Comprehensive Intervention For Type 2 Diabetes And The Assessment Of Individual Treatment Effect In A Real World Community Trail

[BACKGROUND]Type 2 diabetes mellitus(T2DM)as one of the most common chronic diseases in China has caused severe disease burden in recent years.Randomized controlled trials(RCTs)as gold standard of evidence-based medicine have demonstrated the effectiveness of lifestyle and medication intervention on glycemic control,which have been also included into T2DM guidelines as routine intervention strategies.In real world community,however,the implementation of these strategies failed to receive optimal outcomes.Generally,RCTs are precisely designed with strict inclusion and exclusion criteria to recruit a group of homogeneous patients.Real world settings,on the contrary,may differ significantly from RCTs,where many patients are heterogeneously attacked,have multiple comorbidities and with poor adherence.Moreover,the intervention strategies may be not well employed in communities,significantly affecting the intervention outcomes.Therefore,assessing the effectiveness of lifestyle and medication intervention on glycemic control under real world settings and the impact of adherence on their effectiveness is of great public health significance for optimizing the intervention strategies,selecting essential antidiabetic agents and promoting intervention and policy strategies-making for T2DM in communities.In addition,conventional RCTs,with a particular focus on patient populations,attempt to learn average treatment effect(ATE)and determine whether the treatment is effective overall.Yet in medical practices,doctors are faced with individual heterogeneous patient and supposed to choose the optimal medical treatment among a set of alternatives given the patient’s distinct characteristics.Regarding this,real world community trials have the potential to more efficiently provide important information from a broader patient population in respect of their factual disease conditions and medication usage,making it possible to assess the individual treatment effect(ITE)by providing adopted by patients in real-life.Therefore,how to achieve ITE learning using real world data and optimize personalized treatment accordingly remain to be further studied.Driven by this,emulating a hypothetical randomization design and an analogue casual effect using real world non-randomized data has become an increasing concern of researchers.Besides,a variety of approaches targeting ITE learning have been proposed.Of these,uplift modeling is a subfield of machine learning which aims at explicitly predicting differences in outcome probabilities,or,the incremental effect,between treatment and control groups,and thereby allows for identifying individual patient or discovering groups of patients for whom the treatment is most beneficial.Adopting uplift modeling in real world community and estimating its usage in individual evaluation will contribute to making and optimizing individual intervention.[OBJECTIVES]This study designed a T2DM community trial through the T2DM sub-cohort from the Comprehensive Intervention Program of Chronic Diseases in Jiaonan,Shandong.The objectives are as follows.1.To assess the effectiveness of the community-based comprehensive T2DM intervention,including lifestyle and medication intervention,on glycemic control in diabetes patients and evaluate the impact of antidiabetic medication use and adherence to medication on the intervention outcome to inform the promotion of community intervention strategies.2.To emulate a hypothetical RCT study and explore the validity of generating emulated RCTs from real world community trial study through novel techniques,such as propensity score matching(PSM),to contribute to the analytical methodology for real world community intervention trial.3.To detect the ATE of antidiabetic medications based on an emulated RCT,and then estimate ITE using uplift modeling,and identify the benefited individuals from each antidiabetic prescriptions as well as their characteristics,so as to evaluate the availability of uplift modeling in ITE learning and guiding personalized medicine.[METHODS]1.Study design and data source:Comprehensive Intervention Program of Chronic Diseases in Jiaonan,Shandong was implemented by Jiaonan Health Bureau from Jan 1,2011 to Dec 31,2015.Shandong University specially provided technical and statistical supports to this program.The current study designed a T2DM community trial using its T2DM sub-cohort and obtained baseline measurements,intervention and outcome information.Particularly,baseline measurements included demographic,physical,clinical and physiologic parameters.Intervention information indicated the lifestyle and medication intervention and the follow-up visits during the intervention process.Each participant was visited 4 times per year with the visit items of fasting plasma glucose(FPG),blood pressure,smoking,alcohol consumption,exercise status,antidiabetic prescriptions use(e.g.metformin and glipizide)and adherence to medication.Outcome was FPG levels and whether they were controlled at goal.This study recruited 8,499 T2DM patients with 38,557 visits from a total of 15,348 participants in this sub-cohort.2.Assessing comprehensive intervention effect based on the T2DM cohort:Participants were categorized by sex,medication use(users versus non-users)and adherence to medication(adherence group versus nonadherence group).Linear mixed-effect models were adopted to estimate the differences of FPG trajectories between sex and medication subgroups.Outcome was then defined as FPG change per visit(current FPG level-last FPG level)and linear mixed-effect models were adopted to estimate the linear predictions of FPG change per visit and their distributions were displayed and compared between groups using kernel density plots.3.Emulating a hypothetical RCT:We emulated a RCT design based on the follow-up visit per month during Jan 2012 to Dec 2015(48 months in total).For each month,eligible participants were selected according to the following inclusion criteria:1)new registered participants in this month(thus each participant can only be included once);2)exposed to the intervention for at least 6 consecutive months(with prescription of no drugs or any drugs and without adjusting the drug usage);3)with an eligible FPG measurement within half year after 6 months’ intervention.This approach was applied to each of the 48 months and eligible participants were found in 36 months(no eligible participants were found in the left 12 months);thus we emulated 36’trials’and recruited 4,226 participants,of whom 2,723 