Study On Progress Delay Factors And Outcome Prediction Of Randomized Controlled Trials On Traditional Chinese Medicine

BackgroundClinical trials can be delayed for a variety of causes.And clinical trials in traditional Chinese medicine(TCM),due to the particularities of TCM theory,TCM interventions and the study sites,are more susceptible to research delays and barriers.Therefore,researches,which aim to minimize research delay and remove research barriers by exploring impact factors associated with carrying out a TCM clinical trials,and which aim to finish the trials in advance via predicting the outcomes,in particular,for the TCM clinical trials undergone significant delay is needed.Taking a TCM randomised controlled trial(RCT)Effect of an exercise-based cardiac rehabilitation program"Baduanjin Eight-Silken-Movements with self-efficacy building" for heart failure(BESMILE-HF study)as an example,this study aims to explore the factors related to trial delay in TCM,thus keep BESMILE-HF moving forward.The present study also aims to provide innovative methods and pragmatic solutions to alleviate trial delay in the progress of TCM trials.ObjectiveIn view of the enrollment difficulty and trial delay in randomised controlled trials of TCM,and taking BESMILE-HF as a study case,there are two objectives in the present study.One is to analyze the possible factors that may affect the potential patients to participate in the TCM trials,thus developing pragmatic strategies to promote patients recruiting and enrolling.The other is to predict the outcome of the BESMILE-HF study during the course of patients recruiting and enrolling,thus providing evidence to supporting decision-making of stopping the trials before achieving their target sample size.MethodsPart one:Study on Factors Associated with Patients Recruitment in BESMILE-HFThis cross-sectional study collected and analyzed the recorded information of the patients who were recruited to the BEMISLE-HF study.Then according to whether participating in the BESMILE-HF study or not,patients,who were potentially eligible for the BESMILE-HF study,and received telephone interviews and providing information,were divided into the participating group and the rejection group.Basic clinical data and interview information of patients in the two groups were comparatively analyzed,and factors that influenced patient’s decisions on participating in or refusing to the BESMILE-HF study were evaluated using a logistic regression model.Part two:Development and Validation of trial outcome prediction models in the BESMILE-HF study.Based on the data of 47 patients(24 patients in the Baduanjin exercise group and 23 patients in the control group)who had completed the observation of the BESMILE-HF study,the labeled samples formed up by dataset were built in which features(input)were predictors’ overall distribution in dataset,and labels(output,effect size)were the total scores of the Minnesota Living with Heart Failure Questionnaire(MLHFQ)and peak oxygen consumption(peak VO2)in dataset.Multi-layer perceptron(MLP)neural network model and support vector regression(SVR)method were used to respectively construct the prediction models of quality of life and peak VO2 in the BESMILE-HF study.And the candidate variables for the prediction models included demographics,symptoms and signs,comorbidities,foundational treatment,physical function,depression total score from hospital anxiety and depression scale(HADS),exercise ability,quality of life and RCT study characteristics.The prediction value of the four models were evaluated by mean absolute error(MAE),root mean square error(RMSE)and mean absolute percentage error(MAPE).All statistical analysis was done in by using Python 3.7.ResultsPart one:Study on Factors Associated with Patients Recruitment in BESMILE-HF1.Chronic disease management center of heart failure(HF)is the primary site to recruit patients for the BESMILE-HF study.And up to 98.42%of HF patients were recruited through chronic disease management center of HF.However,there are many problems in its registry system,such as information missing and information outdated.The registry system of chronic disease management center of HF is on inadequate management.2.Univariate analysis showed that duration of HF,accompanying for clinical consultation was all significantly different(P<0.05)between the two groups.There were no statistically differences in gender,age,age group,marriage,educational level,work,medical insurance type,nationality,residence,having received TCM treatment,having received treatment with traditional Chinese exercises,holding belief of TCM being effective,health education for HF,arrival means to the site,arrival time to the site,transportation fees for arriving site,ways knowing BESMILE-HF and clinical trial participation experience(P>0.05)between the two group.3.Multivariate logistic regression analysis showed that no requirement of accompanying for clinical consultation(P<0.05,OR=0.011,95%CI:0.001,0.107)and duration of HF less than 5 years(P<0.05,OR=0.060,95%CI:0.007,0.498)were identified as independent factors for participating in the BESMILE-HF study.Patients without requirement of accompanying for clinical consultation and those with duration of HF less than 5 years are more likely to be successfully recruited to participant in the BESMILE-HF study.Part two:Development and Validation