WEIGHT CHANGES FROM GLUCAGON-LIKE PEPTIDE-1 RECEPTOR AGONIST (GLP-1) USE AND DISCONTINUATION: A RETROSPECTIVE COHORT STUDY OF ELECTRONIC HEALTH RECORDS

Abstract

Objective:

To quantify the effect of GLP-1s on weight loss, and weight gain following GLP-1 discontinuation

Methods:

This retrospective cohort study using electronic health records from TriNetX and included individuals age 18 years or older with various lengths of continuous GLP-1 use between September 2014 and November 2023. Applying an intention-to-treat framework, we used linear mixed effects models and propensity score adjustment to model changes in body mass index (BMI) after GLP-1 discontinuation while accounting for correlated within-person observations and potential confounders.

Results:

A total of 78,076 (19.7%) individuals used GLP-1s continuously for three months, 23,861 (6.0%) for 6 months and 12,642 (3.2%) for 9 months. The median (interquartile range [IQR]) BMI prior to initiation was 36.38 (36.30–36.46), which declined by 2.15% (IQR, 2.10–2.21) to 35.81, (IQR, 35.7–35.92) among individuals with three months of continuous GLP-1 use. Median BMI declines were greater among individuals with 6 months (4.38%, IQR 4.27–4.48%) and 9 months (5.56%, IQR 5.34–5.78%) of continuous use. On average, weight loss among discontinuers at months 3, 6 and 9 slowed in comparison to their counterparts.

Conclusion:

Among this real-world sample, GLP-1 use was associated with more modest weight loss than has been demonstrated in randomized controlled trials.

INTRODUCTION

Since exenatide was first introduced in the U.S. marketplace in April 2005, glucagon-like peptide-1 agonists (GLP-1s) have garnered enormous interest as a treatment for obesity.^1,2 In one systematic review and meta-analysis of randomized, placebo-controlled trials, 50.2% (17.5%) of GLP-1 treated individuals, compared with 17.1% (3.1%) of placebo recipients, achieved weight loss of greater than 5% (10%).³ Other randomized evidence also supports conclusions regarding GLP-1 efficacy for weight loss.^4,5,6

Given important differences between randomized trials and real-world use, it’s noteworthy that there has also been a growing body of observational research suggesting GLP-1 effectiveness in reducing weight.^7,8 For example, in a retrospective analysis of electronic health information from a cohort of 258 Italian individuals with diabetes, approximately three-fourths of individuals treated with once-weekly semaglutide achieved weight loss at six and twelve months, with weight loss of ≥10% of BMI occurring in 6.8% (18.2%) after 6 (12) months of treatment.⁹ In another analysis of U.K. individuals with diabetes using injectable GLP-1s with a median baseline BMI of 41.2 kg/m2, weight loss of ≥5% of BMI was achieved by 33.4% (12 months) and 43.5% (24 months).¹⁰

Despite the insights from these studies, their sample have typically been from a single centers or health systems. In addition, little is known regarding the real-world impact of GLP-1 discontinuation on weight regain, which is an important scientific and clinical question since discontinuation is common¹¹ and most individuals will not remain on these medicines lifelong.^12,13 Weight regain upon discontinuation of anti-obesity medicines is expected,^14,15,16 and has been noted in both the STEP 1 extension trial of semaglutide¹⁷ and the SURMOUNT-4 trial of tirzepatide.¹⁸ We performed a retrospective cohort study using a database from a health research network in the United States, using a causal framework to quantify the effect of GLP-1s on weight loss, as well as to characterize weight gain following GLP-1 discontinuation.

METHODS

Data source

We used data from the TriNetX database, a large database comprised of electronic health information derived from 81 academic medical centers in the United States.¹⁹ The data, which includes longitudinal, patient-level electronic medical records, including demographics, diagnoses, medications, and laboratory values, represents approximately 8.5 million individuals and was queried on November 30, 2023.

