The Efficacy Of Acupoint Application In Patients With Diarrhea:A Retrospective Cohort Study

ObjectiveThe present study was a retrospective cohort study based on real-world data.We estimated the efficacy of acupoint application(AA)in patients with diarrhea,the differences of the efficacy between two AA methods and appropriate duration of treatment for the purpose of providing real-world evidence on the efficacy of AA in patients with diarrhea.Also,our study established a machine-learning model(XGBoost)to predict the efficacy of AA in patients with diarrhea on the 7th day to assist clinical decision-making and treatment decisions.MethodsThe study was a registered retrospective cohort study,which consecutively collected 2280 patients with diarrhea from hundreds of primary hospitals nationwide from 22nd August,2020 to 5th November,2020.Participants were divided into AA group and non-AA group according to previous medical records.In this study we evaluated the efficacy of AA in patients with diarrhea on the 3rd,7th,14th and 28th day.Varying degrees of missing data were present across covariates.Missing data of co variates were handled by dummy variable approach.We adopted Generalized Additive Model(GAM)to detect whether there was a nonlinear relationship between numerical variables and diarrheal recovery.If there was a clear nonlinear relationship,these covariates were handled as smooth curve fitting or multiple fractional polynomials.To evaluate the independent effect of AA on diarrhea recovery under different strategies for covariate selection and adjustment,we simultaneously displayed the results from unadjusted,adjusted I,adjusted Ⅱ and fully-adjusted regression models according to STROBE statement.Unadjusted model was not adjusted for any confounding factors;Adjusted Ⅰmodel was adjusted for age and sex;Adjusted Ⅱ model was adjusted for confounders by the change-in-estimate(CIE)and directed acyclic graph(DAG)procedure;Fullyadjusted model was adjusted for region,age,gender,use of TCM decoration,use of Western Medicine,duration of diarrhea,frequency of diarrhea,watery stool,mucous or bloody purulent stool,sour smelled stool and unobvious smelled stool.Also,to more intuitively reflected the different efficacy between AA group and non-AA group,the ORs in four multiple logistic regression models were transformed to RDs using the Delta method to intuitively observe the specific number of the difference out of per 100 patients between AA group and non-AA group.Stratified analyses and interaction test(likelihood ratio test,LRT)were used to explore whether AA and other covariates have interaction effects on the outcomes.Stratified analyses were conducted according to region(North and South),age(<3years,≥3 and<6years,≥6 and<18years,and>18years),sex(male and female),use of TCM decoction(yes and no),use of Western Medicine(yes and no),duration of diarrhea(<2 days and≥ 2 days),frequency of diarrhea(<2 times/day and≥ 2 times/day),loose stool(yes and no),mucous or bloody stool(yes and no),sour smelled stool(yes and no)and unobvious smelled stool(yes and no).Each stratification adjusted for all the variables except the stratification variable itself.LRT was used to explore whether the effect size was statistically different at different levels in subgroups.Additionally,to reduce possible confounding bias caused by nonrandomization in observational study,propensity score matching(PSM)was performed with a 4:1 matching ratio in AA group and non-AA group in this study.Matched variables included all the confounders with a maximum difference of propensity score(PS)of 0.01.In the PS-matched cohort,univariate regression and covariate adjustment using propensity score(CAPS)were presented.Given that PSM may induce sample loss and selection bias,our study implemented inverse probability of treatment weighting(IPTW),standard mortality ratio weighting(SMRW),repeated matching(GenMatch)and other analytic approaches to further evaluate confounding bias caused by nonrandomization.Additionally,our study was an observational study in which unmeasured confounders existed.Consequently,our study calculated E-value to further evaluate what the minimum effect size would have to be for an unmeasured confounder to negate the observed correlation of AA therapy with diarrheal recovery.Secondly,patients with diarrhea who received AA therapy were further divided into wet application group and medication application group to analyze the difference in the efficacy between two groups on the 3rd,7th,14th and 28th day.Multivariate logistic regression was conducted to evaluate the independent effect of different application methods for diarrheal recovery under different strategies for covariate selection and adjustment.We displayed the results from unadjusted,adjusted I,adjusted Ⅱ and fullyadjusted regression models as above.PSM was performed to match all the confounders with a 4:1 matching ratio in wet application group and medication application group and a maximum difference of PS of 0.01 in this study.In the PS-matched cohort,univariate regression and CAPS were presented.The methods were the same as before.Thirdly,multivariate logistic regression was performed to evaluate the independent effect of duration of treatment on diarrheal recovery under different strategies for covariate selection and adjustment.We displayed the results from unadjusted,adjusted Ⅰ,adjusted Ⅱ and fully-adjusted regression models as above.The trend test was applied to test for linear trend relationship between duration of treatment and diarrheal recovery and possible nonlinear relationship between them.Also,we further used smooth curve fitting,recursive method,interaction test and piecewise linear model to explore whether there was saturation and threshold effect between duration of treatment and diarrheal recovery,namely the appropriate duration of treatment for AA.Fourth,our study applied a machine-learning model based on eXtreme Gradient Boosting(XGBoost)to establish a predictive model for efficacy of AA