神经源性膀胱患者肾小球滤过率评估模型的开发和验证

Development and validation an assessing glomerular filtration rate model in patients with neurogenic bladder.

目的通过开发一个用于评估神经源性膀胱(NB)患者肾小球滤过率(GFR)的多层感知(MP)神经网络模型,并与传统线性回归公式进行对比,评价MP神经网络模型预测NB患者GFR的适用性。方法选取2017年3月至2021年9月中国康复研究中心北京博爱医院收集的206例成年NB患者资料作为研究对象,采用随机数字表法选择其中141例患者资料作为开发组,以剩余65例作为测试组,建立MP神经网络GFR评估模型。以99 m锝-二乙烯三胺五乙酸(99 mTc-DTPA)核素肾动态显像法(Gates法)检测的GFR作为参考值(rGFR),对改良版中国基于肌酐的GFR线性回归评估公式(C-GFR cr)、CKD-EPI cr线性回归评估公式(EPI-GFR cr)以及MP神经网络模型(M-GFR cr)评估的GFR进行相关性、精确度、偏差、绝对偏差和准确性的比较。结果M-GFR cr与rGFR相关性最高(r=0.86,P<0.001)。M-GFR cr的绝对偏差低于C-GFR cr和EPI-GFR cr(P<0.05);M-GFR cr的±15%准确性要高于C-GFR cr和EPI-GFR cr,差异均有统计学意义(P<0.05);M-GFR cr的±30%准确性同样高于C-GFR cr和EPI-GFR cr,差异均有统计学意义(P<0.05);M-GFR cr的±50%准确性依然高于C-GFR cr和EPI-GFR cr,差异均有统计学意义(P<0.05)。结论基于肌酐的MP神经网络模型评估的GFR的相关性、精确度、绝对偏差和准确性优于基于肌酐的GFR线性回归评估公式(C-GFR cr、EPI-GFR cr)评估的GFR,推荐在NB患者的GFR评估中使用。

Objective Through the department of a multilayer perceptron(MP)neural network model for assessing glomerular filtration rate(GFR)in patients with neurogenic bladder(NB),and compared it with traditional linear regression formulas to evaluate the Suitability of MP neural network model for predicting GFR of NB patient.Methods The data of 206 adult NB patients collected from the Beijing Bo'ai Hospital,China Rehabilitation Research Center from March 2017 to September 2021 were selected as the research object.According to the existing research,the random number table method was used to select the data of 141 patients as the development group,and the remaining 65 patients were used as the test group,to establish the MP neural network GFR evaluation model.The GFR detected by 99m Tc-DTPA radionuclide dynamic renal imaging method(Gates method)was used as the reference value(rGFR).The GFR from the modified Chinese creatinine-based model(C-GFR cr),the Chronic Kidney Disease Epidemiology Collaboration(CKD-EPI)creatinine-based(EPI-GFR cr)and the MP neural network model(M-GFR cr)were compared for correlation,precision,difference,absolute difference and accuracy.Results M-GFR cr had the highest correlation with rGFR(r=0.86,P<0.001).The absolute difference of M-GFR cr was lower than that of C-GFR cr and EPI-GFR cr,the differences were statistically significant(P<0.05).The±15%accuracy of M-GFR cr was higher than C-GFR cr and EPI-GFR cr,the differences were statistically significant(P<0.05);the±30%accuracy of M-GFR cr was also higher than C-GFR cr and EPI-GFR cr,the differences were statistically significant(P<0.05);the±50%accuracy of M-GFR cr was still higher than C-GFR cr and EPI-GFR cr,the differences were statistically significant(P<0.05).Conclusion The correlation,precision,absolute difference and accuracy of GFR assessed by the MP neural network model are better than the GFR linear regression assessment formula(C-GFR cr,EPI-GFR cr).The MP neural network model are recommended for GFR assessed in NB patients.