面向临床医护场景基于先验机制与深度学习双模型驱动的无创连续血压精准估计方法的研究

Continuous measurement of each patient's arterial blood pressure is an important basis on the patients’ real diagnosis for the medical staff. While the current continuous blood pressure estimation method is still not approved by the clinic due to the not satisfied precision. It is mainly because the complex blood pressure regulation mechanism. Therefore, the existing model cannot cover such many influencing factors, and it is difficult for the traditional mathematical model to realize the sufficient approximation of multiple nonlinear complex functions. Pre-experiment (cold pressor test) confirmed the dynamic association between blood pressure and hemodynamic parameters, while hemodynamic parameters can be non-invasively obtained by thoracic electrical bioimpedance. Therefore, this project further proposed that integration of vascular elastic cavity theory and hemodynamic characteristics to construct the blood pressure mechanism model, while using the deep learning for the feature learning from the multi-modal information. Next, combine the discriminant mechanism characteristics and network characteristics, hoping to solve the problem of incomplete extraction of artificial features and the lack of potential mechanism of computer model. Our project plans to utilize the dual drive mode of data + knowledge to improve the precise. Finally, noninvasive, continuous and precise blood pressure estimation theory will be constructed. It not only provides the new ways and support for the current clinical disease diagnosis, but also can promote the new development of clinical medicine.

连续测量病人每个心动周期的动脉血压，是医护人员对病人进行实时诊断的重要依据。而目前的无袖带连续血压估计方法一直面临精度不高无法走入临床的问题。究其根本在于血压调节机制复杂，现有模型无法涵盖多种影响因素，并且传统数学模型很难实现多元非线性复杂函数的充分逼近。预实验（冷加压实验）证实了血压与血流动力学参数的动态关联，而血流动力学参数可以通过胸阻抗分析无创获取。因此本项目进一步提出，融合血管弹性腔理论及血流动力学特性构建血压机理模型，同时利用深度学习技术从多模态信息中无监督学习特征，将判别性机理特征与经验学习特征进行有效结合，希望解决人工特征提取的不完备性以及计算机模型缺少潜在机制的问题。利用数据+知识的双驱动模式实现连续无创血压估计的精准性，从而构建适合临床应用的无创、连续、精准的血压估计的新理论与方法，不仅可以为当前临床疾病诊断提供新的方法支持，而且可以推动临床医学的新发展。

连续测量病人每个心动周期的动脉血压，是医护人员对病人进行实时诊断的重要依据。而目前的无袖带连续血压估计方法一直面临精度不高无法走入临床的问题。本项目针对目前无创连续血压估计精度低、影响机制复杂难建模的问题，从数据模型和机制分析两个方面展开深入研究。首先，提出了基于深度学习网络提出了融合心电、脉搏波波形信息，人工特征以及个体信息进行连续无创血压的建模方法，这是利用深度学习自动学习的关联特征在血压估计方面的首次尝试，并且取得了较好的血压预测精度。进一步，为了解决模型部分依赖于人工特征，并且血压数据标签难获取的问题，提出了一种基于卷积自编码器的连续血压计算方法，实现了特征的自动提取。其次，针对血压影响机制复杂的问题，提出融合血管弹性腔理论及血流动力学特性构建血压机理模型，构建了一种个体化的0-1维循环系统仿真模型，解析了血压与血流动力学参数之间的关系，为血压的机制建模提供理论基础。最后，结合影响血压的机制因素，建立了机制特征与机器学习结合的无创连续血压估计方法，根据先验知识从体表信号中提取血压关联特征，并利用机器学习技术构建提取的特征与血压之间的关联模型，用于手术中心律失常患者的血压估计，实现了术中血压变化的精准估计，精度符合美国AAMI标准。相关研究共产生了4篇 SCI论文和4篇EI会议论文，3项中国发明专利。