Research On Critical Techniques For Ambulatory Electrocardiogram Signal Analysis Using Deep Learning

The electrocardiogram(ECG)signal is a bioelectrical manifestation of the electrophysiological activity of the heart and represents a critical source of information for diagnosing heart diseases.Ambulatory electrocardiogram(AECG)is an advancement and extension of the standard ECG,which is imperative for the continuous and long-term detection and diagnosis of cardiac electrophysiological rhythms.The complete AECG analysis process is reliant on the seamless cooperation of multiple algorithmic modules.These modules include AECG denoising,AECG QRS wave detection,AECG premature beat detection,and AECG atrial fibrillation detection.Deep learning,a significant branch of machine learning,can automatically learn the complex relationships between features and tasks.It can identify subtle features and potential patterns that may be challenging for human experts to discern,thus improving the accuracy and efficiency of AECG analysis.Drawing on AECG signals as data,utilizing deep learning techniques as a technical means,and incorporating pertinent medical background knowledge,the present study concentrates on the four vital key technologies that underpin the dynamic ECG analysis technology discussed above.The research in this paper has significant theoretical value in denoising AECG signals,QRS waves detection,premature beats detection,and atrial fibrillation detection,and the research findings have important clinical application value in the diagnosis of heart disease.The primary contributions and innovations of this paper are summarized as follows:1.Research on AECG denoising algorithmThe acquisition of AECG signals are under non-clinical constraints and are prone to interference from various types of noise,which can adversely affect subsequent signal analysis and diagnosis.This paper proposes a lightweight deep convolutional network model called ULde-net,which can simultaneously remove three types of noise:Baseline Wander,Muscle Artifact,and Electrode Motion,while retaining the details and characteristic waveforms of the ECG signal.The model employs a combination of conventional convolution,group convolution,and depth convolution,along with additional skip-connections and LMME modules.To improve the model’s performance,knowledge distillation technology is applied to the AECG denoising task,and a two-stage denoising method(Ude-net+DR-net)is designed as the teacher model.The validity of the proposed model is verified under three different types of noise.Experimental results show that the ULde-net model performs well,achieving signal-to-noise ratios of 15.44 dB,13.16 dB,and 15.69 dB,respectively.Furthermore,the introduction of knowledge distillation technology further improves the model’s performance.Finally,the proposed model is subjected to a cross-dataset test using the WEIPA AECG dataset which is collected with the assistance of a cooperative hospital.The good denoising effect and generalization ability of the ULde-net model are qualitatively demonstrated through two aspects:the visual effect of denoising and a comparative analysis of the frequency spectrum.2.Research on AECG QRS wave detection algorithmGiven the diversity and complexity of QRS waves in AECG recordings,this paper proposes the UQRS-net,which is based on the U-net framework,and evaluates its performance on multiple datasets,including the CPSC2019 dataset,the MIT-BIH Arrhythmia database,and the WEIPA AECG dataset.The UQRS-net incorporates key modules,such as the MSFR module at the input stage and the GConv-block at the encoder and decoder.Additionally,a pre-training method based on noise reduction tasks is proposed to enhance the accuracy and generalization ability of the model.The experimental results demonstrate that the pre-trained UQQS-net performs well in terms of accuracy and generalization ability across different datasets.Specifically,in the CPSC2019 hidden test set,the QRSacc and HRacc metrics achieved values of 0.9353 and 0.9606,respectively.In the DS2 test set of the MIT-BIH Arrhythmia database,the F1 score was found to be 99.84%.Furthermore,the performance of the UQRS-net on the hard dynamic ECG dataset outperforms the comparison method,with the F1 score reaching 96.24%.Finally,the average F1 score on the WEIPA AECG dataset was found to be 99.37%.3.Research on AECG premature beats detection algorithmThis present study introduces a dual-input and dual-output deep convolutional network,called USV-net,to address the challenges associated with detecting Supraventricular Premature beats(SPB)and Premature Ventricular Contractions(PVC)in AECG signals,which are commonly affected by complex heartbeats shapes and noise interference.The USV-net combines QRS wave position sequences with AECG signals at the model input and utilizes a plug-and-play MSDC module that is designed based on deformable 2D convolution at the bottom of the model.The model is designed to enhance detection accuracy by generating high-resolution and low-resolution outputs simultaneously during the training phase.The evaluation is conducted using the MIT-BIH Arrhythmia database,CPSC2018 dataset,and WEIPA AECG dataset.In the Patient-Specific mode of the MIT-BIH Arrhythmia database,F1SPB and F1PVC achieved 82.7%and 95.6%,respectively.In the Inter-Patient mode of the CPSC2018 dataset.F1SPB and F1PVC reached 95.21%and 97.37%,respectively.In the cross-dataset test of the WEIPA AECG dataset,F1SPB and F1PVC were 96.25%and 97.94%,respectively.The proposed USV-net shows promising performance in accurately detecting SPB and PVC in AECG signals.4.Research on AECG atrial fibrillation detection algorithmIn this study,an accurate and fast detection method for the atrial fibrillation(AF)in long-term AECG is proposed using a deep learning approach.The method comprises two stages,achieving a balance between operating efficiency and detection accuracy.In the first stage,a lightweight convolution model named RDAF-net is designed for the initial and rough detection,which enables quick location of suspected AF regions and improves the sensitivity to AF while maintaining sufficient accuracy.In the second stage,a ConvTransAF-net model is proposed by combining convolution and self-attention to perform fine detection on the suspected AF segment,determining the precise location of the AF segment.The proposed two-stage AF detection method is evaluated using the Intra-Patient mode of the MIT-BIH Atrial Fibrillation database.The efficacy of the two-stage approach for detecting AF was evaluated in the Intra-Patient mode of the MIT-BIH Atrial Fibrillation dataset.The RDAF-net exhibited a sensitivity of 99.14%and an accuracy of 98.53%,while the ConvTransAF-net showed a specificity of 98.77%and an accuracy of 98.65%.Furthermore,a cross-dataset evaluation was carried out on three distinct cohorts of the WEIPA AECG dataset,utilizing a combination of the MIT-BIH Atrial Fibrillation database and the CPSC2018 dataset.The outcomes of this evaluation confirmed the accuracy of the first stage of detection in identifying persistent AF.In the paroxysmal AF group,the proposed method achieved an average accuracy rate of 99.21%and a Dice coefficient of 97.56%.In the non-AF group,the average specificity was found to be 99.72%.