prescribed with drugs(metformin,glipizide or the combination therapy of metformin and glipizide)were included in the treatment group and 1,503 without prescriptions were included in the control group.Outcome was defined as the first FPG measurement in this half year.Propensity score matching(PSM)was adopted to rebuild treatment and control groups and balance the differences between groups.Specifically,propensity score was calculated by logistic model including T2DM-related factors as matched variables.Participants were exact matched on baseline FPG and nearest-neighbor matched on other covariates in a one-to-one manner.Consequently,832 participants in metformin group and 832 paired participants in control group,190 participants in glipizide group and 190 paired participants in control group,451 participants in combination therapy group and 451 paired participants in control group were matched and selected into analysis4.Assessing medication intervention effect based on the emulated RCT:for the emulated RCT,analysis of covariance(ANCOVA)was adopted to further adjust the matched variables(to address the selection bias originating from the PSM)and compare the difference of change of FPG level(post-intervention FPG level-baseline FPG level)between treatment group and control group.This part of analysis was regarded as the ATE evaluation for three prescriptions.Uplift modeling,with the casual conditional inference forest as estimate algorithm and Euclidean distance as split criteria,was adopted to predict the difference between the treatment and control success probability—defined as whether the FPG outcome was controlled at goal(≤7mmol/L)-on the level of individuals and identify benefited individual and their characteristics.Model performance was assessed by Qini curve and Qini coefficient q.Model validation was performed using 5-fold cross validation procedure.This part of analysis was regarded as the ITE evaluation.[RESULTS]1.Descriptive characteristics of T2DM cohort:8,499 participants were included in analysis with mean age of 64.88 years;37%men,63%women;59%use medication,and 41%did not use medication.Participants who used medication were younger,more smokers and CVD patients,with higher glucose,blood pressure and lipid levels compared those without medication use.Over the follow-up of 4 years,the crude mean decline in FPG from baseline was-0.36 mmol/L(95%CI-0.38 to-0.35)in overall participants;-0.40 mmol/L(95%CI-0.47 to-0.34)in men versus-0.49 mmol/L(95%CI-0.54 to-0.43)in women(P=0.058);-0.56 mmol/L(95%CI-0.62 to-0.50)in medication users versus-0.29 mmol/L(95%CI-0.34 to-0.24)in medication non-users(P<0.001).2.Comprehensive intervention effect based on the T2DM cohort:1)Mean FPG levels of overall participants showed a significant decline trend along with the intervention process.Mean FPG levels of men and medication users were lower than that of women and non-users at each time point,which may due to the discrepant baseline FPG levels.FPG mean change from baseline was increasing along with the process,and no sex difference was observed through adjusted linear mixed-effect model(P=0.101),while medication users showed greater FPG decrease versus medication non-users(P<0.001).Adjusted FPG mean change was-0.124 mmol/L(95%CI-0.125 to-0.122)in medication users versus-0.065 mmol/L(95%CI-0.067 to 0.063)in medication non-users,with the difference of-0.059 mmol/L(95%CI-0.061 to-0.056).2)Kernel density plots suggested FPG change per visit was greater in medication users(median-0.106 mmol/L)than in medication non-users(median-0.049 mmol/L),and in adherence group(median-0.102 mmol/L)than in nonadherence group(median-0.069 mmol/L).3.Emulated RCT:Before PSM,major variables including age,sex,education status,glucose,blood pressure,lipid levels were significantly different between treatment and control group,whereas after PSM a good balance was successfully achieved between metformin group and control groups,glipizide group and control group,metformin&glipizide group and control group.4.Medication intervention effect based on the emulated RCT:1)Difference of FPG mean change between metformin group and control group was-0.04 mmol/L(95%CI-0.09 to 0.16),indicating null effect of metformin intervention(P=0.544)from ATE evaluation;difference of FPG control rate between two groups was 6.97%(95%CI 2.58 to 11.36).ITE evaluation by uplift modeling indicated 60%of the participants could benefit from metformin use.Qini coefficient from cross validation was 0.0266,indicating uplift model performed better than random treating strategy in population.Further recognizing the characteristics of the benefited individuals found those who were younger,better educated,non-smokers,with lower waist circumference,baseline glucose,blood pressure and lipid levels could benefit from metformin use.2)Difference of FPG mean change and FPG control rate between glipizide group and control group were-0.06 mmol/L(95%CI-0.17 to 0.29)and 1.05%(95%CI-7.62 to 9.73)respectively,indicating null intervention effect from neither ATE evaluation(P=0.594)nor rate comparison(P=0.812).ITE evaluation indicated 50%of the participants could benefit from glipizide use,and Qini coefficient from cross validation was 0.0213.Participants who were better educated,non-smokers and with greater waist circumference could benefit from glipizide use.3)Difference of FPG mean change and FPG control rate between metformin&glipizide group and control group were 0.05 mmol/L(95%CI-0.24 to 0.15)and 3.99%(95%CI-1.48 to 9.46)respectively,indicating null intervention effect from neither ATE evaluation(P=0.604)nor rate comparison(P=0.812).ITE evaluation indicated 40%of the participants could benefit from using this combination therapy,and Qini coefficient from cross validation was 0.0245.Men who were better educated,non-smoker and with more exercise could benefit from this combination therapy.[CONCLUSIONS]1.The community-based lifestyle and antidiabetic medication intervention were effective in controlling FPG levels on a population level.Antidiabetic medication and good medication adherence were helpful for glycemic control.2.This study emulated an analogue casual medication treatment effect in a hypothetical RCT from real world community trial through PSM and demonstrated the availability and feasibility of emulating RCT from non-randomized data in real world community trial.3.The ITE evaluation based on three antidiabetic prescriptions indicated ATE evaluation may mask the actual intervention effect.ITE evaluation through uplift modeling could identify benefited individual and their characteristics,and thereby optimize the personalized medicine.