毕业生知识库

随着计算机技术与医学研究及临床实践的结合日益紧密，医疗大数据呈现出爆发式的发展态势，其重要性和价值日益凸显，已经成为国家重要的基础性战略资源。如何从医疗大数据中提取出具有代表性、区分度以及医学意义的患者表征，是医疗大数据智能化应用的重要前提；如何依据患者表征来确定恰当的相似性度量进而识别出具有相似特征的患者群体，是医疗大数据智能化应用的关键环节。然而，医疗大数据所固有的规模大、维度高、多缺失值、模态多样、扩增迅速和结构复杂的特性，为其分析和利用带来了数据特征缺失、数据动态更新和数据模态协同难题。如何有效处理上述难题，实现高质量的患者表征和高效、精确的患者相似性学习，已经成为医疗大数据研究领域的迫切需求。

图学习模型在聚合邻域节点信息、挖掘医学知识共性结构特征和捕捉跨模态数据语义关联方面优势突出，为应对上述难题提供了切实可行的思路。因此，本文以患者表征学习为核心，以图学习模型和方法作为切入点，聚焦患者相似性学习任务及其衍生的相似患者检索任务、患者数据跨模态检索任务，深入研究面向患者表征的图学习模型与方法。应用任务由分类过渡到检索，适用场景从单模态拓展到多模态。本文的创新性研究成果主要有：

1. 提出了结构感知孪生图神经网络模型(Structure-Aware Siamese Graph neural Networks, SSGNet)。通过电子医疗记录(Electronic Health Records, EHR)挖掘患者表征，进而确定相似患者，可为人工智能辅助诊疗提供重要的证据支持。针对EHR中数据缺失导致患者相似性学习方法性能不可靠的问题，提出将EHR数据高效组织为患者图的形式，设计了图结构自适应低秩优化的孪生属性图神经网络算法，实现患者鲁棒表征与成对相似性的端到端联合学习，缓解EHR中普遍存在的特征缺失对模型性能造成的负面影响。在威斯康星乳腺癌、克利夫兰冠心病两个国际主流的EHR数据集和IgA肾病的自建EHR数据集上的实验结果证实了SSGNet在5个指标上的性能优势，统计检验结果证实了性能提升的显著性，且在数据缺失率为0%-90%时相似性判别准确率无明显下降，证实了SSGNet能有效应对EHR数据缺失。

2. 提出了图指导深度哈希网络模型(Graph-guided Deep Hashing Networks, GDHN)。通过在高维、大规模、异构的EHR数据库中检索与给定患者相似的患者，可为临床诊断提供重要参考。针对EHR数据相似患者检索中存在检索效率低、准确性差的问题，提出将药物类别作为标签构建标签图，通过图卷积标签编码有效挖掘多标签语义信息和关联依赖，设计了多标签图嵌入与患者哈希表征的耦合对比学习算法，实现患者哈希编码信息量和判别力的显著增强。GDHN将高维EHR数据转化为低维二值哈希编码，有效提升了大规模数据情境下的相似患者检索效率，在IgA肾病和MIMIC-III两个数据集上的相似患者检索实验结果证实了多个编码位数下GDHN在3个指标的性能优势，且GDHN能够使与查询更为相似的实例在检索结果中排在更靠前的位置，更具临床应用价值。

3. 提出了引入跨模态注意力的图增强哈希检索模型(Graph Enhanced Hashing with Cross-modal Attention, GEHCA)。通过在图像、文本等跨模态的患者数据进行高效检索，有助于全方位剖析患者病程状态。针对现有跨模态哈希方法多标签信息利用不佳、模态协同建模不足的问题，提出了图信息增强的多标签诊断编码，获得标签级关联的嵌入表征，并利用模态内-模态间关联互补性，设计了图像、文本与多标签嵌入表征间的跨模态注意力语义融合算法，有效缓解多标签矩阵稀疏问题，丰富融合表征的语义信息，提升了哈希编码判别力。在胸部X光成像与诊断报告数据集MIMIC-CXR上与3个经典跨模态哈希方法和4个先进的深度跨模态哈希方法进行了对比实验，实验结果证实了多个编码位数下GEHCA在“X光片检索诊断报告”和“诊断报告检索X光片”两项跨模态检索任务上具有性能优势。

With the increasingly close integration of computer technology with medical research and clinical practice, medical big data has shown an explosive growth trend, highlighting its significance and value as a crucial national strategic resource. Extracting representative, discriminative, and medically meaningful patient representations from medical big data serves as a critical prerequisite for intelligent medical applications. Determining appropriate similarity metrics based on these representations to identify patient groups with similar characteristics emerges as a pivotal step in the intelligent utilization of medical big data. However, the inherent characteristics of medical big data, including its large scale, high dimensionality, numerous missing values, modal diversity, rapid growth, and complex structure, pose challenges in terms of data missing, dynamic data updating, and multi-modal data collaboration. Effectively addressing these challenges to achieve high-quality patient representations and efficient, accurate patient similarity learning has become an urgent need in the field of medical big data research.