Cohort

We derived a cohort of adults who received at least one prescription for any of the six GLP-1s that were approved by the U.S. Food and Drug Administration (FDA) at the time of our cohort derivation (Semaglutide, Liraglutide Trizeptide, Exenatide, Lixisenatide and Dulaglutide). For each individual, we defined their index date as the first date of the first GLP-1 prescription that we observed, which we required to have occurred between September 2014 and November 2023.

We then sequentially excluded: (1) individuals less than 18 years of age on their index date; (2) individuals missing information about sex; (3) individuals with a home address outside of the United States; and (4) individuals with a non-sensical date of GLP-1 receipt (i.e., receipt on a date prior to the product’s FDA approval). Of the remaining individuals, we then excluded those without a BMI measured within six months prior to their index date and those without a BMI measured within 12 months after their index date. We also excluded individuals using a GLP-1 for less than three continuous months, as well as those who received a prescription for another FDA-approved anti-obesity medication within 12 months of their index date, including naltrexone/bupropion (Contrave), orlistat (Xenical), phentermine/topiramate (Qysmia), or setmelanotide (Imcivree).

Exposure and outcome

Our exposure was GLP-1 use. For each month of follow-up, we assigned each individual as either having GLP-1 on hand or not. Individuals who received treatment for fewer than or equal to 15 days within a given month were classified as not being treated for that month. Since nearly all prescriptions with a days of supply were for 30 days, we assumed those with missing information about days supply were also for 30 days. We considered prescription written within 3 days of one another as duplicate transactions, while in all other cases, we allowed for indefinite stockpiling when calculating where or not pills were on hand.²⁰ While our index date reflected the date of the first GLP-1 initiation, we also derived a discontinuation date, which we defined as the date when GLP-1s were no longer on hand, as long as it was followed by no fewer than 30 days of such non-use (eFigure 1).

Outcome

Our primary outcome was percentage change in BMI from baseline. We defined our baseline BMI as the most recently assessed BMI within six months prior to each individual’s index date. For each month, we derived their BMI using both the weight and height of a given person to calculate BMI, along with the BMI measurements. In settings where individuals had two or more BMI’s in a given month, we derived an average BMI.

Covariates

We examined baseline variables including individuals’ sociodemographic (age, sex, race, ethnicity, marital status, regional location) and clinical factors (comorbidities, comedications) in our analysis. We identified 17 comorbidities within 1 year before the index date and used Charlson weights to calculate the Charlson Comorbidity Index (CCI).²¹ We identified sulfonylureas, metformin, DPP-4 inhibitors, SGLT2 inhibitors, and insulin as concomitant medications if they were received within 1 year before the index date. All covariates were treated as time-fixed in the analysis.

Statistical Analysis

We characterized the overall cohort using percentage for categorical variables and mean with SD, or median with interquartile range (IQR) when appropriate, for continuous variables. We compared the effect of continued vs. discontinued GLP-1 medications among individuals who consecutively took the medications for 3, 6, and 9 months using longitudinal structural mixed models³⁸ with multiple imputation.³⁹ This study examines the average difference in the longitudinal progression of BMI percentage over time between the two intervention options (discontinuation or continuation) as the target causal estimate. The framework employs a structural model for longitudinal outcomes and a propensity model that describes the baseline and time-dependent factors influencing the discontinuation of GLP-1 medications, thereby characterizing the complete longitudinal data process and quantifying the causal target parameter as a function of the model coefficients. In addition, to address the missing data problem, which is common in observational studies, we integrate the longitudinal causal framework with a multiple imputation scheme such that the uncertainty induced from the missing BMI is also accounted for. Given that this was an intent-to-treat analysis, we focus on the effect of medication discontinuation in a given month (3, 6 and 9). Individuals may change their treatment exposure in the subsequent follow-up period, i.e., discontinuous users may re-initiate treatment, or continuous users may stop treatment. We used descriptive statistics to characterize this treatment pattern.

