Biomedical signal quality analysis

我们目前正在经历数据的爆炸式增长。这种增长包括生物医学数据（例如，心电图（ECG）、肌电图（EMG）、脉搏血氧仪、血压），这些数据提供了关于身体状态和功能的有价值的信息，并在各种应用（例如，健康/保健、生物识别、游戏和运动/健身）中很有用。关于生物医学信号的获取，以及提取有用信息的信号处理方法，存在着一个庞大且不断增长的知识体系。然而，由于噪声、伪影和测量设置误差，生物医学信号可能受到污染；在无人监督的机构（例如远程保健）中尤其如此，因为那里没有训练有素的操作人员。记录中的污染物可能导致误解、不准确和错误，包括误诊。尽管生物医学仪器取得了进步，但污染物经常出现在记录中。目前，生物医学信号质量分析依赖于人类专家。这既耗时又昂贵，而且容易出现人为错误。此外，由于数据量大，普遍、连续和/或多通道监测的增加使得手动或半自动生物医学信号质量分析方法变得不切实际。

We are currently experiencing an explosive growth in data. This growth includes biomedical data (e.g., electrocardiogram (ECG), electromyogram (EMG), pulse oximetry, blood pressure) which provide valuable information regarding the status and function of the body and are useful in a variety of applications (e.g., health/wellness, biometrics, gaming, and sports/fitness). There exists a large and continually growing body of knowledge regarding the acquisition of biomedical signals, as well as signal processing methods to extract useful information. Biomedical signals, however, can be contaminated due to noise, artifacts, and measurement setup errors; this is particularly true in unsupervised setups (e.g., telehealth), where highly trained operators are not present. Contaminants in the recordings can lead to misinterpretations, inaccuracies, and errors, including misdiagnoses. Despite advances in biomedical instrumentation, contaminants are frequently present in recordings. Currently, biomedical signal quality analysis relies on human experts. This is time-consuming, costly, and prone to human error. In addition, the increases in pervasive, continuous and/or multi-channel monitoring are making manual or semi-automated biomedical signal quality analysis methods impractical due to the amount of data. The objective of this research is to develop novel automated biomedical signal quality analysis methods to detect, identify, quantify, and mitigate contaminants. The proposed research is organized into three main research themes: 1) Multi-scale analysis, 2) Multi-variate analysis, and 3) Pattern recognition. Multi-scale approaches are well-suited to signals that arise from complex interconnected systems, such as biological systems. Recent research indicates strong potential in this approach, compared to conventional approaches that are either time or frequency based. Multi-variate approaches take advantage of redundant and complementary information within multi-channel recordings (i.e., multiple leads for the same signal type) and/or multi-modal recordings (i.e., recordings of different signal types). Pattern recognition methods can be employed to discover and leverage trends within the data; this can be used to detect and identify contaminants in biomedical signals, as well as classify the quality of data (e.g., excellent, good, poor, unacceptable). Methods will be evaluated in terms of performance (e.g., correctly detecting and identifying contaminations) and generalizability (e.g., methods work for various contaminants and combinations of contaminants). The exponential growth in biomedical data is associated with various challenges (e.g., acquisition, transferring, storage, and visualization). This research tackles a key, under-researched, area of quality analysis. There is utility within the large datasets being developed, but the capacity to discern which data has adequate quality, and avoid overly contaminated data, is essential. Outcomes of the proposed research will provide engineering contributions in signal processing and data quality analysis, and in the long-term be applied in the context of biomedical signal instrumentation and measurement. For example, automatic biomedical signal quality analysis methods will enable acquisition setups to be validated, alerting operators of issues and directing them on how to resolve these issues. It will also improve the performance of signal processing methods that extract information from these signals (e.g., increase accuracy of clinical decision support systems, reduction of false alarms). While this research is focused on biomedical data, the concepts and frameworks developed in this research are applicable to other data types.