【作者】 杨旭；

【导师】 吴旭；

【作者基本信息】 北京邮电大学 ， 网络与信息安全（专业学位）， 2025， 硕士

【摘要】 医疗健康领域关乎国家发展与社会稳定,互联网技术与医疗健康的深度结合必将产生大规模多源异构数据,这些数据对于医疗健康领域的研究具有重要指导作用。然而,医疗数据的高度隐私性和敏感性限制了研究人员对其访问和使用的权限,导致能够获取的数据规模较小且样本不平衡,难以满足研究需求。因此,如何生成高质量的医疗健康多模态数据从而激活数据的价值,成为当前亟待解决的问题。数据生成技术通过算法或工具来创建模拟真实世界的数据,可以帮助解决数据不足、数据隐私保护等问题,这项技术在大数据时代具有关键意义。针对医疗文本记录具有多样性、超时空性等特征,现有的文本生成模型难以捕捉潜在逻辑关系的问题,提出了融合时空与隐状态转移的医疗文本生成模型。该模型在对抗训练框架下,结合泄露特征优化策略,采用多层LSTM网络并增加隐藏层和嵌入层数量,以增强生成器对长距离依赖关系的捕捉能力。同时,基于马尔可夫链设计了时空数据生成模块,实现对电子病历记录中时空属性的建模与生成。所提模型在NLLoracle、EmbSim等多个文本质量评价指标上取得了提升,其中NLLoracle提高了 2.038。实验结果证明该模型不仅可以解决数据隐私问题和数据样本不足的问题,而且增强了数据的可用性。针对患者拍摄的症状图像缺乏公开数据集且图像质量较低,已有的生成模型难以满足生成需求的问题,研究并提出了基于几何与像素混合增强的医疗图像生成模型。该模型以深度卷积网络和WGAN-DIV为基础,设计了几何增强与像素增强模块,通过几何变换增强图像的多样性,利用像素级优化提升图像的细节表现力,从而显著提高了生成模型的泛化能力和生成图像的视觉质量。实验通过多个图像评价指标对生成结果进行分析,验证了各模块的有效性。结果表明,该模型在FID得分上提升了 33.092,同时在SSIM、PSNR等指标上也优于现有模型。针对现有的医疗健康数据集彼此间缺乏关联性,难以形成用户、内容、时空等多维度属性的立体表示与信息融合的问题,研究并提出了基于图文对齐的多模态电子病历数据生成方法。该方法设计了用户信息生成模块和医疗信息提取模块,并基于PLIP模型生成图像与文本相关联的医疗图文数据,从而构建高质量的多模态电子病历数据集。实验结果表明,生成的数据在内容丰富度与仿真度上表现优异,三级李克特量表平均得分2.515(满分3.000),能够有效支持医疗健康领域的分析与研究。最后,基于上述模型方法实现了医疗健康多模态数据仿真生成系统,系统经过全面测试能够满足医疗领域对高质量仿真数据的需求。数据利用者可通过该系统自动化生成高度逼真的多模态医疗健康数据,为医疗数据分析与安全管理等工作的开展提供了有力的支持。更多还原

【Abstract】 The healthcare field is related to national development and social stability,and the deep integration of Internet technology and healthcare will inevitably generate large-scale multi-source heterogeneous data,which are important guides for research in the healthcare field.However,the high degree of privacy and sensitivity of medical data restricts researchers from accessing and using them,resulting in small and unbalanced samples of data that can be accessed,making it difficult to meet research needs.Therefore,how to generate high-quality healthcare multimodal data so as to activate the value of the data has become an urgent problem to be solved.Data generation technology can help solve the problems of insufficient data and data privacy protection by creating simulated real-world data through algorithms or tools,and this technology is of key significance in the era of big data.Aiming at the problem that medical text records are characterized by diversity and supra-temporality,and existing text generation models are difficult to capture potential logical relationships,a medical text generation model that integrates spatio-temporal and hidden state transfer is proposed.The model incorporates a leakage feature optimization strategy under an adversarial training framework,using a multilayer LSTM network and increasing the number of hidden and embedded layers to enhance the generator’s ability to capture long-distance dependencies.Meanwhile,a spatio-temporal data generation module is designed based on Markov chain to realize the modeling and generation of spatio-temporal attributes in electronic medical record records.The proposed model achieves improvement in several text quality evaluation metrics such as,EmbSim,etc.,in which the NLLoracle is improved by 2.038.The experimental results demonstrate that the model not only solves the data privacy problem and the problem of insufficient data samples,but also enhances the usability of the data.Aiming at the problem that symptom images taken by patients lack public datasets and have low image quality,and existing generation models can hardly meet the generation requirements,an adversarial generation model for medical images based on hybrid geometric and pixel enhancement is researched and proposed.The model is based on deep convolutional network and WGAN-DIV,and the geometric and pixel enhancement modules are designed to enhance the diversity of the images through geometric transformations,and to improve the detailed expression of the images using pixel-level optimization,which significantly improves the generalization ability of the generative model and the visual quality of the generated images.The experiment analyzes the generated results by several image evaluation metrics to verify the effectiveness of each module.The results show that the model improves the FID score by 33.092,and also outperforms the existing models in terms of SSIM and PSNR.Aiming at the problem that existing medical and health datasets lack correlation with each other,and it is difficult to form a three-dimensional representation and information fusion of multi-dimensional attributes such as user,content,and spatio-temporal attributes,a multimodal electronic medical record data generation method based on graphic alignment is researched and proposed.The method designs a user information generation module and a medical information extraction module,and generates medical graphic data based on the PLIP model that associates images with text,thus constructing a high-quality multimodal electronic medical record dataset.The experimental results show that the generated data excel in content richness and simulation,with an average score of 2.515(out of 3.000)on a three-level Likert scale,which can effectively support analysis and research in the field of healthcare.Finally,a healthcare multimodal data simulation generation system is realized based on the above modeling approach,and the system has been fully tested to meet the demand for high-quality simulation data in the medical field.Data users can generate highly realistic multimodal healthcare data through the system,which provides strong support for medical data analysis and safety management.更多还原