【作者】 刘毅；

【导师】 李灯熬；

【作者基本信息】 太原理工大学 ， 电子科学与技术， 2024， 博士

【摘要】 心力衰竭(心衰)全球患者人数已达6400万,并且随着人口老龄化进程的加快,心衰的患病率持续上升,已成为主要的死亡原因之一,给个体和社会都带来了巨大的负担。尤其是在近五年来的大流行期间,病毒感染引发了一系列心血管疾病的并发症,并进一步推动了心衰的发生。心衰患者因感染面临更高的死亡风险。此外,从大流行中恢复的患者中,心脏结构变化的发病率也有所增加。因此准确的检测心衰患者并对其细致的管理显得尤为重要。为了应对这一挑战,本文致力于研究基于不同来源多种模态医学数据的创新诊断方法,研究通过整合心电图、文本数据和胸片三种在心衰诊断中常用的数据类型,探索其在心衰辅助诊断中的应用潜力。本文围绕心衰患者及其在病毒感染期间的护理,研究了心电图模态融合理论、文本特征挖掘方法、胸片特征集成技术和多源数据融合建模的心衰辅助诊断方法,研究不同来源多种模态数据的功能特点、特征提取、模型构建和分析方法。主要研究工作包括:(1)针对心电图中不同疾病在时域和频域的关注点差异,本文提出了一种结合时频特性的双分支并行注意力融合方法,以有效地整合时域和频域信息。在时域模态方面,设计了卷积神经网络和双向LSTM的特征提取方案;在频域模态方面,分析了频域双模型及双模态的融合方法。然后结合心电图时域和频域的特性,设计了自注意力机制以优化心电图特征融合。通过多个实验的分析对比,研究了不同组合模型的优势,实验结果表明注意力机制的时频融合方法具有最优的效果,能够有效捕捉心电图中的关键特征,且在心衰二分类任务中表现出良好性能。该方法可适用于心内科分诊,随着可穿戴设备的普及,有望提高心衰早期诊断的效率和准确性。(2)针对心衰在不同人口群体中的流行病学尚未明确的问题,本文基于公开数据集设计了一套由七个步骤组成的技术路径,以分析共病和生物标志物在不同患者群体中的作用。首先通过单因素分析发现,心衰差异可能被人群因素掩盖,将患者分为心衰组与非心衰组,结果显示血尿素氮等因素均存在显著差异,进一步的研究发现心衰在组别3中具有显著性,接着在生存曲线中揭示了不同组别的生存率差异。最后对显著因素进行了共同特征和总和特征归纳,结合人工神经网络进行分析,探索了不同组合特征对准确率产生的影响。研究发现了组别间的特征差异,为个性化诊断和治疗提供了指导。(3)针对当前方法在胸片局部到全局纹理观察能力上的不足,本文提出了一种结合局部和全局特征优势的Trans Convnet建模方法,以提升病变诊断能力。模型包含对四个模块进行多层次的特征集成,通过多种模态对比分析、预训练影响分析、冻结性能影响分析、不同配置对比分析以及泛化与可视化分析等方法开展研究,结果显示所提出的模型可以有效提取和整合胸片中的重要纹理信息,提升了病变识别的精度。此外,该模型在心衰的二分类中取得了99%的准确性,有助于心衰的临床决策和治疗优化。(4)针对心衰多源数据融合中模态不足的问题,本文结合心电图、文本和胸片数据进行实验设计,模拟临床医生的决策流程,展示了数据融合的显著优势。本文结合三种大批量数据的研究经验,通过整合心电图、文本和胸片数据开展实验,为实现不同来源数据的有效融合,设计了三种数据的特征提取建模方案,并进行了特征级融合和分数级融合策略的实验验证。结果显示,多源数据的融合显著提高了诊断准确性,尤其在应对噪声干扰和数据增强方面表现出色。随后,本文设计了结合两种策略特点的FASMNet融合模型,并进一步探索了不同组合的影响。通过所提出的方法,展现了多源数据在辅助诊断中的应用潜力。综上所述,本文对多种来源和多模态医学数据在心衰辅助诊断中的应用进行了深入研究。同时,本文也为将这些方法推广到其它复杂疾病的诊断提供了新的可能性。这项研究为医学数据分析开辟了新的思路,有助于推动医学模型的构建和辅助诊断技术的进一步发展。更多还原