Without loss of generality, we describe in more detail the method for studying the discontinuation effect for individuals who continuously took GLP-1 for 3 months then either continued vs. discontinued treatment in month 4. The discontinuation effect was evaluated by comparing across the two groups, stratified by their propensity to discontinue, the trajectory of change in BMI percentage relative to baseline between month 4 and month 12 (eMethods). Specifically, we imputed missing BMI percentages relative to baseline for each month before month 4 using a linear mixed effects model. We modeled BMI percentage between month 1 and month 3 as a function of all baseline covariates and natural splines of time. From this model, we conducted 5 rounds of random imputation for missing BMIs which generated 5 datasets with the complete trajectory of BMI percentage for all patients between month 1 and month 4. Second, for each of the 5 imputed datasets, we used gradient boosting trees (xgboost) to calculate propensity scores, i.e., the probability of continuing vs discontinuing GLP-1 at month 4 conditional on all baseline characteristics and the imputed complete dynamic history of BMI percentage. Third, we fit a linear mixed effects model for the observed longitudinal trajectories of BMI percentage between month 1 and month 12 with indicator of GLP-1 discontinuation, natural splines of time, the discontinuation and time splines interaction, and the estimated propensity scores as independent variables. This was done separately for each of the 5 datasets. Based on the treatment effect coefficients, we calculated the estimated average treatment effect as the difference in the trajectory of BMI percentage between the two groups. We averaged the estimated treatment effect obtained from the 5 sets of regression and derived its uncertainty estimates for the averaged effect via Rubin’s rules for multiple imputation. Finally, we calculated the average treatment effect defined as the difference in BMI percentage for each month under the two treatment regimes, discontinuation versus no discontinuation at each month. Because the average treatment effect reflects the average difference in trajectories between the two groups, it does not directly explain how BMI percentages progress over time for the two groups. To address this, we provided the group-specific smoothed trajectories of the change in average BMI percentages from baseline.

We performed similar analyses for individuals who continued GLP-1 for 6 months and continued or discontinued on month 7, and those who used GLP-1 for 9 months and continued or discontinued on month 10.

Given that weight is frequently used as a metric, we quantified weight loss as a percentage of the initial body weight in a subset of patients for whom weight data were available. To explore the heterogeneity of treatment effect, we performed stratified analysis by GLP-1 drugs, by diabetes status and by both variables among individuals who continued GLP-1 for 6 months and continued or discontinued on month 7.

All analyses were conducted using SAS (version 9.4) and R (version 4.4.2). The study was approved by the Johns Hopkins Medicine Institutional Review Board. Patients or the public were not involved in the design, or conduct, or reporting, or dissemination plans of our research

RESULTS

Cohort derivation

Our initial sample included 1.1 million individuals, of which approximately 1 million were included as adult users of GLP-1s in the United States. Of these, we excluded 598,077 individuals without a BMI during the six months prior to GLP-1 initiation, as well as 32,992 individuals without a BMI within one year of GLP-1 initiation. Of the remaining 403,556 individuals, we further excluded 7,583 using other FDA-approved anti-obesity treatments, and 317,897 individuals discontinuing GLP-1s within 3 months of initiation, leaving 78,076 individuals in our final cohort (Figure 1). A total of 78,076, 23,861, and 12,642 individuals continuously used GLP-1s for 3, 6 and 9 months, respectively. The percentage of missing BMI ranged from 55% to 80% across the three cohorts over months 4 through 12 (eFigure 2).

Figure 1.

Flowchart depicting cohort derivation.

Characteristics of individuals

Overall, the mean age was 54.3 years (standard deviation [SD] ±13.4 years), 64.3% were female, 65.2% were White and 10.9% were LatinX. The mean Charlson Comorbidity Index was 2.6 (SD 3.0), approximately two-thirds (65.8%) of the cohort had diabetes, and approximately two-fifths of individuals (44.4%) used metformin during the study period (Table 1).

