Machine Assisted Analysis of Biosignals

生物相关的时间序列数据通常由包含与决策相关的高度信息的局部部分组成，这些信息隐藏在具有无信息信号数据的区域之间。我感兴趣的是提取信号的重要部分，使用信息度量对这些区域的重要性进行评分，以便让人类决策者了解与其决策相关的信息在数据中的位置。我工作的总体目标是开发工具来帮助人类用户理解他们的数据，并确定数据的哪些部分包含重要信息。这类数据描述了许多有趣的问题，包括我目前正在探索的问题：肌电图疾病表征，情绪调节的电生理标志物，动态姿势状态风险表征和人类驾驶员性能评估。关联挖掘是机器学习的子领域，处理识别模式及其相关信息内容，在这些基于风险的表征中特别感兴趣，因为关联挖掘提供了一种将统计置信度与机器学习探索生成的规则相关联的方法。这些规则又提供了一种捕获使用波形模板定义的特定信号模式的描述的方法。在典型的最大似然模型中，唯一的描述可能以“最佳”表征或建议的行动过程的形式出现，与此相反，信息度量和基于规则的模式的组合允许我提供一个数据景观，允许交互式探索来评估不同潜在表征的相对支持和权衡。我对完整决策探索系统背景下的这项工作感兴趣：从原始测量的收集到可视化的决策支持，提供一个从数据到最终人类洞察的相关信息的分析链。 在短期内，这项研究将提高对感兴趣数据的理解：姿势状态，基于EMG的肌肉表征，驾驶员注意力以及以生物测量的时间序列数据为特征的其他领域。从长远来看，通过基于模板的分析，适用于更一般的时态数据分析的工具将有助于理解任何领域的类似数据。影响：更好、更深入地理解时态数据的影响。

Biologically related time-series data frequently is made up of localized sections containing information with a high degree of information pertinent to decision making hidden among regions with uninformative signal data. I am interested in extracting the important portions of the signal, scoring these regions with their importance using information measures, to allow a human decision maker insight into where within their data the information associated with their decision rests. The overall objective of my work is to produce tools to assist human users in understanding their data, and identify which parts of the data contain the important information. Many interesting problems are described by data of this type, including those I am exploring currently: electromyographic disease characterization, electrophysiological markers for emotional regulation, dynamic postural state risk characterization, and human driver performance estimation. Association mining, the sub-field of machine learning dealing with identifying patterns and their associated information content, is particularly of interest in these risk-based characterizations, as association mining provides a means to associate statistical confidence with rules generated by machine learning exploration. These rules in turn provide a means of capturing the description of a particular signal pattern defined using a waveform template. In contrast to a typical maximum likelihood model, in which the only description may come in the form of a "best" characterization or suggested course of action, the combination of information metrics and rule based patterns allows me to provide a data landscape that allows interactive exploration to assess the relative support and trade offs of different potential characterizations. I am interested in this work in the context of full decision exploration systems: from the collection of raw measures through to visualized decision support, providing a soup-to-nuts analysis chain that carries the relevant information from data through to the final human insight. In the short term, this research will yield improved understanding of the data of interest: postural state, EMG based muscle characterization, driver attention, and other domains characterized by time-series data of biological measures. In the longer term, tools applicable to more general analysis of temporal data through template based analysis will aid the understanding of similar data in any field with this problem type. The impact: better and deeper understanding of the impact of temporal data.