Evaluation Of Traditional Chinese Medicine’s Chronic Kidney Disease Management’s Benefits And Prognosis—A Machine Learning Approach

ObjectiveThis study intends to collect information on chronie kidney disease management related diagnosis and treatments to construct a disease risk prediction model.Then through analyses and validation of this prediction model,we hope this prediction model can provide a theoretical basis for better disease management.And the ultimate purpose is to improve the prognosis of patients with chronic kidney disease.MethodsIn this study,a retrospective data analysis method was used.From January 1,2012 to December 31,2018,patients with chronic kidney disease stage 4 in the chronic disease management clinic of Guangdong Provincial Hospital of Traditional Chinese Medicine were enrolled.Data like demographic characteristic,laboratory test results,TCM symptoms and syndrome differentiation were collected.Moreover,data like the implementation of TCM chronic disease management,medications and other related variables were also collected.Then,we recorded their primary outcomes,including:dialysis,kidney transplantation and death.According to the outcomes,patients were grouped for further analyses.By combining knowledge maps representation learning and applying machine learning methods,we established classification models of knowledge,relationships,and path vectors based on multiple dimensions such as disease,syndrome,and risk factors.Therefore,we could make predition of the benefits and prognosis for the traditional Chinese medicine’ s chronic disease management.The performance analysis,validation and optimization of different prediction models are carried out through various experimental methods such as feature selection and parameter adjustment.In terms of discrimination,the performance of different models is evaluated by the Area Under the Receiver Operating Characteristic curve(AUROC)and the F1 coefficient.ResultsA total of 256 patients with chronic kidney disease stage 4 were enrolled in the study,of which 155 were in the control group and 101 in the endpoint group.A total of 21 variables were screened by Boruta algorithm,including:crack,enlarged tongue,imprints of the teeth,antihypertensive drugs,categories of drugs administrated,pharmacodynamics,nephrotoxic drugs,unhealthy diet,high quality protein,food taboos,salt intake,drug withdrawal because of deterioration,WeChat articles,ALT/AST,ALB,Urea,TC02,Cr,body weight score,BMI and percentage of upper arm muscle circumference.Then a total of 8 machine learning models were used for training.In terms of discrimination,the Stabilized Linear Discriminant Analysis(SLDA)model achieved the best performance,with the AUROC 0.8;Partial Least Squares(PLS)model achieved sub-optimal performance,with AUROC 0.78;In addition,the performance of the K-Nearest Neighbors(KNN)model was not satisfying,with the lowest AUROC 0.54.The remaining models’ AUROC had no significant difference.Similar to the performance of AUROC,the F1 coefficients of different models were also analyzed.The results showed that SLDA was still the best performance model,with the F1 coefficient 0.84.The sub-optimal model was still the PLS model,with the F1 coefficient 0.83.But the result was not significantly different.The KNN model is not very satisfying,with the F1 coefficient 0.67.ConclusionComputer-based machine learning algorithms,such as the use of SLDA and PLS models,can identify practical-level intervention effects in real time.Moreover,they can correct the deviations in chronic disease management accurately.Therefore,the machine learning method provides a common strategy management method for traditional Chinese medicine’ chronic kidney disease management,which may be extended to the management of chronic diseases of other chronic non-communicable diseases.