Table 1.

Baseline characteristics of individuals with 3 months of continuous GLP-1 use for intention-to-treat analyses (n=78,076).

Characteristics	Total (n=78,076)	Continuous users (n=48,956)	Discontinuous Users (n=29,120)	SMD
Age (mean, standard deviation)	54.3 (±13.5)	54.2 (±13.4)	54.4 (±13.6)	0.019
Sex, %				0.020
Female	64.3%	64.7%	63.7%
Male	35.7%	35.3%	36.3%
Race, %
White	65.2%	66.6%	62.8%	−0.080
Black or African American	20.1%	19.0%	21.8%	0.070
Asian	3.1%	3.2%	3.0%	−0.010
Native Hawaiian or Other Pacific Islander	0.8%	0.8%	0.9%	0.010
American Indian or Alaska Native	0.5%	0.5%	0.5%	−0.005
Other race	4.0%	3.8%	4.3%	0.024
Unknown	6.3%	6.0%	6.7%	0.028
Ethnicity, %
Hispanic or Latino	10.9%	11.4%	10.1%	−0.043
Not Hispanic or Latino	78.2%	79.7%	75.6%	−0.099
Unknown	10.9%	8.9%	14.3%	0.170
Marital status, %
Married	36.5%	36.1%	37.2%	0.022
Single	25.0%	24.6%	25.8%	0.029
Unknown	38.5%	39.3%	37.0%	−0.048
Regional location, %
Northeast	12.5%	10.7%	15.6%	0.146
Midwest	27.0%	27.6%	25.9%	−0.040
South	46.4%	47.5%	44.6%	−0.058
West	14.1%	14.2%	13.9%	−0.009
CCI (mean, standard deviation)	2.6 (±3.0)	2.7 (±3.1)	2.4 (±2.8)	−0.100
Diabetes, %	65.8%	65.2%	66.7%	0.032
Medication, %
Metformin	44.4%	45.8%	42.0%	−0.077
Sulfonylureas	17.8%	18.6%	16.6%	−0.050
DPP-4 inhibitors	11.8%	12.4%	10.9%	−0.046
SGLT2 inhibitors	11.4%	11.4%	11.4%	−0.000
Insulin	30.6%	30.6%	30.5%	−0.002

CCI Charlson Comorbidity Index; SMD standardized mean difference

Table 1 also depicts characteristics of continuers and discontinuers after three months of continuous treatment using an intention-to-treat framework, suggesting that with the exception of a few characteristics such as unknown ethnicity, northeast location, and mean number of Charlson comorbidities, these two groups were similar with respect to observed characteristics, with standardized mean differences (SMDs) less than 0.1.

Use of GLP-1s

Figure 2 depicts changes in the use of GLP-1s over time among continuers and discontinuers after three months of continuous GLP-1 use. Among individuals who continued GLP-1 at month 4 (n=48,956), two-thirds continued GLP-1 at month 5 and only 18% continued at month 12, indicating a significant discontinuation over time. Likewise, among individuals who discontinued at month 4 (n= 29,120), 54% were discontinuers by month 12, indicating nearly half restarted treatment (Figure 2).

Figure 2.

Change in GLP-1 use over time among individuals who used GLP-1 for 3 continuous months, stratified by continuation and discontinuation status

There were substantively similar patterns of stopping among continuous users and starting among discontinuous users in the 6- and 9-month GLP-1 cohorts (eFigure 3).

Effect of GLP-1 use on weight loss

The effects of GLP-1 use after three, six and nine months is depicted in Figure 3. For example, Panel 3A depicts the effects of GLP-1 use among individuals who used the product continuous for three months. The median BMI among such individuals at baseline was 36.38 (95% confidence intervals [CI] 36.30–36.46). After three months of treatment, there was a 2.15% reduction in BMI (IQR 2.10–2.21%).

Figure 3.

