Prognostic Models Of IgA Nephropathy

Background and Objective:Immunoglobulin A nephropathy(IgAN)is the most common type of chronic glomerulonephritis and one of the main causes for end-stage renal diseases in China.The heterogeneity of IgAN leads to a great individual difference in prognosis.Prognosis assessment is an essential factor for precise treatment and patient self-management.It has always got intense attentions from physicians and scientists.Current prognosis evaluation models are from European and American.There are very few models to reflect the prognosis in Chinese IgAN patients.In the present study,we aimed to establish various prognosis evaluation models using pathological characteristics and clinical indicatorsMethods:(1)The retrospective study cohort concluded adult IgAN patients who were enrolled from January 1st,2000 to December 31th,2010 at PLA General Hospital.The prospective cohort included multicenter prospective data with follow-up starting in May 2014.After the data screened by the box chart and the casewise diagnosis of regression analysis,we randomly divided the retrospective cohort into the modeling group and the internal verification group at a ratio of 7:3.The prospective cohort was used as an independent external validation group.(2)eGFR reduction>50%or ESRD was the endpoints of pathological and the clinicalpathological model.The pathological model was created using COX regression,Kaplan-meier survival curve analysis,and log-rank test.We combined the clinical indexes and pathological indicators and applied the R programming language to set up the clinical pathology nomogram model for prognosis risk prediction.The models were evaluated through the C-index(Harrell's concordance index),the area under ROC curve(AUC),and Akaike information criterion(AIC).(3)The artificial intelligence model for predicting the endpoint events of patients was based on clinical and pathological information.Its prediction ability was assessed and compared by AUC,prediction accuracy,and other indexes.Results:(1)After data denoising,the modeling group had 423 cases(79.5±32.5 months of follow-up),and the internal verification group had 198 patients(81.9±28.8 months of follow-up).566 patients were included as the external validation cohort(19.23 ± 9.53 months of follow-up).(2)A hierarchical pathological model was constructed by combining survival curves.Membrane proliferation and tubulointerstitial damage were selected as key pathological prognosis indicators.Clinicalpathological model was established using mesangial proliferation,tubulointerstitial disease,mean arterial pressure,and baseline urinary protein quantification.The R programming was used to build nomogram.In the modeling group,AUC of the pathological model was 0.8,and it was 0.88 for the clinicalpathological model.The performance comparison was done in the validation groups.The results indicated that the clinicalpathological model showed better prediction efficiency than other pathological models in the internal validation cohort(p<0.05).In the external validation cohort,the clinical pathology model showed a trend of increased prediction performance as compared to other pathological classifications.(3)The artificial neural network model was established and assessed.A multi-layer perception(MLP)model and a radial basis function(RBF)model was constructed.The variables was selected by weight>0.05.The AUC of the MLP model was 0.92 for the modeling cohort,0.91 for the internal validation cohort,and 0.85 for the external validation cohort.The AUC of the RBF model was 0.9 for the modeling cohort,0.94 for the internal validation cohort,and 0.84 for the external validation cohort.There was no statistically significant difference in the predictive efficiency between these two models(p>0.05).However,the predictive efficiency of both models was higher than current pathological models(p<0.05).Conclusions:The hierarchical pathological model predicts as effectively as Oxford classification does.In addition,the former is more intuitive than the latter.In the prospective cohort,the prediction efficiency of the clinical pathology model tended to be increased than that of pathological models.The artificial neural network model can use a variety of clinical and pathological data and has better prediction performance than other pathological models.