【Abstract】 Heart failure(HF)affects approximately 64 million people worldwide,and its prevalence continues to rise with the acceleration of population aging,making it one of the leading causes of death and imposing a significant burden on individuals and society.Particularly during the past five years of the pandemic,viral infections have triggered a series of cardiovascular complications,further promoting the onset of heart failure.HF patients face higher mortality risks due to infections.Moreover,there is an increased incidence of cardiac structural changes among patients recovering from the pandemic.Therefore,accurate detection and meticulous management of HF patients are crucial.To address this challenge,this dissertation explores innovative diagnostic methods based on multi-modal medical data from different sources,investigating the potential of integrating electrocardiograms(ECG),textual data,and chest X-rays—three commonly used data types in HF diagnosis—for their application in auxiliary HF diagnosis.This research focuses on HF patients and their care during viral infections,studying ECG modality fusion theories,text feature extraction methods,chest X-ray feature integration techniques,and multi-source data fusion modeling for HF auxiliary diagnosis.It explores the functional characteristics,feature extraction,model construction,and analysis methods of multi-modal data from various sources.The main research work includes:(1)Addressing the differences in time-domain and frequency-domain focus for various diseases in ECG,this dissertation proposes a dual-branch parallel attention fusion method that combines time-frequency characteristics to effectively integrate temporal and frequency domain information.In the temporal domain,a feature extraction scheme using convolutional neural networks and bidirectional LSTM is designed.In the frequency domain,the dual-model and dual-modality fusion methods are analyzed.A self-attention mechanism is then incorporated to optimize ECG feature fusion based on temporal and frequency domain characteristics.Through multiple experiments and comparative analyses,the advantages of different combined models are examined.The results demonstrate that the attention mechanism-based time-frequency fusion method yields optimal results,effectively capturing key features in ECG and showing excellent performance in the binary classification task of heart failure.This method can be applied in cardiology triage and,with the proliferation of wearable devices,is expected to enhance the efficiency and accuracy of early HF diagnosis.(2)To address the unclear epidemiology of HF across different demographic groups,this dissertation designs a seven-step technical pathway based on publicly available datasets to analyze the roles of comorbidities and biomarkers in various patient populations.Initial univariate analysis reveals that population factors may obscure HF differences,dividing patients into HF and non-HF groups,with significant differences observed in factors like blood urea nitrogen.Further research identifies significant HF presence in group 3,and survival curves reveal survival rate differences among groups.Finally,significant factors are summarized into common and total features,analyzed with artificial neural networks to explore the impact of different feature combinations on accuracy.The study identifies characteristic differences among groups,providing guidance for personalized diagnosis and treatment.(3)Addressing the limitations of current methods in observing local to global textures in chest X-rays,this dissertation proposes a Trans Conv Net modeling method that combines the advantages of local and global features to enhance lesion diagnostic capabilities.The model incorporates multi-level feature integration across four modules,with studies conducted through multi-modal comparative analysis,pre-training impact analysis,frozen performance impact analysis,different configuration comparisons,and generalization and visualization analyses.Results indicate that the proposed model effectively extracts and integrates important texture information from chest X-rays,improving lesion recognition accuracy.Additionally,the model achieves 99%accuracy in HF binary classification,aiding clinical decision-making and treatment optimization for HF.(4)To tackle the issue of insufficient modalities in HF multi-source data fusion,this dissertation designs experiments combining ECG,textual,and chest X-ray data to simulate clinicians’decision-making processes,demonstrating the significant advantages of data fusion.Leveraging extensive research experience with large-scale datasets,experiments are conducted by integrating ECG,text,and chest X-ray data.For effective fusion of data from different sources,feature extraction and modeling schemes for each data type are designed,with experiments validating feature-level and score-level fusion strategies.The results show that multi-source data fusion significantly improves diagnostic accuracy,particularly excelling in handling noise interference and data augmentation.Subsequently,a FASMNet fusion model is developed,combining the characteristics of both strategies,and further exploring the effects of different combinations.The proposed methods showcase the potential of multi-source data in aiding diagnostic processes.In conclusion,this dissertation provides an in-depth exploration of the application of multi-source and multi-modal medical data in HF auxiliary diagnosis.It also opens new possibilities for extending these methods to the diagnosis of other complex diseases.This research paves new avenues for medical data analysis,contributing to the advancement of medical model construction and the development of auxiliary diagnostic technologies.更多还原