基于改进的非负矩阵分解技术的抗运动干扰心电信号感知方法

Continuous electrocardiogram (ECG) monitoring is crucial for effectively preventing and diagnosing cardiovascular diseases. However, existing ECG monitoring methods are limited by their reliance on expensive equipment unavailable to common users, the stringent requirements of the monitoring process, and confined application scenarios, making them insufficient to meet the urgent need for long-term continuous ECG monitoring of the general population in their daily lives. Given these limitations, this study proposes a motion-robust ECG signal sensing method based on modified non-negative matrix factorization (NMF). The basic idea is to leverage a gyroscope embedded into a low-cost wrist-worn wearable to characterize cardiac activities encoded into body vibrations and interpret them to generate fine-grained ECG signals accurately. As eliminating body motion interference is inherently hard, this work innovatively employs modified NMF to tackle the problem; this can effectively handle body motion interference, even if untrained, and extract the cardiogenic body vibrations from noisy gyroscope data. Due to the lack of clear pattern of cardiogenic body vibrations in each cardiac cycles, current cardiac cycle segmentation solutions cannot be applied. Thus, this work deeply analyses the morphological features of cardiogenic body vibrations and utilizes machine learning techniques for the identification of spike points for segmentation. Finally, cycle generative adversarial network (CycleGAN) framework is employed to construct a correlation mapping model between the cardiogenic body vibrations and the ECG signals. With innovative construction, this model can accurate generation of the ECG signals without the need for a huge amount of training data. Extensive experiments with 18 volunteers confirm the effectiveness of the proposed method, with the average amplitude errors of 7.92% and 9.02% for stationary and moving scenarios, respectively. These values fall well within the acceptable range of medical standards for error tolerance of less than 10%.