心血管疾病是一种严重威胁人类的常见病，发病时长往往在分秒之内，除了临床治疗，日常监测也至关重要。在经济和科技的飞速发展的背景下，可穿戴设备与人工智能技术结合在市场上产生了众多用于日常监测的产品。目前临床使用的可穿戴动态心电图机可满足长时间心电信号采集的需求，但无法进行实时监测；商用的可穿戴心电监测设备能够满足佩戴舒适的需求，但是使用的导联数较少，无法达到临床要求的检测精度；再者因为商用可穿戴系统的算力有限，需要将心电数据上传到云端进行处理，这种方式会带来延时，无法实时进行分析；最后，在训练心电分析模型时，目前只有去隐私化的公开数据集可以使用，如何将基于公开数据集的心电分析系统迁移到可穿戴设备上也是一大难题。 本文首先基于公开数据集对心电数据分别进行心拍的分类研究以及心律异常的分类研究。利用卷积神经网络和时序网络中的长短记忆网络的多模型融合实现了高性能的分类模型，同时减少了模型的参数量，提升了模型的推理速度。随后对有优良导电性以及较高生物相容性的MXene电极采集得到的十二导联心电数据进行了降噪、QRS波定位等操作，并且将前述的分类模型迁移到基于MXene电极的可穿戴心电监测设备采集到的十二导联心电数据上，实现了对可穿戴设备采集到的数据的有效分类；此外，为了减少模型参数量压缩模型以及对模型进行可解释性研究，本文使用了全卷积网络自动编码器对心电节拍进行了向量化以及特征提取，对编码后的心拍数据使用一维卷积和长短记忆网络设计了心拍分类模型，对编码后的长序列心电数据迁移自然语言处理领域的文本分类模型设计了可变长的心律异常分类模型，基本可以达到使用原始数据时的分类性能，并减少了不低于60%的数据量和不低于31%的模型参数量。 本文将模型部署到可用于边缘计算的树莓派4B和IOS系统上，实现了实时分析。树莓派上每个心拍的分类时间不超过50ms，每个长序列的心律异常的分类时间不超过90ms，并对代码进行基于树莓派的优化，通过并行化多进程的方式实现了加速。IOS系统上可以通过应用程序的交互性界面获取心拍数据、绘制心拍信号以及实时展示分类结果。
Cardiovascular disease is a common disease that seriously threatens human beings. The onset time is often within minutes and seconds. In addition to clinical treatment, daily monitoring is also crucial. In the context of the rapid development of economy and technology, the combination of wearable devices and artificial intelligence technology has produced many products for daily monitoring in the market. The current clinical wearable Holter monitor can meet the needs of long-term ECG signal acquisition, but cannot perform real-time monitoring. Commercial wearable ECG monitoring equipment can be worn comfortably, but the number of leads used is small and the required detection accuracy is hard to achieve. In addition, because the computing power of the commercial wearable system is limited, ECG data needs to be uploaded to the server for processing, which will bring delays and cannot meet the needs of real-time monitoring. Finally, in training artificial intelligence models, only deprived public datasets can be used. How to migrate the ECG analysis models based on public datasets to wearable devices is also a big problem.In this paper, based on public datasets, the multi-lead ECG data are respectively studied for the classification of cardiac beats and the classification of abnormal cardiac rhythms. The multi-model fusion of the long-short memory network(a kind of time-series network) and the convolutional network realizes a high-performance classification model, while reducing the number of parameters of the model and improving the inference speed of the model. Subsequently, the twelve-lead ECG data collected by MXene electrodes with excellent conductivity and high biocompatibility are subjected to noise reduction, QRS wave localization, etc. On the 12-lead ECG data collected by wearable ECG monitoring equipment, the effective classification of the data is realized. In addition, in order to reduce the amount of model parameters, compress the model and conduct model interpretability research, the fully convolutional network auto-encoder is used to vectorize ECG beats and extract features, and a one-dimensional convolution and long-short memory network is used to design a beat classification model for the encoded heartbeat data. A text classification model in the field of natural language processing is migrated to design a variable-length abnormal heart rhythm classification model, which can achieve similar classification performance when using the original data, while compressing the data volume by no less than 60\% and reducing the number of model parameters by no less than 31\%.Then the model is deployed on edge computing devices like Raspberry Pi 4B and IOS systems, and realizes real-time analysis. The classification time of each heartbeat on the Raspberry Pi does not exceed 50ms, and the classification time of each long sequence of abnormal heart rhythm does not exceed 90ms. The code is optimized on Raspberry Pi, and acceleration is achieved by multi-process parallelism. On the IOS system, the heartbeat data can be obtained through the interactive interface of the application. The heartbeat signal can be drawn, and the classification results can be displayed in realtime.
