College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, China.
College of Medicine and Biological Information Engineering, Eindhoven University of technology, Eindhoven, The Netherlands.
Biomed Eng Online. 2021 Jan 28;20(1):12. doi: 10.1186/s12938-021-00848-w.
Atrial fibrillation (AF) represents the most common arrhythmia worldwide, related to increased risk of ischemic stroke or systemic embolism. It is critical to screen and diagnose AF for the benefits of better cardiovascular health in lifetime. The ECG-based AF detection, the gold standard in clinical care, has been restricted by the need to attach electrodes on the body surface. Recently, ballistocardiogram (BCG) has been investigated for AF diagnosis, which is an unobstructive and convenient technique to monitor heart activity in daily life. However, here is a lack of high-dimension representation and deep learning analysis of BCG.
Therefore, this paper proposes an attention-based multi-scale features fusion method by using BCG signal. The 1-D morphology feature extracted from Bi-LSTM network and 2-D rhythm feature extracted from reconstructed phase space are integrated by means of CNN network to improve the robustness of AF detection. To the best of our knowledge, this is the first study where the phase space trajectory of BCG is conducted.
2000 segments (AF and NAF) of BCG signals were collected from 59 volunteers suffering from paroxysmal AF in this survey. Compared to the classical time and frequency features and the state-of-the-art energy features with the popular machine learning classifiers, AF detection performance of the proposed method is superior, which has 0.947 accuracy, 0.935 specificity, 0.959 sensitivity, and 0.937 precision, for the same BCG dataset. The experimental results show that combined feature could excavate more potential characteristics, and the attention mechanism could enhance the pertinence for AF recognition.
The proposed method can provide an innovative solution to capture the diverse scale descriptions of BCG and explore ways to involve the deep learning method to accurately screen AF in routine life.
心房颤动（AF）是全球最常见的心律失常，与缺血性卒中和全身性栓塞的风险增加有关。筛查和诊断 AF 对改善终生心血管健康至关重要。基于心电图的 AF 检测是临床护理的金标准，但受到需要在体表贴电极的限制。最近，弹道心动描记术（BCG）已被用于 AF 诊断研究，这是一种在日常生活中监测心脏活动的非侵入性和方便的技术。然而，BCG 的高维表示和深度学习分析还很缺乏。
因此，本文提出了一种基于注意力的多尺度特征融合方法，用于 BCG 信号。通过 CNN 网络，将从 Bi-LSTM 网络提取的 1-D 形态特征和从重构相空间提取的 2-D 节律特征进行融合，以提高 AF 检测的鲁棒性。据我们所知，这是首次对 BCG 的相空间轨迹进行研究。
在这项研究中，从 59 名阵发性 AF 患者中采集了 2000 段 BCG 信号（AF 和 NAF）。与经典的时频特征以及最先进的能量特征和流行的机器学习分类器相比，该方法的 AF 检测性能更优，在相同的 BCG 数据集上，其准确率为 0.947，特异性为 0.935，灵敏度为 0.959，精确率为 0.937。实验结果表明，组合特征可以挖掘更多潜在特征，注意力机制可以增强对 AF 识别的针对性。
该方法可以为捕捉 BCG 的多种尺度描述提供一种创新的解决方案，并探索将深度学习方法引入到日常生活中准确筛选 AF 的方法。
