Research And Application Of Sleep Apnea Detection Based On Ballistocardiogram

Sleep apnea is a common and nonnegligible sleep disorder in contemporary society.It may affect numerous aspects of life and may cause a litany of complications.In clinical medicine,Polysomnography（PSG）is considered the gold standard for the diagnosis of sleep apnea,but it will bring both psychological and physical burden to the patients on account of the complex diagnostic process and it is failed to meet the demand for long-term care of sleep apnea.In contrast,recent studies have shown that detection methods based on a single physiological signal can achieve accurate detection of sleep apnea while reducing the complexity of detection,thus making up for the shortcomings of PSG in long-term monitoring at home,which is of great significance in promoting medical universalization.In this paper,we take the multichannel ballistocardiography（BCG）signal as the object of study,and take the granularity of sleep apnea detection and the realization of terminal lightweight detection as the entry point,and mainly accomplish the following works:First,a multichannel BCG signal sleep apnea dataset was constructed.To address the lack of BCG dataset,this paper used an 18-channel BCG sensor system embedded in a mattress to collect short sleep data from 4 healthy subjects in different sleeping positions and 11 sleep apnea patients throughout the night,and used a 90-second overlapping sliding window to organize the data.Second,a fine-grained sleep apnea detection model based on multi-channel BCG signals,M-Breath,was proposed to address the problems of coarse granularity and low channel utilization of sleep apnea detection,which uses a channel attention mechanism to adjust the contribution weights of different channel BCG signals and fuse the channel data.The fused single-channel data are sequentially passed through a convolutional neural network-based event classifier and a U-Net3+-based one-dimensional semantic segmentation network to finally output the sampling point-level sleep apnea occurrence time period.Then,a lightweight sleep apnea detection algorithm Q-Breath based on multi-scale entropy and peaks is proposed,which is optimized to reduce the memory requirement and improve the detection performance to address the problem that the resources of terminal devices are difficult to meet the deep learning model,and firstly,wavelet decomposition and reconstruction are used to pre-process the data,and then the multi-scale entropy and the number of peaks are calculated and compared with the corresponding threshold to find out whether the current window has apnea.Then,the multi-scale entropy value and the number of peaks are calculated and compared with the corresponding thresholds.Finally,an intelligent sleep apnea monitoring mini-program is implemented.The miniprogram can visualize the BCG signal during sleep and evaluate the sleep apnea status,and also has the ability to output professional sleep apnea evaluation reports to meet the needs of both daily home monitoring and professional doctors’ diagnosis.In summary,this paper proposes a fine-grained sleep apnea detection model and a lightweight detection method based on multichannel BCG signals.The former can achieve sleep apnea segmentation at the sampling point level,while the finer-grained detection results are more informative for clinical diagnosis;the latter provides a new idea for accurate detection of sleep apnea events on resource-limited terminal devices and real-time detection tasks.