Weight loss patterns with tirzepatide compared to semaglutide

This study analyzed de-identified Electronic Health Record (EHR) data from academic medical centers in the United States via the nference nSights Analytics Platform. Prior to analysis, all data underwent expert determination de-identification satisfying Health Insurance Portability and Accountability Act (HIPAA) Privacy Rule requirements [45 Code of Federal Regulations (CFR) §164.514(b)(1)], employing a multilayered transformation approach for both structured data (cryptographic hashing of identifiers, date shifting, and geographic truncation) and unstructured clinical text (ensemble deep learning and rule-based methods with >99% recall for personally identifiable information detection) (10, 11). nference established secure data environments within each participating center, housing these de-identified patient data governed by expert determination. These de-identified data environments were specifically designed to enable data access and analysis without requiring Institutional Review Board (IRB) oversight, approval, or exemption confirmation. Accordingly, informed consent and IRB review were not required for this study.

From a longitudinal database of 23 million patients, we identified 640,794 with at least one glucagon-like peptide-1 receptor agonist (GLP-1RA) order. Requiring ≥3 orders spanning ≥1 month yielded 278,471 patients; restricting to semaglutide or tirzepatide prescriptions yielded 215,004 patients. We constructed drug-exclusive cohorts comprising patients who received only semaglutide brands (Ozempic, Wegovy, and Rybelsus) or only tirzepatide brands (Mounjaro and Zepbound); patients could switch between brands of the same drug but not between drugs. We excluded patients with bariatric surgery history and required weight measurements during baseline (≤90 days pretreatment), the first year posttreatment (year 1), and the second year posttreatment (year 2), yielding 34,143 semaglutide-treated and 16,034 tirzepatide-treated patients (Fig.). S1

Patients in both semaglutide and tirzepatide cohorts were stratified into five response groups (Fig. 1): high response (>15% weight loss during the entire study period), moderate response (5–15% during the entire study period), late response (<5% during year 1, >5% during year 2), weight regain (>5% during year 1, ≥5% regain during year 2), and minimal weight loss (<5% weight loss during the entire study period). The index date was the date of the first GLP-1RA prescription baseline weight was the measurement closest to the index date within 90 days pretreatment. Weight changes were calculated relative to baseline and aggregated into 30-day windows (±15 days). Population-level trajectories were generated by computing patient-level mean percentage changes within each window, then smoothed using Savitzky–Golay filtering (window = 5, polynomial order = 3). Uncertainty was quantified as SEM (SEM = σ/√n); shaded regions represent ±1 SEM.

To compare treatment responses between tirzepatide and semaglutide while controlling for baseline differences, we performed propensity-score matching for the overall tirzepatide and semaglutide populations as well as separately within each of the five weight-loss outcome categories (high response, moderate response, weight regain, late response, and minimal weight-loss groups). Propensity scores were estimated using multivariable logistic regression with treatment assignment (tirzepatide vs. semaglutide) as the dependent variable and the following covariates: age at index date, sex, type 2 diabetes mellitus (T2DM) status (defined as ≥3 diagnoses within 5 years prior to index), baseline body mass index (BMI; most recent measurement within 90 days before index), baseline weight (most recent measurement within 90 days before index), year of treatment initiation, and follow-up duration in days. Because race itself is not known to be associated with GLP-1RA treatment selection, it was not included as a matching covariate. Nearest-neighbor matching without replacement was performed using a 1:1 matching ratio and a caliper width of 0.1 SDs of the logit of the propensity score. Patients with index dates before January 1, 2023 were excluded to ensure comparable follow-up periods and account for differential market availability of the medications. Balance diagnostics were assessed by comparing standardized mean differences for all matching covariates between matched cohorts within each response category, with values <0.1 considered indicative of adequate balance. We characterized brand name distributions within matched cohorts (Table) among the tirzepatide (Mounjaro and Zepbound) and semaglutide (Ozempic, Wegovy, and Rybelsus) branded formulations. S1

Summary statistics were generated for semaglutide and tirzepatide cohorts across all response groups (Tables S2–S7). Demographics included age at first prescription, gender, and race/ethnicity. Race and ethnicity data were derived from patient self-report documented in the electronic health record. We categorized patients as White (non-Hispanic), Black/African American (non-Hispanic), Hispanic (any race), or other/unknown, with the latter category combining all other racial groups and missing data due to small sample sizes that precluded separate statistical analysis (12). Baseline clinical characteristics included T2DM, BMI, and weight. BMI and weight were defined as the nearest recorded values within 90 days prior to the index date. T2DM status was defined by at least three diabetes diagnostic codes recorded in the 5 years prior to the index date. We characterized weight measurement distributions across a 10-year pretreatment period and the first 2 years posttreatment. GLP-1RA prescription intervals were calculated as the time between the first and last prescription for each patient (Fig. S2), and prescription frequency was defined as the total number of prescriptions received per patient. Follow-up duration was computed as days from index date to last clinical encounter for that patient. Baseline and posttreatment monthly clinical note counts were quantified. Statistical comparisons between propensity-matched tirzepatide and semaglutide cohorts were performed using two-proportion z-tests for binary outcomes and χ² tests or Fisher's exact tests (for cell counts <5) for categorical outcomes.

