Zhang Qingxue, Zhou Dian, Zeng Xuan
Department of Electrical Engineering, University of Texas at Dallas, 800 W Campbell Rd, Richardson, TX 75080, USA.
Physiol Meas. 2016 Nov;37(11):1945-1967. doi: 10.1088/0967-3334/37/11/1945. Epub 2016 Sep 28.
This paper proposes a novel machine learning-enabled framework to robustly monitor the instantaneous heart rate (IHR) from wrist-electrocardiography (ECG) signals continuously and heavily corrupted by random motion artifacts in wearable applications. The framework includes two stages, i.e. heartbeat identification and refinement, respectively. In the first stage, an adaptive threshold-based auto-segmentation approach is proposed to select out heartbeat candidates, including the real heartbeats and large amounts of motion-artifact-induced interferential spikes. Then twenty-six features are extracted for each candidate in time, spatial, frequency and statistical domains, and evaluated by a spare support vector machine (SVM) to select out ten critical features which can effectively reveal residual heartbeat information. Afterwards, an SVM model, created on the training data using the selected feature set, is applied to find high confident heartbeats from a large number of candidates in the testing data. In the second stage, the SVM classification results are further refined by two steps: (1) a rule-based classifier with two attributes named 'continuity check' and 'locality check' for outlier (false positives) removal, and (2) a heartbeat interpolation strategy for missing-heartbeat (false negatives) recovery. The framework is evaluated on a wrist-ECG dataset acquired by a semi-customized platform and also a public dataset. When the signal-to-noise ratio is as low as -7 dB, the mean absolute error of the estimated IHR is 1.4 beats per minute (BPM) and the root mean square error is 6.5 BPM. The proposed framework greatly outperforms well-established approaches, demonstrating that it can effectively identify the heartbeats from ECG signals continuously corrupted by intense motion artifacts and robustly estimate the IHR. This study is expected to contribute to robust long-term wearable IHR monitoring for pervasive heart health and fitness management.
本文提出了一种新颖的机器学习框架，用于在可穿戴应用中，从严重受随机运动伪影干扰的手腕心电图（ECG）信号中持续且稳健地监测瞬时心率（IHR）。该框架包括两个阶段，即心跳识别和细化。在第一阶段，提出了一种基于自适应阈值的自动分割方法来选出心跳候选信号，其中包括真实心跳以及大量由运动伪影引起的干扰尖峰。然后，针对每个候选信号在时间、空间、频率和统计域中提取26个特征，并通过稀疏支持向量机（SVM）进行评估，以选出能够有效揭示残余心跳信息的10个关键特征。之后，使用所选特征集在训练数据上创建的SVM模型，用于从测试数据中的大量候选信号中找出高置信度的心跳。在第二阶段，通过两个步骤进一步细化SVM分类结果：（1）一个基于规则的分类器，具有名为“连续性检查”和“局部性检查”的两个属性，用于去除异常值（误报）；（2）一种心跳插值策略，用于恢复漏检心跳（漏报）。该框架在一个由半定制平台采集的手腕ECG数据集以及一个公共数据集上进行了评估。当信噪比低至 -7 dB时，估计IHR的平均绝对误差为每分钟1.4次心跳（BPM），均方根误差为6.5 BPM。所提出的框架大大优于现有方法，表明它能够有效地从严重受剧烈运动伪影干扰的ECG信号中识别心跳，并稳健地估计IHR。这项研究有望为普及的心脏健康和健身管理中的稳健长期可穿戴IHR监测做出贡献。
