Development And Preliminary Clinical Application Of Prediction Model Of Chronic Disease Management In Chinese Medicine Of CKD Based On Decision Tree Technology

ObjectiveFirstly,we used the method of systematic review to systematically search and sort out the current research evidence on the prediction model of disease progression in chronic kidney disease(CKD)non-dialysis patients based on artificial intelligence technology to understand the research progress and evidence quality in this field and to summarize the relevant technologies and methods.Then,based on the data of CKD patients collected by our center,the decision tree was used to screen the Western medicine factors,Chinese medicine factors,and chronic disease management factors affecting the disease progression of CKD and build a prediction model of chronic disease management in Chinese medicine of CKD.Compare the prediction performance of the decision tree model with the traditional modeling methods of Lasso Cox regression and Cox regression.The optimal model was developed into a visual predictive mini program for clinical chronic disease management.While predicting the disease progression of CKD patients,patients were visually demonstrated that the risk of disease progression decreased significantly after the control of risk factors so as to improve patient adherence to chronic disease management.Finally,the prediction model was applied in clinical practice on a small scale,and the qualitative research method was used to interview patients’ understanding and acceptance of this model,which provided reference evidence for its wide application in clinical practice in the later stage.Methods1.A systematic review was used to sort out the research evidence on the prediction model of disease progression in CKD non-dialysis patients constructed by artificial intelligence technology.We used disease progression,RRT,death,dialysis,ESRD,ESKD,Chronic Kidney Disease,artificial intelligence,machine learning,deep learning,and other Chinese and English search terms to search the relevant literatures published from PubMed,Embase,Cochrane Library.Chinese National Knowledge Infrastructure,Wanfang,VIP Database,and the Chinese Biomedicine Literature Database from the establishment of these databases until October 11,2022.Literature screening was conducted according to inclusion and exclusion criteria.Data extraction and quality assessment were conducted for the included literature according to the TRIPOD declaration and the CHARMS guidelines.2.The decision tree was used to construct the prediction model of chronic disease management of Chinese medicine of CKD:this was a retrospective cohort study.The study subjects were CKD patients who visited the Chronic Disease Management Clinic at the Nephrology Department of the Guangdong Provincial Hospital of Chinese Medicine and agreed to join the chronic disease management,based on the chronic disease management system,electronic case system,and follow-up data,Chinese and Western.medicine clinical data and the occurrence of endpoint events were collected.Endpoint events were defined as renal replacement therapy(peritoneal dialysis,hemodialysis,or kidney transplantation)or estimated glomerular filtration rate(eGFR)<5 ml/(min·1.73m2).A single imputation method based on the model was used to impute the missing values.Pearson correlation analysis was used to select variables.Firstly,four balance training sets were created respectively for 1-4 years using the retractable sampling method.The original data of 1-4 years were taken as the testing set,and decision tree was used to screen the influencing factors of Chinese and Western medicine and chronic disease management factors of the endpoint events in 1-4 years,and the prediction models for predicting the occurrence of endpoint events in 1-4 years were constructed respectively.Then,based on the original data,we used the traditional methods of Lasso Cox regression and Cox regression to construct the prediction model,then drew a nomogram graph.Compared the predictive performance of the decision tree model with the Lasso Cox regression and Cox regression models,the evaluation indexes were sensitivity,specificity,accuracy,and AUC(Area under the receiver operating characteristic curve).Then,the optimal prediction model after the above comparison was developed into a visual WeChat mini program for clinical chronic disease management.Finally,data from CKD patients who were followed up for more than half a year in the Ambispective Cohort Study(SMP-CKD cohort)were collected.The optimal model was used for predicting the endpoint events.The predicted results were compared with the actual occurrence of endpoint events,and the accuracy,sensitivity,and specificity of the prediction model were calculated.3.We selected a small number of CKD patients from the chronic disease management outpatient department of Guangdong Hospital of Chinese Medicine to use the above predictive mini program for chronic disease management in Chinese medicine