Independent component analysis based algorithms for high-density electromyogram decomposition: Systematic evaluation through simulation
Independent component analysis based algorithms for high-density electromyogram decomposition: Systematic evaluation through simulation
复制标题
DOI：
10.1016/j.compbiomed.2019.04.033
复制
发表时间：
2019-06-01
影响因子：
7.7
通讯作者：
Hu, Xiaogang
中科院分区：
文献类型：
--
作者：
Dai, Chenyun;Hu, Xiaogang
运动单位活动为神经肌肉控制的不同方面提供了重要的理论和临床见解。基于高密度肌电图 (HD EMG) 记录，我们系统地评估了三种基于独立分量分析 (ICA) 的肌电分解算法（Infomax、FastICA 和 RobustlCA）的性能。这些算法在具有一系列肌肉收缩水平和一系列信号质量的模拟高清肌电信号上进行了测试。我们的结果表明，所有三种算法都可以输出准确的（85%-100%）运动单位放电时序。具体而言，在各种信号条件下，尤其是在信号质量较低和收缩水平不同的情况下，RobustlCA 在三种算法中始终表现出较高的分解精度。但在高收缩水平下，RobustlCA 的分解产率往往较低。相比之下，FastICA 的准确度往往最低，但可以检测到最多数量的运动单位，尤其是在高收缩水平下。我们的结果还表明，FastICA 和 RobustlCA 的计算时间相似，比 Infomax 短。此外，每种算法的准确性与轮廓距离度量的聚类指数适度相关，并且与算法对的一致率密切相关。总的来说，我们的研究结果为根据对分解精度/产量有不同要求的特定应用选择特定分解算法提供了指导。
Motor unit activities provide important theoretical and clinical insights regarding different aspects of neuromuscular control. Based on high-density electromyogram (HD EMG) recordings, we systematically evaluated the performance of three independent component analysis (ICA)-based EMG decomposition algorithms (Infomax, FastICA and RobustlCA). The algorithms were tested on simulated HD EMG signals with a range of muscle contraction levels and with a range of signal quality. Our results showed that all the three algorithms can output accurate (85%-100%) motor unit discharge timings. Specifically, the RobustlCA consistently showed high decomposition accuracy among the three algorithms under a variety of signal conditions, especially with a low signal quality and varying contraction levels. But the yield of decomposition of RobustlCA tended to be low at high contraction levels. In contrast, FastICA tended to show the lowest accuracy, but can detect the largest number of motor units, especially at high contraction levels. Our results also showed that the computation time was similar for FastICA and RobustlCA, which was shorter than Infomax. Additionally, the accuracy of each algorithm correlated moderately with the clustering index the silhouette distance measure, and correlated strongly with the rate of agreement of the algorithm pairs. Overall, our findings provide guidance on selecting particular decomposition algorithms based on specific applications with different requirement on the accuracy/yield of the decomposition.