interest for further examination. As an illustration, we apply this approach to a data set of local field potentials collected using stereoelectroencephalography from a human subject performing a reaching task. Using the proposed procedure, we produced a comprehensive set of brain regions and frequencies related to our specific behavior. We demonstrate how this tool can produce preliminary results that capture neural patterns related to behavior and aid in formulating data-driven hypotheses, hence reducing the time it takes for any scientist to transition from the exploratory to the confirmatory phase.

中文翻译：

---

一种探索性数据分析方法，用于从大规模神经记录中识别感兴趣的大脑区域和频率。

高分辨率的全脑记录具有发现未知功能的潜力，但是当呈现大量区域和频谱频率时，也提出了如何在大脑与行为之间找到这种关联的挑战。在本文中，我们提出了一种探索性数据分析方法，该方法可以对大量的多变量神经记录进行分类，以快速提取一部分大脑区域和频率来编码行为。这种方法结合了现有工具，无需先验即可利用矩阵的低秩逼近选择区域和频带进行分析。详细地，跨每个记录接触的所有频率的神经活动的频谱含量被计算并表示为矩阵。然后，使用奇异值分解计算矩阵的秩1近似，并提取关联的奇异矢量。然后，将捕获频谱图显着特征的时间奇异矢量与不断变化的行为信号相关。有效地计算每个大脑区域的相关性分布，并将其用于查找感兴趣的区域和频带的子集以进行进一步检查。作为说明，我们将此方法应用于使用立体脑电图从执行伸手任务的人类受试者收集的局部场电势数据集。使用建议的程序，我们生成了与我们的特定行为相关的一组全面的大脑区域和频率。我们演示了该工具如何产生初步结果，以捕获与行为相关的神经模式，并帮助制定数据驱动的假设，从而减少任何科学家从探索阶段过渡到确认阶段的时间。