Effect of GLP-1 continuation and discontinuation on weight loss

ATE Average treatment effect. This represents the difference in the trajectory of BMI percentage between the two groups, i.e., continued GLP-1 vs. discontinued GLP-1.

Analogous information is presented for individuals using GLP-1s continuously for six (Panel 3B) and nine (Panel 3C) months. Individuals using GLP-1s for longer periods of time achieved greater weight loss – at six months the median weight loss was 4.38% (IQR 4.28–4.48%) of BMI, whereas at nine months the corresponding values were 5.56% (IQR 5.34–5.78%). Among the 74,419 individuals where weight data was available, results for average weight loss at 3-months (2.15%), 6-months (3.98%) and 9-months (4.55%) were largely similar to the primary analysis where BMI was used as a metric (eFigure 4).

Among individuals who used GLP-1 continuously for 6 months, those using semaglutide (average weight loss was 5.5%) and tirzepatide (weight loss was 10%) had greater weight loss compared to those using liraglutide, dulaglutide, lixisenatide or exenatide (weight loss ranging from 0.1% to 3.2%) (eFigure 5). Likewise, the effect of GLP-1 use on weight loss was greater among individuals without diabetes (weight loss was 6.1%) compared to those with diabetes (weight loss was 2.9%) (eFigure 6). When we stratified by both diabetes and GLP-1 drugs, we found the effectiveness of individual GLP-1 drug was greater among those without diabetes compared to those with diabetes (eFigure 7).

Effect of GLP-1 discontinuation on weight regain

Panels 3D to 3F allow for a contrast between individuals discontinuing and continuing GLP-1s at a given time point in this intention-to-treat framework. Among individuals who used GLP-1 for 3 months, the average reduction in BMI at month 12 was 3.7% among discontinuers vs 4.5 % among continuers with an average treatment effect of 0.86 units (CI, 0.72–0.99) (Panels 3A and 3D). The difference between discontinuers and continuers was relatively small in 6-months cohort (Panels 3B and 3E), and nearly null in 9-months cohort (Panels 3C and 3F). Similar findings were noted when we used weight as a metric (eFigure 4).

DISCUSSION

Despite a surge of interest in GLP-1s for the treatment of obesity, their real-world effect on weight has not been well characterized, nor is it clear how much weight is regained upon GLP-1 discontinuation. In this retrospective cohort study of a U.S. electronic health record repository, the median (interquartile range) body mass index (BMI) decreased 2.15% (2.10–2.21%), 4.38% (CI 4.27–4.48%) and 5.56% (CI 5.34–5.78%) after 3, 6 and 9 months of continuous use, respectively. Many discontinuers at a given month reinitiated GLP-1s subsequently, and thus discontinuers regained weight, though the average weight of discontinuers remained less than at baseline. In contrast to several other observational studies of the effect of GLP-1s,^22,23,24,25 we used robust causal inference methods to quantify the effect of discontinuation on weight loss. Our findings are important because of how commonly GLP-1s are prescribed, their costs to individuals and health systems, and persistent gaps in understanding their real-world effectiveness.

Our estimate of GLP-1 effectiveness using an intention-to-treat framework – approximately 4.4% after 6 months of continuous use – is consistent with randomized evidence. For example, in a pivotal trial of liraglutide, the ADJUMCT 2 trial, individuals in the liraglutide group losing between 3.0% and 6.1% (2.5 to 5.1 kiligrams [kg]) of baseline weight at week 26, as compared with 0.2% (0.2 kg) among the placebo group.²⁶ While other well-controlled trials have yielded larger treatment effects,^27,28 such trials have highly controlled settings, selective participant inclusion, and rigid protocols that limit extrapolation to real world settings. In addition to randomized evidence, our findings are also consistent with other real-world studies of GLP-1 effectiveness, some of which suggest a mean reduction in weight of 2% at six months to 3% at one year.^29,30 While one observational study has yielded 6-month weight loss estimates as high as 11%, the analysis was based on 175 patients from a single weight management clinic.³¹ There is substantial evidence that the effectiveness of GLP-1 medications in producing weight loss varies across products.³² By contrast, while individuals in our analyses without diabetes had greater weight loss, several randomized trials examining heterogeneity of treatment effects among those with and without diabetes found similar effectiveness across these two groups.^33,34 We also observed greater effectiveness in weight loss with newer drugs such as semaglutide and tirzepatide compared to other GLP-1 receptor agonists, a finding that is consistent with current evidence from both randomized trials and real-world studies.