of CKD.Then the qualitative research method was used to interview these patients through face-to-face in-depth interviews and collect the demographic data and clinical information after obtaining informed consent.After transcribing the interview data,NVivo software was used to organize the data.Results1.This study finally included 21 pieces of literature,18 in English and 3 in Chinese.The subjects were mainly IgA nephropathy,diabetic nephropathy,and CKD patients.There were 7 in IgA nephropathy patients,3 in diabetic nephropathy patients,and the remaining 11 in CKD patients.There were ten first authors from China,five from Italy,two from the United States,and one each from Canada,the Netherlands,Germany,and Australia.Most studies’ predictive factors mainly included demographic data,medical history,and laboratory data from blood and urine samples.Artificial intelligence modeling methods:the most commonly used included XGBoost,artificial neural network,random forest,decision tree,support vector machine,K-nearest neighbor,logistic regression,Bayes,etc.Among the 21 studies,16 mainly adopted cross-validation or proportional random sample division for internal verification,while only six conducted external verification.The evaluation indexes of model performance mainly included accuracy,sensitivity,specificity,negative predictive value,AUC,F1-score/f-measure,C statistic,etc.For clinical applications,only six studies developed prediction models into web calculators or decision support systems.2.A total of 1373 CKD patients were included in this study,and the median follow-up time was 25(16.37)months.By the end of the follow-up,122 patients(8.89%)occurred endpoint events.The artificial intelligence method of the decision tree was used to build the models.The first year’s model included eGFR,Urea,hemoglobin(Hb),total dialectical score,fatigue,gender,the score of exercise cognition,and urine protein(PRO).Variables included in the second-year model were eGFR,Urea,albumin(ALB),age,the score of Nephrotoxic drug cognition,urinary protein creatinine ratio,damp-heat syndrome,fatigue,total carbon dioxide(TCO2),and low density lipoprotein cholesterol(LDL-C).Variables included in the third-year model were:Urea,ALB,the score of Nephrotoxic drug cognition,and the score of cognition of causes of aggravation of kidney disease.Variables included in the fourth-year model were eGFR,dampness-heat syndrome,triglyceride(TG),and nocturia.11 variables were finally included in the model established by the Lasso Cox regression,including eGFR,extracellular water/total body water(ECW/TBW),PRO,Urea,red blood cell(RBC),diuretics,α antihypertensive drugs,calcium channel blocker(CCB)antihypertensive drugs,compound α-ketoic acid tablets,folic acid tablets,and erythropoietin(EPO).11 variables were finally included in the model established by the Univariate and multivariate Cox regression,including eGFR,hypertension,epigastric hypermnesia,TBW,Waist-Hip ratio(WHR),PRO,Urea,diuretic,angiotensin converting enzyme inhibitor/angiotensin receptor blocker(ACEI/ARB),compound α-ketoic acid tablets and hormones.The performance of the above three models was evaluated.In the first year,the sensitivity of the decision tree,Cox regression,and Lasso Cox regression models were 1,0.970,and 1,respectively.The specificity was 0.943,0.898,and 0.896,respectively.Regarding accuracy,they were 0.945,0.900,and 0.899,respectively.Regarding AUC,they were 0.976,0.963,and 0.964,respectively.In the second year,the sensitivity of the decision tree,Cox regression,and Lasso Cox regression models were 0.951,0.877,and 0.963,respectively.The specificity was 0.954,0.955,and 0.862,respectively.In terms of accuracy,they were 0.954,0.948,and 0.871,respectively.Regarding AUC,they were 0.979,0.963,and 0.955,respectively.In the third year,the sensitivity of the decision tree,Cox regression,and Lasso Cox regression models were 0.939,0.889,and 0.929,respectively.The specificity was 0.888,0.920,and 0.883,respectively.In terms of accuracy,they were 0.899,0.914,and 0.893,respectively.In terms of AUC,0.936,0.947,and 0.945,respectively.In the fourth year,the sensitivity of the decision tree,Cox regression,and Lasso Cox regression models were 0.848,0.907,and 0.907,respectively.The specificity was 0.814,0.887,and 0.893,respectively.In terms of accuracy,they were 0.827,0.895,and 0.898,respectively.The AUC values were 0.899,0.910,and 0.926.respectively.Based on the comprehensive judgment of the results,the model constructed by the decision tree is the optimal prediction model,which was developed into a visual WeChat predictive mini program for clinical use.Finally,the data of 47 CKD patients who were followed up for more than half a year in the SMP-CKD cohort study were collected.The first-year model constructed by the decision tree was used for prediction.The accuracy of the first-year model was 97.87%,the sensitivity was 1,and the specificity was 0.978.3.The qualitative study discussed CKD patients’ views on the prediction model,on the