University of Texas, Department of Computer Science, San Antonio, Texas 78249, United States of America. US Army Research Laboratory, Human Research and Engineering Directorate, Aberdeen Proving Ground, Maryland 21287, United States of America.
J Neural Eng. 2018 Dec;15(6):066015. doi: 10.1088/1741-2552/aadc1c. Epub 2018 Aug 22.
EEG spindles, narrow-band oscillatory signal bursts, are widely-studied biomarkers of subject state and neurological function. Most existing methods for spindle detection select algorithm parameters by optimizing agreement with expert labels. We propose a new framework for selecting algorithm parameters based on stability of spindle properties and elucidate the dependence of these properties on parameter selection for several algorithms.
To demonstrate this approach we developed a new algorithm (Spindler) that decomposes the signal using matching pursuit with Gabor atoms and computes the spindles for each point in a fine grid of parameter values. After computing characteristic surfaces as a function of parameters, Spindler selects algorithm parameters based on the stability of characteristic surface geometry.
Spindler performs well relative to several common supervised and unsupervised EEG sleep spindle detection methods. Spindler is available as an open-source MATLAB toolbox (https://github.com/VisLab/EEG-Spindles). In addition to Spindler, the toolbox provides implementations of several other spindle detection algorithms as well as standardized methods for matching ground truth to predictions and a framework for understanding algorithm parameter surfaces.
This work demonstrates that parameter selection based on physical constraints rather than labelled data can provide effective, fully-automated, unsupervised spindle detection. This work also exposes the dangers of applying cross-validation without considering the dependence of spindle properties on parameters. Parameters selected to optimize one performance metric or matching method are not optimized for others. Furthermore, elucidation of the stability of predicted indicators with respect to algorithm parameter selection is critical to practical application of these algorithms.
脑电纺锤波，窄带振荡信号爆发，是广泛研究的主体状态和神经功能生物标志物。大多数现有的纺锤波检测方法通过优化与专家标签的一致性来选择算法参数。我们提出了一种基于纺锤波特性稳定性选择算法参数的新框架，并阐明了这些特性对几种算法参数选择的依赖性。
为了演示这种方法，我们开发了一种新的算法（Spindler），该算法使用匹配追踪与 Gabor 原子分解信号，并在参数值的细网格中计算每个点的纺锤波。在计算特征曲面作为参数函数后，Spindler 根据特征曲面几何形状的稳定性选择算法参数。
Spindler 相对于几种常见的有监督和无监督 EEG 睡眠纺锤波检测方法表现良好。Spindler 可作为一个开源的 MATLAB 工具箱（https://github.com/VisLab/EEG-Spindles）获得。除了 Spindler，该工具箱还提供了几种其他纺锤波检测算法的实现，以及将地面真实值与预测值匹配的标准化方法和理解算法参数曲面的框架。
这项工作表明，基于物理约束而不是标记数据的参数选择可以提供有效的、全自动的、无监督的纺锤波检测。这项工作还暴露了在不考虑纺锤波特性对参数的依赖性的情况下应用交叉验证的危险。为了优化一个性能指标或匹配方法而选择的参数，并不针对其他指标进行优化。此外，阐明预测指标相对于算法参数选择的稳定性对于这些算法的实际应用至关重要。
