基于特征选择的自适应模糊神经网络在肾小球滤过率中的应用

Application of Adaptive Fuzzy Neural Network Based on Feature Selection in Glomerular Filtration Rate

临床上广泛使用肾小球滤过率(GFR)评价肾功能指标,医生根据GFR预测出慢性肾病(CKD)阶段进而制定相应的治疗方案。菊粉清除率和同位素标记物清除率一直为测定GFR的主要标准。但菊粉价格昂贵、同位素标记方法具有放射性,限制了它们用于GFR的检测。提出一种特征选择的自适应模糊神经网络的进展过程GFR估计方法,分别对6个月、12个月及18个月后的慢性肾病患者进行GFR估计。先对29个特征进行相关性分析,将筛选出来的5个特征进行模糊化、初始化隶属度函数和模糊规则生成,得到模糊神经网络(AFNN),然后用参数训练AFNN模型,得到最优AFNN,最后用新样本数据进行GFR估计,得到误差结果并进行评估。实验结果表明,运用该方法,GER估计误差均小于其它方法,其中最小标准化误差达到1.079 5×10-6,泛化能力增强。

In clinical diagnosis,Glomerular filtration rate（GFR）is widely used to evaluate renal function.Doctors predict the progress of chronic kidney disease（CKD）stages and then make the appropriate treatments according to GFR.Inulin clearance and isotope marker clearance have been considered as the gold standard for GFR detection,but the high cost of inulin limits its routine clinical application and the usage of isotope marker clearance is limited by radiation.This paper proposes a GER estimation method based on adaptive fuzzy neural network with feature selection and makes GER estimation on patients with chronic kidney disease at 6months,12 months and 18 months.Firstly,29 features are analyzed by correlation analysis and then five features are selected.Secondly,the five features are fuzzified get fuzzy neural network（AFNN）,the membership functions are initialized and the fuzzy rules are maked.After that,the AFNN model is trained by the parameters to get the optimal AFNN.Finally,GFR estimation is performed with new sample data to get the error results.The experiment shows that this GER estimation method is better than other methods for the minimum standard error is 1.079 5×10^-6,and generalization ability is enhanced.