Our findings regarding weight gain with GLP-1 use are also important, since many individuals who are prescribed these treatments will not take them long-term,³⁵ whether due to coverage and reimbursement restrictions, intolerance, or other reasons.³⁶ In the STEP 1 trial that randomized individuals with overweight and obesity to once weekly semaglutide versus placebo, plus a lifestyle intervention, participants who stopped treatment regained approximately two-thirds of their lost weight within one year after stopping, though still maintaining a 5.6% reduction from baseline weight.^{37 38} Similarly, within 12 weeks of discontinuing liraglutide, pre-diabetic patients had a mean weight regain of 1.9 kg, accounting for about 25% of their lost weight (weight loss pre vs post discontinuation: 7.5kg vs 5.6kg).³⁹ Weight regain upon discontinuation of anti-obesity treatments has also been demonstrated with non-GLP-1 therapies. For example, a prospective cohort study showed significant weight gain in patients who discontinued topiramate, returning to pretreatment levels within six months.⁴⁰ Similarly, in a randomized controlled trial assessing the efficacy of orlistat, a gastrointestinal lipase inhibitor, the authors noted “a marked rebound effect” among those stopping orlistat therapy, with individuals switched to placebo regaining twice as much weight as those continuing on treatment.⁴¹

Our study has limitations. First, we had varying information across individuals regarding our key outcome of interest, body mass index. We used several approaches to maximize our ability for causal inference in this setting, including: requiring a minimum number of measures for cohort entry, performing multiple imputation to address item missingness, and employing longitudinal structural models. Second, our data did not include information about the number of days of supply for most GLP-1 prescriptions. Because of this, based on data that was available for a subset of prescriptions, knowledge that insurers may impose 30-day supply limits,⁴² and the frequency of 30-day supply for retail pharmacy fills,⁴³ we assumed that each fill was for a 30-day supply. Third, individuals may have dropped out of our open, real-world cohort – or switched between treatment and non-treatment - for a number of reasons. Therefore, our intention to treat analyses assigned individuals to a treatment group based only on their initial treatment (or non-treatment), which provided the most conservative treatment effect estimate. There could be a variety of treatment patterns for GLP-1 use and discontinuation. For example, individuals can use GLP-1 continuously for three months in a row, or three months spread out over a year (i.e., one month every four months), or any number of other patterns. Although we could perform descriptive analysis among these groups of people and calculate weight loss, the results cannot be interpreted causally. Therefore, we did not do such analysis, and future research should explore as-treated effect of GLP-1 continuation vs. discontinuation by performing careful longitudinal modelling of exposure. Fourth, although we controlled for observable variables using propensity score analysis, we cannot rule out the possibility of unobserved differences affecting continuation vs. discontinuation of GLP-1s. Finally, because this is a real-world data, we do not know if someone truly discontinued GLP-1 versus data is simply missing in the EHR.

CONCLUSIONS

GLP-1s have not only transformed the treatment landscape of Type 2 diabetes, they have also generated enormous enthusiasm because of their clear efficacy in producing weight loss among those with obesity. In this retrospective cohort study of nationwide U.S. electronic health data using an intention-to-treat framework, we find more modest evidence of effectiveness than the efficacy demonstrated in randomized, placebo-controlled trials. We also demonstrate weight regain following GLP-1 discontinuation, underscoring the importance of maximizing access to therapy, and adherence, among individuals in whom GLP-1s are strongly indicated.