Prediction of renal insufficiency in Pima Indians with nephropathy of type 2 diabetes mellitus

作     者：Goldfarb-Rumyantzev, AS Pappas, L 

作者机构：Univ Utah Sch Med Div Nephrol Salt Lake City UT 84112 USA Univ Utah Sch Med Div Hypertens Salt Lake City UT 84112 USA 

出 版 物：《AMERICAN JOURNAL OF KIDNEY DISEASES》 (美国肾病杂志)

年 卷 期：2002年第40卷第2期

页      面：252-264页

核心收录：

学科分类：1002[医学-临床医学] 100201[医学-内科学(含：心血管病、血液病、呼吸系病、消化系病、内分泌与代谢病、肾病、风湿病、传染病)] 10[医学] 

主　　题：prediction model artificial neural network tree-based model generalized additive model Pima Indians diabetic nephropathy renal insufficiency outcome 

摘      要：Background: A high prevalence and early onset of type 2 diabetes in Pima Indians is well known. Our objective is to use several statistical models to identify predictors of glomerular filtration rate (GFR) deterioration and develop an algorithm to predict GFR 4 years after the initial evaluation. Methods: All records (n = 86) were randomly assigned to a training set (n = 60) and a testing set (n = 26). Linear regression, generalized additive, tree-based, and artificial neural network models were used to identify predictors of outcome and develop a prediction algorithm. Results: Proteinuria remained the single most important predictor of long-term renal function;other predictors included baseline GFR, blood pressure, plasma renin activity, lipid profile, age, weight/body mass index, and diabetes duration. All four models achieved a good correlation (r = 0.73 to 0.78) between observed and predicted 4-year GFRs on a separate (testing) data set. Best results in predicting the value of GFR were achieved using a tree-based model with six terminal nodes (r = 0.78;root mean squared prediction error = 38.9). The tree-based and generalized additive models achieved high positive (91%) and negative (100%) predictive values in identifying subjects, who developed depressed GFRs in 4 years. An artificial neural network achieved the highest area under the receiver operating characteristic curve (0.91). Conclusion: GFR depression within 4 years can be predicted with a precision that suggests potential clinical utility. A tree-based model with six terminal nodes has shown the best results in predicting the actual value of GFR, whereas an artificial neural network is the model of choice to identify the group of patients that will develop renal insufficiency. (C) 2002 by the National Kidney Foundation, Inc.