基于房颤代谢组学时序数据的进化循环神经网络及应用研究

Stroke caused by thrombus abscission is the most common prognostic complication of atrial fibrillation (AF), and is also the main cause of disability and death in patients with AF. Accurate assessment and prediction of the risk of prognosis in patients with AF is of great significance. Metabolomics can reveal the progression of AF based on endogenous small molecular substance. This study is to select significant biomarkers for the prognosis of patients with AF and construct prediction models, by dynamically monitoring changes of the endogenous small molecules in the blood at different time points in patients with AF, using dynamic metabolomics of longitudinal design and analysis method of high dimensional time-series data. For time-series data of metabolomics, we propose a novel model of evolutionary algorithm based recurrent neural network (EA-RNN), which is constructed by the evolvement and improvement of genetic algorithms for parameter estimation, based on recurrent neural network. EA-RNN can be used for feature selection and prediction of high dimensional time-series data. The main research contents include: construction and optimization of EA-RNN model, identification of metabolites and biological interpretation, construction of prediction models, functional verification of metabolic biomarkers. This project combines the statistical model with the actual data to provide the metabolic biomarkers and prediction models that can predict the risk of stroke in patients with atrial fibrillation.

血栓脱落引起的脑卒中是房颤最常见的预后并发症，也是房颤患者致残、致死的主要原因。准确地评估和预测房颤患者的预后风险是极为关键的临床问题。代谢组学可从内源性小分子层面揭示房颤→脑卒中的进展规律。本项目利用纵向设计的动态代谢组学和相应的高维时序数据分析方法，通过动态监测房颤患者不同时间点的血液中内源性小分子的变化，筛选对预测脑卒中有作用的代谢标志物，并建立有效的预测模型。针对代谢组学时序数据的特点，本项目提出一种进化循环神经网络（EA-RNN）算法，该方法在循环神经网络思想的基础上，采用遗传算法进化和改善网络模型的参数估计性质，主要解决高维时序数据的特征筛选和预测问题。研究内容包括：EA-RNN模型建立和参数优化、物质鉴定和生物学解释、建立预测模型、代谢标志物功能验证等。本项目将统计模型与实际数据相结合，期望给出能够标识房颤患者继发脑卒中风险的代谢标志物及预测模型。

房颤是临床上最常见的持续性心律失常，房颤患者缺血性脑卒中发生风险是一般人群的5倍。目前临床上常用评分量表来为房颤患者进行缺血性脑卒中的危险分层，然而其预测效果尚不准确，目前国际上缺少有效的生物标志物来预测房颤患者的脑卒中继发风险。本项目通过房颤和脑卒中血浆代谢组学数据分析，筛选出脑卒中血浆代谢组学生物标志物；基于遗传算法和神经网络建立的EA-RNN模型，能够准确预测房颤患者继发脑卒中的风险；本项目通过房颤和脑卒中蛋白质组学数据分析，筛选出脑卒中相关的免疫炎症通路；通过生物学实验验证了脑卒中的炎症反应相关机制；本项目开发的可视化预测软件能够准确预测房颤患者继发脑卒中的风险，可转化用于各级医疗机构及家庭自测。本项目从深度学习算法研究、多组学研究、生物学实验研究等多个层次揭示了房颤→脑卒中的生物学机制，为房颤个体化治疗策略的制定提供参考依据。

内容获取失败，请点击重试