基于人工神经网络的肾小球滤过率估算研究

发布时间：2018-04-05 12:22

本文选题：肾小球滤过率　切入点：慢性肾脏病　出处：《华南理工大学》2012年硕士论文

【摘要】：慢性肾脏病（CKD）是一类严重影响人类健康的疾病。临床上广泛采用肾小球滤过率（GFR）来评价肾功能及作为辅助诊断的依据。目前基于99mTc-DTPA的同位素成像方法准确性高，被认为是GFR检测的金标准，但需要昂贵的配套设备，限制了其临床推广。为了寻找一种方便、快捷、价廉、在普通医院也能实施的GFR检测方法，有美国学者提出GFR估算经验方程，并在美国得到临床推广。但该方程移植到中国人群的适用性差，远远达不到临床要求，而国内相关研究进展缓慢。本文是与中山大学附属第三医院肾内科合作，在前期大量随访数据和相关研究的基础上，采用了人工神经网络方法建立GFR估算模型。本研究积累了共1180例样本，分为训练集、内部验证集和外部验证集。所有数据经过预处理后，分别采用BP神经网络、GRNN、Legendre神经网络、多项式神经网络建立模型，并对BP神经网络采用了常见的遗传优化算法，尝试采用S函数将Legendre神经网络由单输入变量推广至多输入变量，提出了多项式神经网络幂次组合的自适应确定方法。同时还采用MIV方法基于遗传算法优化的BP网络进行输入变量的筛选，以提高模型性能。各模型的性能主要从准确性和一致性两个方面来进行评估，其中准确性用偏差、偏差百分数、符合率等指标来反映，而一致性由Bland-Altman图来体现，并且通过相应的统计学检验方法来检验各指标之间的差异是否具有统计学意义。通过各神经网络模型和经验方程的性能比较，最终选用6输入变量的遗传算法优化BP神经网络模型作为较优模型，并设计了相应的可视化界面以实现模型展示。本文提出的模型，其性能优于临床常用的经验方程，，可进行临床推广以实现进一步的验证和修正。同时人工神经网络的性能均要优于经验方程，这一结果表明以人工神经网络为代表的机器学习方法在生物医学数据处理方面，相对于经典的统计学方法而言具有无可比拟的优势。
[Abstract]:Chronic kidney disease (CKD) is a kind of disease which seriously affects human health.Glomerular filtration rate (GFR) is widely used to evaluate renal function and serve as a basis for auxiliary diagnosis.At present, the isotope imaging method based on 99mTc-DTPA has high accuracy and is considered to be the gold standard of GFR detection, but it needs expensive matching equipment, which limits its clinical popularization.In order to find a convenient, fast and inexpensive method for GFR detection in general hospitals, some American scholars put forward the empirical equation of GFR estimation, and it has been popularized in the United States.However, the applicability of the equation to Chinese population is poor, which is far from the clinical requirement, and the domestic research progress is slow.In this paper, in cooperation with the Department of Renal Medicine of the third affiliated Hospital of Sun Yat-sen University, on the basis of a large number of follow-up data and related studies, an artificial neural network method is used to establish the GFR estimation model.A total of 1180 samples were collected in this study, which were divided into training set, internal verification set and external verification set.After all the data are preprocessed, BP neural network (GRN) and polynomial neural network are used to establish the model, and the common genetic optimization algorithm is used for BP neural network.This paper attempts to generalize Legendre neural network from single input variable to input variable by S function, and presents an adaptive method for power combination of polynomial neural network.At the same time, MIV method is used to select input variables based on BP neural network optimized by genetic algorithm to improve the performance of the model.The performance of each model is mainly evaluated from two aspects of accuracy and consistency, in which the accuracy is reflected by deviation, deviation percentage, coincidence rate, and the consistency is reflected by Bland-Altman diagram.And through the corresponding statistical test methods to test whether the differences between the indicators have statistical significance.By comparing the performance of each neural network model and the empirical equation, the genetic algorithm with 6 input variables is selected to optimize the BP neural network model as the better model, and the corresponding visual interface is designed to realize the model display.The model proposed in this paper is superior to the commonly used empirical equations and can be popularized for further validation and correction.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：R318.0

【相似文献】