A multi-cascade deep learning method for physiological micro-vibration signal enhancement: A case for OSA monitoring

Physiological micro-vibration signals (PMVS) always contain a lot of basic vibration information of organs and tissues, thereby providing favorable judgment information for human physiological index monitoring. Meanwhile, wearable equipment based on flexible sensors, such as polyvinylidene fluoride (PVDF) film, can provide comfortable and long-time wearing experience and has become a hot research topic. In this study, we design a wearable monitoring device based on PVDF film with comfortable wearability and precise measurement to collect PMVS data. However, the collected PMVS data suffers from intense and complicated interference, bringing challenges for the following diagnostic procedure. To improve the quality of the captured data, A PMVS multi-cascade denoising network (PMD-Net), which combines multi-cascade scheme and attention mechanism is proposed. Specifically, the PMD-Net is composed of three cascades to capture informative features from high, middle, and low resolutions. Moreover, self-attention mechanism and spatial attention mechanism are also introduced to refine the obtained features. On this basis, we use synthetic and real monitored data as the analyzed data and compare the results with other popular methods to evaluate its performance. It is shown that PMD-Net represents better denoising performance than the other methods being compared methods and obviously improve the quality of PMVS data. Simultaneously, the monitoring results proposed in this study were compared with those obtained from medical polysomnography (PSG) for the accurate diagnosis of obstructive sleep apnea (OSA). The change characteristics of physiological signals exhibited a high level of consistency, while vibration signals provided more intricate information. Consequently, the flexible wearable monitoring device presented in this paper demonstrates attributes such as comfortable wearability, precise measurement, and comprehensive signal acquisition. Furthermore, its signal processing efficacy has been validated through clinical OSA monitoring using PMD-Net, thereby offering a reliable and feasible approach for advancing human health signal monitoring technology.