results of the prediction model,their suggestions on the application of the prediction model,and the medical help they hope to obtain.The views on the prediction model could be summarized as positive and negative effects:the positive effects included:(1)helping patients and doctors to know the disease progression;(2)helping to prevent the disease progression;(3)promoting behavioral change;(4)help to maintain the compliance of self-chronic disease management;(5)increase the convenience of disease management;(6)enhance the confidence of self-managing the disease;(7)improve the psychological burden of patients.Negative effects included:(1)discouraging patients’ confidence in the disease treatment;(2)increasing the psychological burden of patients.Views on the results of the prediction model could be summarized as disbelief/doubt or belief:the reasons for disbelief/doubt were mainly:(1)the questioning of the development of modern medicine;(2)lack of understanding of the principle of prediction model;(3)whether there is data of prospective verification;(4)there may be emergencies in life;(5)more trust in physical symptoms;(6)more trust in doctors;(7)the disease is dynamic;(8)the human body is influenced by many environmental factors.The main reasons for belief were:(1)the prediction results align with psychological expectations;(2)belief in the development of science.Suggestions for applying the prediction model included suggestions for the prediction model itself and the application for chronic disease management.Suggestions for the prediction model itself included:(1)increasing the inclusion of other risk factors;(2)improving the prediction accuracy of the model;(3)increasing the predictive time;(4)developing disease prediction models for early-stage CKD patients.Suggestions for the application of the prediction model in chronic disease management included:(1)using other communication methods instead of predictive models to manage patients;(2)the permissions of the predictive mini program are not opened to patients;(3)help patients understand and accept the prediction model;(4)choose the right time to predict;(5)predict the outcome and inform the results according to the different conditions of patients;(6)give targeted disease guidance based on the predicted results;(7)the prediction results do not show the probability,only the change curve;(8)timely psychological counseling for patients with poor prognosis;(9)the mini program to add the patient’s personal annual management report.The medical help CKD patients want includes:(1)medical staff can be reached in case of emergency;(2)provide individualized professional guidance;(3)remind abnormal indicators when seeing a doctor;(4)provide various forms of disease science;(5)provide dietary advice/develop recipes;(6)provide psychological counseling.Finally,this study also summarized the factors that influence the mood of patients,including(1)psychological quality;(2)state of mind;(3)experienced a period of more severe symptoms;(4)the cognition of their disease outcome;(5)age;(6)good or bad prediction results;(7)duration of disease;(8)how to think of the prediction results;(9)whether the prediction results can be improved by controlling risk factors.Conclusion1.The artificial intelligence model provides a reliable and accurate method for personalized prognosis prediction in CKD patients and for identifying high-risk patients.However,the actual clinical application of the research is still very few,which may need to strengthen the clinical application further.2.The result of the decision tree model shows that gender,age,eGFR,Urea,PRO,urinary protein creatinine ratio,Hb,ALB,TCO2,TG,LDL-C,fatigue,nocturia,total dialectical score,damp-heat syndrome,the score of exercise cognition,the score of Nephrotoxic drug cognition,and the score of cognition of causes of aggravation of kidney disease were essential factors affecting the disease progression in CKD patients.Compared with the Lasso Cox regression and Cox regression models,the decision tree model showed better predictive performance in 1-2 years but did not show better predictive performance in 3-4 years,which may require further improvement in the artificial intelligence method.3.The qualitative study results showed that most patients believed that applying the predictive mini-program developed by the prediction model of chronic disease management in Chinese medicine of CKD to clinical chronic disease management has certain positive effects.However,it may also have negative effects,mainly because the bad prediction results may increase the psychological burden on patients.At the same time,patients also gave suggestions and requirements for applying the prediction mini-program to chronic disease management according to their own ideas.Therefore,in the later stage,how to further optimize and improve the process of applying the prediction mini-program to chronic disease management according to patients’ suggestions and needs is the focus of our medical staff.