Study On Prognosis Models Of IgA Nephropathy Based On Clinicopathological Features And TCM Syndrome Elements

AimExploring whether the prognostic factors of Chinese medicine can help to predict IgAN prognosis,and comparing the predicting ability among algorithms including decision tree,adaptive boosting,random forest and support vector machine.MethodsThis study is a retrospective cohort study on IgA nephropathy patients proven by renal biopsies.Baseline demographic data,clinicopathological indicators,TCM syndrome elements,and follow-up information were collected.The end points were defined as doubled serum creatinine,eGFR decreased larger than 50%,ESRD,dialysis or death.The Cox proportional hazards model was conducted to screen for factors associated with poor prognosis of IgA nephropathy.The prognostic factors were divided into Western medicine factors,traditional Chinese medicine(TCM)factors.Decision tree,adaptive boosting,random forest and support vector machine were conducted to build predictive models to predict whether IgAN patients have a combined end point within 5 years after biopsies,respectively basing on Western medicine factors,TCM factors and joint Chinese and Western medicine factors.And the prognostic abilities of models were evaluated and compared.ResultsThe enrolled 402 patients were 32.5 years-old(inter-quartile range,25-41 years),including 214(53.2%)females,with a median disease course of 6(1-24)months.After a median 2.2(1.1-3.9)years long follow-up course,45 patients came to the end points.Cox regression univariate analysis showed that 31 variables were closely related to the prognosis of IgA nephropathy,including systolic blood pressure,diastolic blood pressure,onset fatigue,onset edema,asymptomatic onset,history of hypertension,urine protein,blood urea nitrogen,serum creatinine,estimated glomerular filtration rate,uric acid,serum potassium,phosphorus,total serum protein,serum albumin,Haas grade,Lee grade,Katafuchi scores,MEST scores,crescent proportion,fatigue,edema,nocturia,chilly,nausea,pink tongue,pale tongue,white greasy tongue coating,yang deficiency of spleen and kidney,stasis and amount of subordinate syndromes.According to collinear diagnosis,4 variables were excluded,including baseline systolic blood pressure,baseline diastolic blood pressure,total serum protein and Lee grade,and mean arterial pressure was added in.Finally,there were 28 variables joined the multivariate analysis and prediction models modeling.Cox regression multivariate analysis showed baseline serum creatinine,the T score of MEST scores,yang deficiency of spleen and kidney were independent prognostic risk factors.The data of patients came to end points within 5 years and did not come to end points after 5 years long follow-up,was randomly divided into training set and test set by 7:3 ratio.Twelve predictive models were built by decision tree,adaptive boosting,random forest and support vector machine respectively basing on Western medicine factors,TCM factors,joint Chinese and Western medicine factors.The analysis capabilities of each model on the training set were evaluated.Under the same algorithm,the error rates were TCM factors>Western factors>joint Chinese and Western medicine factors,while the R2 and areas under the curve(AUG)of receiver operating characteristic(ROC)curves were joint Chinese and Western medicine factors>Western factors>TCM factors.When the factors were the same,random forest performed best with the lowest error rate,largest R2 and AUG of ROC.The decision tree performed worst with the highest error rate,smallest R2 and AUG of ROC curve.Adaptive boosting and support vector machine were tied for second.The outcomes of the test set were predicted by 12 models.All the models were good enough to predict the outcomes,as the AUGs of ROC curves were all lager than 0.7.Notably,the AUG of ROC curve was more than 0.9,reaching 0.91,when modeled joint Chinese and Western medicine factors by support vector machine.The joint Chinese and Western medicine factors performed best under the same algorithm,while the Western medicine factors got the second and the TCM factors got third.When the factors were the same,support vector machine performed best with the largest precision,R2 and AUG of ROC curve.Random forest had the same rank of error rate with support vector machines,and the R2 of which was second,while the AUG of ROC curve was the third.The overall performance of random forest was second.The adaptive boosting models had the highest overall error rate,the third R2 and the second AUG of ROC curve.The overall prediction performance of adaptive boosting was the third.The decision trees had the third error rate,but the R2 and AUG of ROC curve were the last,thus the the overall prediction performance was the fourth.ConclusionAll models built by TCM factors or Western medicine factors showed good ability in predictive performance in this study,but models built by Western medicine factors performed better.Modeling with joint Chinese and Western medicine factors was superior to just TCM factors or Western medicine factors.When conducting a similar study,if the TCM data was available,the method of modeling with joint Chinese and Western medicine factors is recommended.Four algorithms,decision tree,adaptive boosting,random forest and support vector machine,all has been proven the ability to build a good predictive model.Under the same variables,the performance of support vector machine is optimal,followed by random forest,while the adaptive boosting got the third and the decision tree got the fourth.Because each algorithm has its advantages and limitations,the selection of algorithm should base on actual needs.