in patients with diarrhea on the 7th day.The model was evaluated by area under the receiver operating characteristic curve(AUC).We compared the classification performance of XGBoost with that of four mainstream machine-learning and ensemble learning methods including Artificial Neural Network(ANN),ANN+Boosting,ANN+Bagging and Support Vector Machine(SVM)in terms of accuracy(ACC),Precision(P)and F1 score.Results1 Study on the efficacy of acupoint applicationIn unadjusted and adjusted I models,the efficacy in AA group was significantly more improved than that in non-AA group on the 7th,14th and 28th day.In adjusted Ⅱand fully-adjusted models,AA group showed better efficacy than non-AA group on the 14th and 28th day.Among above,OR,95%CI and P value are listed as follows.On the 14th day:RD=0.07,95%CI:0.02-0.12,P=0.0060;On the 28th day:RD=0.08,95%CI:0.03-0.13,P=0.0001.After adjusting for confounders,the direction of the estimates of AA on diarrheal recovery stayed the same in the subgroups including regions(North and South),age(<3 years old,≥3 and<6 years old,≥ 6 and<18 years old and>18 years old),gender(male and female),use of TCM decoction(yes and no),use of Western Medicine(yes and no),duration of diarrhea(<2 days and≥ 2 days),frequency of diarrhea(<2 times/day and≥ 2 times/day),watery stools(yes and no),mucous or bloody purulent stool(yes and no),sour smelled stool(yes and no)and unobvious smelled stool(yes and no)on the 3rd,7th,14th and 28th day.P for interaction tests were all greater than 0.05.In present study,we matched region,age,gender,use of TCM decoration,use of Western Medicine,duration of diarrhea,frequency of diarrhea,watery stool,mucous or bloody purulent stool,sour smelled stool and unobvious smelled stool using PSM which was performed with a 4:1 matching ratio in AA group and non-AA group and a maximum difference of PS of 0.01.In the PS-matched cohort,the directions of estimates in univariate regression and CAPS were consistent with that in multivariate regression,but the effect size estimates were different form that in multivariate regression.The direction and magnitude of the estimates broadly agreed across IPTW,SMRW and multivariate regression.We matched all the baseline characteristics using a repeated matching method(GenMatch)which was performed with a 1:1 matching ratio in AA group and non-AA group and a maximum difference of PS of 0.05.In the repeated matched cohort,the directions of estimates in univariate regression and CAPS were consistent with that in multivariate regression,but the effect size was different from that in multivariate regression.In adjusted Ⅱ model E values for the efficacy of AA therapy on the 14th and 28th day were 1.83 and 2.20,indicating an unmeasured confounding variable would require an association of at least RR=1.83 and RR=2.20 respectively with AA and diarrheal recovery to nullify the observed correlations.2 Study on the efficacy of different acupoint application methodsNo statistical significance in efficacy was found between two AA methods for diarrheal recovery on the 3rd,7th,14th and 28th day in unadjusted,adjusted I,adjustedⅡ and fully-adjusted regression models.PSM was performed with a 4:1 matching ratio in wet application group and medication application group and a maximum difference of PS of 0.01 in this study.Matched variables included region,age,gender,use of TCM decoration,use of Western Medicine,duration of diarrhea,frequency of diarrhea,watery stool,mucous or bloody purulent stool,sour smelled stool and unobvious smelled stool.In the PS-matched cohort,the directions of the estimates on 3rd,7th,14th and 28th day in univariate regression and CAPS were consistent with that in multivariate regression,but the effect size estimates were different from that in multivariate regression.3 Study on the duration of treatmentIn adjusted Ⅱ model,OR and 95%CI were 1.08 and 1.03 to 1.13(P=0.0013).The efficacy of AA in diarrheal recovery improved with increasing duration of treatment within the range of the 2nd to 28th day.The trend test effectively showed diarrheal recovery tended to increase gradually and the trend was significant with increasing tertiles of duration of treatment in the four regression models.The best cutoff point was the 4th day between the 2nd and 28th day using smooth curve fitting,recursive method and LRT,etc.The gap between ORs of anterior section and posterior section of smooth curve gradually became narrowed and P value for LRT tended to be insignificant.When the inflection point was the 19th day,P value for LRT=0.076>0.05 and the nonlinear relationship between duration of treatment and diarrheal recovery was no longer significant.Therefore,it can be inferred that the appropriate duration for AA treatment ranged from the 4th to 19th day.4 The predictive model for the efficacy of acupoint application in patients with diarrheaIn the predictive model for the efficacy of AA in patients with diarrhea,the top three variables with the highest importance were age,diarrhea time and region(North).The discriminative ability of the model was good,based on an AUC of 0.81.It was concluded that the model was highly accurate and precise for prediction.However,XGBoost was not superior to other four mainstream machine-learning and ensemble learning algorithms including ANN,ANN+Boosting,ANN+Bagging and SVM.Conclusion1 The efficacy in AA group was significantly more improved than that in non-AA group on the 14th and 28th day.The sensitivity analysis suggested that results were possibly biased toward the null(toward negative).2 No significant difference was found in the efficacy of two AA methods for the treatment of diarrhea on the 3rd,7th,14th and 28th day.3 Between the 2nd and 28th day,the appropriate duration for the treatment of diarrhea with AA ranged from the 4th to 19th day.4 The predictive model for the efficacy of AA in patients with diarrhea based on XGBoost was highly accurate and precise for prediction,which can be used to predict the efficacy of diarrheal recovery with AA.