Reconstruction of two-dimensional movement trajectories from selected magnetoencephalography cortical currents by combined sparse Bayesian methods
Reconstruction of two-dimensional movement trajectories from selected magnetoencephalography cortical currents by combined sparse Bayesian methods
复制标题
DOI：
10.1016/j.neuroimage.2010.09.057
复制
发表时间：
2011-01-15
期刊：
影响因子：
5.7
通讯作者：
Sato, Masa-aki
中科院分区：
文献类型：
--
作者：
Toda, Akihiro;Imamizu, Hiroshi;Sato, Masa-aki
从非侵入性记录的大脑活动中重建运动是脑机接口（BMI）的一项关键技术。然而，脑电图（EEG）或脑磁图（MEG）不可避免地会记录来自多个皮质区域的信号的混合，因此它不仅不如侵入性方法有效，而且也给融入神经科学知识带来了更大的困难。我们结合了两种稀疏贝叶斯方法来克服这个困难。首先，通过分层贝叶斯 MEG 逆方法以毫米和毫秒的量级估计数千个皮质电流，然后稀疏回归方法自动仅选择相关的皮质电流，通过时间序列的线性加权和精确重建运动。使用组合方法，我们通过移动腕关节来重建食指尖在指向各个方向时的二维轨迹。测试数据集观察到良好的泛化（重建）性能：预测位置和实际位置之间的平均误差为 15 毫米，是所需运动路径长度的 7%。该方法的重建精度明显高于直接使用MEG传感器信号。此外，权重值的空间分布和时间特征表明，初级感觉运动、高级运动和顶叶区域主要有助于预期时间过程的重建。这些结果表明，组合的稀疏贝叶斯方法提供了有效的方法来预测与感觉运动控制直接相关的非侵入性大脑活动的运动轨迹。 (C) 2010 Elsevier Inc. 保留所有权利。
Reconstruction of movements from non-invasively recorded brain activity is a key technology for brain-machine interfaces (BMIs). However, electroencephalography (EEG) or magnetoencephalography (MEG) inevitably records a mixture of signals originating from many cortical regions, and thus it is not only less effective than invasive methods but also poses more difficulty for incorporating neuroscience knowledge. We combined two sparse Bayesian methods to overcome this difficulty. First, thousands of cortical currents were estimated on the order of millimeters and milliseconds by a hierarchical Bayesian MEG inverse method, and then a sparse regression method automatically selected only relevant cortical currents in accurate reconstruction of movements by a linear weighted sum of their time series. Using the combined methods, we reconstructed two-dimensional trajectories of the index fingertip during pointing movements to various directions by moving the wrist joint. A good generalization (reconstruction) performance was observed for test datasets: mean error between the predicted and actual positions was 15 mm, which was 7% of the path length of the required movement. The reconstruction accuracy of the proposed method was significantly higher than directly using MEG sensor signals. Moreover, spatial distribution and temporal characteristics of weight values revealed that the primary sensorimotor, higher motor, and parietal regions mainly contributed to the reconstruction with expected time courses. These results suggest that the combined sparse Bayesian methods provide effective means to predict movement trajectory from non-invasive brain activity directly related to sensorimotor control. (C) 2010 Elsevier Inc. All rights reserved.