Graph learning models have outstanding advantages in aggregating information from neighborhood nodes, mining common structural features of medical knowledge, and capturing semantic associations of cross-modal data, providing practical and feasible insights for addressing the aforementioned challenges. Therefore, this dissertation bases on patient representation learning, using graph learning models and methods as the breakthrough point, and focuses on the tasks of patient similarity learning, similar patient retrieval tasks, and cross-modal retrieval of patient data. The target tasks transition from classification to retrieval, and the applicable scenarios expand from uni-modality to cross-modality. The main contributions of this dissertation are summarized as follows:

1. We propose a novel deep learning framework, Structure-Aware Siamese Graph Neural Networks (SSGNet), for patient similarity learning. This model learns patient representations through Electronic Health Records (EHR) to identify similar patients, providing crucial evidence support for artificial intelligence-assisted diagnosis and treatment. To address the reliability issues in patient similarity learning methods caused by missing data in EHR, we propose the efficient organization of EHR data into a patient graph. We have designed a siamese property graph neural network algorithm, where the graph structure can be optimized adaptively according to the low-rank property, to achieve joint learning of robust patient representations and pairwise similarities in an end-to-end manner. This mitigates the negative impact on model performance caused by missing data in EHR. Experimental results on two internationally recognized EHR datasets, Wisconsin Breast Cancer and Cleveland Coronary Heart Disease, as well as a self-built EHR dataset for IgA Nephropathy, confirm the performance advantages of SSGNet across five metrics. Statistical tests validate the significance of these performance improvements. Notably, there is no significant decrease in Accuracy of the pairwise similarity discrimination when the data missing rate ranges from 0% to 90%, demonstrating the effectiveness of SSGNet in handling missing values in EHR data.

2. We propose a novel deep supervised hashing framework, Graph-guided Deep Hashing Networks (GDHN), for similar patient retrieval. This model retrieves similar patients from high-dimensional, large-scale, heterogeneous EHR databases, providing significant references for clinical diagnosis. Addressing the issues of low retrieval efficiency and accuracy in similar patient retrieval from EHR data, we propose to construct a label graph using medication categories as labels. Through encoding the multi-labels with graph convolutional networks, we effectively mine multi-label semantic information and correlation dependencies. We have designed a contrastive learning algorithm that combines multi-label graph embedding with patient hash representations, achieving more informative and discriminative hash codes of patients. GDHN transforms high-dimensional EHR data into low-dimensional binary hash codes, effectively improving the efficiency of similar patient retrieval in large-scale medical data. Experimental results on similar patient retrieval from two datasets, IgA Nephropathy and MIMIC-III, confirm the performance advantages of GDHN across three metrics at various code lengths. Moreover, GDHN ensures that instances more similar to the query are ranked higher in the retrieval results, enhancing its clinical application value.

3. We propose a novel deep cross-modal patient hashing method, called Graph Enhanced Hashing with Cross-modal Attention (GEHCA). This model enables efficient retrieval across multi-modal medical data, including images and text, providing a comprehensive analysis of the patient's condition. Addressing the limitations of existing cross-modal hashing methods in utilizing multi-label information and inadequate modal collaborative modeling, we propose a graph-enhanced diagnosis multi-label encoder to obtain label-level correlated embeddings. Leveraging both intra-modal and inter-modal correlations and complementarities, we have designed a cross-modal attention semantic fusion algorithm between image, text, and multi-label embeddings. This effectively mitigates the sparsity issue in multi-label matrices, enriches the semantic information of fused representations, and enhances the discriminative capacity of hash codes. Comparative experiments with 3 classic cross-modal methods and 4 state-of-the-art deep learning based cross-modal methods on the MIMIC-CXR dataset, which comprises chest X-ray images and diagnostic reports, demonstrate the performance advantages of GEHCA across multiple code lengths on two cross-modal retrieval tasks, i.e. ``retrieving diagnostic reports using X-ray images'' and ``retrieving X-ray images using diagnostic reports''.