【作者】 张鑫；

【导师】 陈超敏；

【作者基本信息】 南方医科大学 ， 生物医学工程， 2025， 硕士

【摘要】 心电图(Electrocardiogram,ECG)在心脏疾病的临床诊断中至关重要,但在实际应用时仍面临着诸多挑战。首先,临床心电数据量庞大,尤其是在动态心电图等长时间连续监测场景下,单纯依赖人工视觉判读不仅效率低下,还可能导致短暂性心律失常等异常信号的漏诊。其次,由于心电信号存在显著个体差异且病理特征复杂,不同医生对相同心电信号的诊断可能会存在主观性差异。最后,心电信号的微弱特性使其易受噪声干扰,这进一步增加了诊断复杂性。近年来,人工智能技术在心电信号分析领域展现出巨大的应用潜力。相较于机器学习在特征设计和分类器选择上的严重人工依赖性,深度学习凭借其强大的自动特征提取能力和端到端的学习模式呈现出显著优势。因此,开发基于深度学习的心电信号智能分类与降噪系统具有重要的临床价值和研究意义。本文两项研究工作主要如下:(1)基于多尺度卷积和位置编码注意力的心电信号分类研究。本研究提出了一种融合多尺度卷积神经网络与改进Transformer编码器的混合架构心电信号分类模型。该模型具有以下创新性设计:首先,研究构建了一种特殊的三层多分支卷积结构,采用不同尺度的卷积核并行提取心电信号的关键局部波形特征。其次,Transformer编码器同时整合了绝对位置编码与相对位置编码,以增强模型对心电信号时序依赖关系的建模能力。其中,相对位置信息被偏置嵌入概率稀疏注意力结构中,在降低传统注意力机制的计算复杂度的同时进一步提高了分类性能。最后,对编码后的多尺度特征进行全局均值池化并输入全连接层以实现最终的心电信号分类。(2)基于去噪自编码器和生成对抗网络的心电信号去噪研究。针对传统信号处理技术在消除心电信号非平稳噪声方面的局限性,本研究提出了一种ATDAE-GAN组合去噪模型,它由改进的编码器、解码器和判别器三个部分组成。首先,编码器中创新性地引入了可学习的阈值化降噪层,通过反向传播自动优化阈值参数,在编码后的特征空间实现噪声抑制。其次,解码器采用对称的转置卷积结构以逐步重建去噪后的心电信号,同时采用跳跃连接保留信号细节。最后,线性判别器对重建信号和干净信号进行差异评估,并将损失反馈至解码器,以进一步提高去噪信号质量。两项研究均基于美国麻省理工心律失常数据库开展实验验证。在心电信号分类研究中,提出的MSFT分类模型通过5折交叉验证取得优异性能:准确率、精确度、灵敏度和F1评分均达到99.40%(±0.05%),特异度达到99.85%(±0.02%),显著优于现有方法。在去噪研究中,模型在12种噪声背景测试集下均表现出色,定性评估显示去噪后的关键波形特征保持完整,定量评估整体性能优于传统方法和其他深度学习去噪技术。更多还原

【Abstract】 The electrocardiogram(ECG)plays a crucial role in the clinical diagnosis of cardiac diseases,yet its practical application faces several challenges.First,the vast volume of clinical ECG data—particularly in long-term continuous monitoring scenarios such as ambulatory ECG—renders manual visual interpretation inefficient and prone to missing transient arrhythmias and other abnormalities.Second,due to significant inter-individual variability in ECG signals and the complexity of pathological features,diagnostic subjectivity may arise among different physicians analyzing the same ECG recording.Third,the low-amplitude nature of ECG signals makes them susceptible to noise interference,further complicating diagnosis.In recent years,artificial intelligence(AI)has demonstrated substantial potential in ECG signal analysis.Compared to traditional machine learning methods,which heavily rely on manual feature engineering and classifier selection,deep learning offers notable advantages through its powerful automatic feature extraction and end-to-end learning framework.Therefore,developing intelligent ECG classification and denoising systems based on deep learning holds significant clinical and research value.The two main research contributions of this work are as follows:(1)ECG classification based on multi-scale convolution and positional encoding attention.This study proposes a hybrid architecture for ECG signal classification that integrates multi-scale convolutional neural networks with an improved Transformer encoder.The model incorporates the following innovative designs:First,a specialized three-layer multi-branch convolutional structure was constructed,employing parallel convolutional kernels of varying scales to extract key local waveform features from ECG signals.Second,the Transformer encoder simultaneously incorporates both absolute and relative positional encodings to enhance the model’s ability to capture temporal dependencies in ECG signals.Notably,the relative positional information is embedded with bias in a probabilistic sparse attention structure,which reduces the computational complexity of traditional attention mechanisms while further improving classification performance.Finally,global average pooling is applied to the encoded multi-scale features before feeding them into a fully connected layer for final ECG signal classification.(2)To address the limitations of conventional signal processing techniques in eliminating non-stationary noise from ECG signals,this study proposes a novel ATDAE-GAN hybrid denoising model comprising three key components:an enhanced encoder,decoder,and discriminator.First,the encoder innovatively incorporates a learnable thresholding denoising layer,where threshold parameters are automatically optimized through backpropagation to achieve noise suppression in the encoded feature space.Second,the decoder employs a symmetrical transposed convolutional architecture to progressively reconstruct denoised ECG signals,while skip connections are adopted to preserve critical signal details.Finally,a linear discriminator evaluates the discrepancy between reconstructed and clean signals,with the resulting loss being backpropagated to the decoder to further enhance denoising quality.Both studies were experimentally validated using the MIT-BIH Arrhythmia Database.In the ECG classification research,the proposed MSFT classification model achieved outstanding performance through 5-fold cross-validation,with accuracy,precision,sensitivity and F1-score all reaching 99.40%(±0.05%),and specificity attaining 99.85%(±0.02%),significantly outperforming existing methods.For the denoising study,the model demonstrated excellent performance across 12 different noise-condition test sets.Qualitative evaluation confirmed the preservation of key waveform characteristics after denoising,while quantitative assessment showed superior overall performance compared to both traditional methods and other deep learning-based denoising techniques.更多还原