A Study On The Deep Learning-based Models For Intelligent Analysis Of Wearable ECG Signals

A wearable electrocardiogram(ECG)recorder can remotely monitor the ECG of the user,detecting abnormal events in time.After recording ECG signals with a wearable ECG recorder,the user can pay to upload the ECG signals to the cloud platform for medical diagnostic services.The complex interference or strong noise of wearable ECG signals may prevent doctors from obtaining helpful information,wasting the medical expense of the user.The quality assessment of wearable ECG signals on mobile devices can filter out the ECG signals without diagnostic value,reducing medical costs.Additionally,doctors need to take much time to analyze the long-term ECG signals(e.g.,24-hour ECG),which means a huge workload.The heart rate analysis and the automatic recognition of arrhythmia events can assist in ECG diagnosis.Therefore,this thesis aims at studying the deep learning-based models for intelligent analysis of wearable ECG signals.We need to construct two lightweight models for wearable ECG signals,one for the real-time quality assessment and another for the real-time R-peak detection.Besides,a robust model for the automatic diagnosis of atrial fibrillation events is also needed.The main work and innovative achievements of this thesis are as follows:(1)In view of the limited computing resources of mobile devices,a lightweight model was obtained by neural network architecture search to assess the quality of wearable ECG signals in real time.Most of the existing methods for the quality assessment of ECG signals only gave one overall label for a multi-lead ECG signal.Differently,we proposed a novel and effective strategy,any lead,to assess the quality of all leads using a single model,reminding the user to adjust the wearing posture.The proposed method was evaluated on a large-scale(10,709 signals)wearable 12-lead ECG dataset and a public dataset named Physionet Cinc Challenge 2011.The parameters and FLOPs of our model were about 66.76 K and 36.44 M.Our model achieved excellent performance on the datasets above,with AUC of 98.32%and 97.64%,F1 scores of 94.36%and 93.52%,and the inference time on an Android emulator of approximately 78 ms.Extensive experimental results demonstrate the effectiveness of our method in assessing the quality of all leads of ECG signals on mobile devices in real time.(2)To deal with the complex interference and the strong noise of wearable ECG signals,we proposed a heartbeat-aware convolutional neural network to detect R peaks.We applied an encoder-decoder network to predict the R-peak heatmap of ECG data.Then the peaks of the predicted heatmap were located to obtain the R-peak positions.To enhance the model’s capability in extracting the global context,a module,HeartbeatAware(HA),was introduced.The encoder’s output feature map was fed into the HA module to predict the heartbeat number of the signal.The encoder-decoder module and the HA module were trained by multi-task learning.Further,we adopted the lightweight convolution to reduce the model’s parameters and computational complexity.The experiments were conducted on a wearable ECG dataset and a public dataset named LUDB.At a tolerance window interval of 150 ms,the sensitivities of our method reached 100%,and the true positive rates exceeded 99.9%,which significantly overwhelmed many existing algorithms.As for the ECG signal of 10 s duration,our method’s CPU time of R-peak detection was about 23.2 ms,which meets the demand for real-time R-peak detection.The experimental results suggest that the proposed method can perform well on both wearable ECG signals and routine ECG signals for R-peak detection.(3)Considering the weak ability of the convolutional neural network to utilize the time information and the recurrent neural network’s weak ability to extract the morphological features,we used Transformer for the atrial fibrillation recognition of wearable ECG signals(SiT).The experiments were compared SiT with multiple deep neural networks(DNNs)on our wearable ECG dataset.The AP of SiT was 95.68%.The lead-off robustness,noise robustness,and ability in discriminating atrial fibrillation and atrial flutter of SiT all exceeded DNN.Extensive experimental results show that SiT is significantly effective for the atrial fibrillation detection of wearable ECG signals.