of Outcome Prediction Models in BESMILE-HF1.There was no significant difference in the candidate variables,including demographics,symptoms and signs,comorbidities,foundational treatment,physical function,depression total score from HADS,exercise ability and quality of life,between the control group and the Baduanjin exercise group(P>0.05).2.When developing the prediction model of peak VO2 in the BESMILE-HF study,six important characteristics,including the sample size of each dataset,proportion of Baduanjin intervention,consuming Chinese medicine,level 2 of International Physical Activity Questionnaire-Short Form(IPAQ-SF),baseline systolic blood pressure(SBP),and baseline depression score from HADS were screened out through the decision tree.The MLP optimal model was 6-46-1,in which the model input layer contained six neurons,the hidden one layers included 46 neurons and the output layer contained 1 neuron,with a learning rate to be 0.0056.The SVR optimal model was obtained when penalty coefficient C was 2.5154 and the kernel function paramet nu was 0.1938.The SVR optimal model’s MAE,RMSE and MAPE were lower than that of the MLP optimal model(0.4094 vs 1.3137,0.4697 vs 1.3159,40.9251vs 136.1673)on the test set.Their predictive value of peak VO2 effect size was-0.801 ml/kg/min and 7.694 ml/kg/min,respectively.And with increasing sample size,the peak VO2 effect size predicted by MLP optimal model tended to decline,while that predicted by SVR optimal model was towards rising.The former did not show the minimal sample size for achieving the target outcome of BESMILE-HF(i.e.,Baduanjin exercise resulting in better peak VO2 improvment than the conrtol).Whereas the latter revealed that 60 is the minimal sample size to obtaining the target outcome of BESMILE-HF,and at which,the peak VO2 effect size was 1.379 ml/kg/min.3.When developing the prediction model of quality of life in the BESMILE-HF study,five important characteristics,including the sample size of each dataset,proportion of Baduanjin intervention,consuming Chinese medicine during the trial,consuming angiotensin converting enzyme inhibitor and angiotensin Ⅱ receptor blocker(ACEI/ARB)at baseline,and the baseline left ventricular ejection fraction(LVEF),were screened out through the decision tree.The MLP optimal model was 5-91-1,in which the model input layer contained five neurons,the hidden one layers included 91 neurons and the output layer contained one neuron,with a learning rate to be 0.0442.The SVR optimal model was obtained when penalty coefficient C was 9.9607 and the number of support vectors nu is 0.4447.The SVR optimal model’s MAE,RMSE and MAPE were lower than that of the MLP optimal model(1.8370 vs 3.9570,1.8997 vs 3.9817,74.2136 vs 160.0733)on the test set.Their predictive effect size of MLHFQ’s total scores was 22.769 and 6.902,respectively.And both predicted MLHFQ’s total scores effect size was increasing with sample size growing,however,not revealing the minimal sample size for achieving the target outcome of BESMILE-HF.ConclusionAccompanying for clinical consultation,duration of HF,arrival means to the site,and clinical trial participation experience were the important factors that influenced patients’decisions on participanting in the BESMILE-HF study.No requirement of accompanying for clinical consultation and duration of HF less than 5 years was independent factor for participating in the BESMILE-HF study.It indicated that researchers should strengthen the recruitment among patients without requirement of accompanying for clinical consultation and those with HF duration less than 5 years.There are several vital method to improve HF patients recruitment for the BESMILE-HF study,such as having a deep conversation on the BESMILE-HF study with patients and their caregivers,highlighting the non-pharmacological intervention of the BESMILE-HF study,helping book parking space and providing subsidies for parking fee,inviting patients who had participated in clinical trials to share their experience,optimizing the chronic disease management system for HF.BESMILE-HF’s outcomes in target sample size(n=200)predicted by SVR optimal model are more accurate than that predicted by MLP optimal model.In addition,the SVR optimal model for peak VO2 showed that patients in Baduanjin exercise group demonstrated better peak VO2 improvement than those in the control group when the sample size increasing to 60.However,the predicted outcomes from optimal model developed by limited data were heterogeneous.Thus,the decision that finishing BESMILE-HF when the sample size reaching 60 should be prudently making,and even giving up.The accuracy of the outcome prediction models of the BESMILE-HF study remained to be promoted by using more labeled samples to train.Keeping the BESMILE-HF study moving forward should focus primarily on improving recruitment strategy according to the factors which can impact patients to participate in.To our knowledge,this is the first study that focused on methodological research on improving recruitment strategy and outcome prediction in TCM,to explore the research delay and barriers and its solutions when conduction TCM clinical trials.The present study is the first to try to apply machine learning to improve the implementation of TCM clinical trials.The methods applied in our study can provide deep insights to the similar delayed TCM clinical trials.