We employed a multistage information extraction (IE) pipeline using fine-tuned Bidirectional Encoder Representations from Transformers (13) models to extract clinical phenotypes from unstructured clinical notes. The pipeline first applies Named Entity Recognition to identify clinical entities, then uses sequential qualifier models to assess clinical context by determining subject (patient vs. other), temporality (current, past, or hypothetical), and certainty (confirmed, negated, or suspected). This contextualization framework disambiguates confirmed current patient symptoms from negated findings, suspected conditions, or prior diagnoses. This IE approach has been validated in prior studies (8) with F1 scores of 0.93 and 0.95 for conditions and medications, respectively.

We performed a comparative analysis of adverse-event prevalence between propensity-matched tirzepatide and semaglutide cohorts within each weight-loss response category. We analyzed 16 commonly reported adverse events identified from Food and Drug Administration prescribing information for both medications, including gastrointestinal, systemic, and metabolic events. Adverse-event prevalence was determined exclusively from unstructured clinical notes during the 2-year posttreatment period using the validated IE pipeline described above. The pipeline identified adverse-event mentions from all available clinical documentation, restricting output to confirmed, current patient-attributed events while excluding negated, suspected, or historical mentions. For each adverse event within each response group, we compared the prevalence in tirzepatide versus semaglutide cohorts using χ² tests for adequate cell counts (≥5 expected events per cell) or Fisher's exact test for smaller counts. We performed 16 independent tests within the high response, moderate response, and minimal weight-loss categories separately and reported unadjusted P-values at P < 0.05 (Fig. 2).

From initial cohorts of 34,143 and 16,034 patients with at least three prescriptions for semaglutide or tirzepatide, respectively, we created 1:1 propensity-matched cohorts of 10,339 patients each (see Methods). The standardized mean differences for all variables used in matching were <0.1, indicative of adequate balance. Over 2 years of follow-up, tirzepatide was associated with markedly greater maximal weight reduction than semaglutide (Fig. 1, Tables S2–S4). The mean of maximal 2-year weight reduction was 14.7% (16.0 kg) for patients treated with tirzepatide compared with 10.8% (11.6 kg) for patients treated with semaglutide (P < 0.001, two-proportion z-test, Fig. 1a). Further, the proportion achieving their highest recorded weight loss of ≥5% at any time point was 90% with tirzepatide and 82% with semaglutide (P < 0.001). Differences widened at higher thresholds: 68 versus 47% achieved ≥10% loss (P < 0.001), and 43 versus 22% achieved ≥15% loss (P < 0.001), respectively (Fig. 1a–c). Maximal reductions of ≥20% were observed in 26% of tirzepatide-treated patients compared with 12% of semaglutide-treated patients (P < 0.001), and ≥25% reductions were observed in 14 versus 5% (P < 0.001; Fig. 1a–c). We compared the rate of weight loss between tirzepatide and semaglutide across all the weight-loss groups (Fig. 1d–h). During the first 6 months after treatment initiation, tirzepatide was associated with more rapid weight reduction than semaglutide in the high-response group (Fig. 1d). The mean monthly rate of weight loss was 2.54% for tirzepatide and 2.18% for semaglutide (P < 0.001) in the high-response group. Taken together, these findings indicate that compared with semaglutide tirzepatide produced overall higher weight loss and among high responders faster weight loss.

We evaluated the weight-loss response distributions by race/ethnicity and sex for both tirzepatide and semaglutide (Table S2). White individuals were more represented in the high-response group than the minimal response group for both tirzepatide (91.2 vs. 79.8%, P < 0.001, two-proportion z-test) and semaglutide (91.2 vs. 79.1%, P < 0.001). On the contrary, Black patients were more represented in the minimal weight-loss group than the high-response group for both tirzepatide (10.9 vs. 6.0%, P < 0.001) and semaglutide (10.6 vs. 5.8%, P < 0.001), with a similar pattern observed among Hispanic patients for tirzepatide (3.3 vs. 1.6%, P < 0.001) and semaglutide (4.1 vs. 1.3%, P < 0.001). Evaluating response distribution by sex, females were more represented in the high-response group than the minimal response group for both tirzepatide (74.9 vs. 59.2%, P < 0.001) and semaglutide (80.3 vs. 58.5%, P < 0.001).

We compared the prevalence of 16 commonly reported adverse events between response-group–specific propensity-matched cohorts. The baseline and follow-up characteristics of the cohorts are summarized in Tables S5–S7. In the high-response group, the posttreatment prevalence of multiple gastrointestinal phenotypes was significantly lower among patients treated with tirzepatide compared with semaglutide, including nausea (27.2 vs. 31.4%, P = 0.002), vomiting (12.8 vs. 17.1%, P < 0.001), constipation (16.9 vs. 20.2%, P = 0.005), diarrhea (16.9 vs. 19.2%, P = 0.04), and dyspepsia (0.7 vs. 1.6%, P = 0.006; Table S8). Additionally, tirzepatide-treated patients in the high-response group had lower prevalence of fatigue (26.1 vs. 29.7%, P = 0.007) and headache (23.5 vs. 28.4%, P < 0.001; Table S8). In the moderate-response groups (n = 2,962), tirzepatide was also associated with a lower posttreatment prevalence of nausea (16.4 vs. 19.4%, P = 0.003), vomiting (7.8 vs. 9.6%, P = 0.02), and dizziness (8.9 vs. 11.0%, P = 0.007; Table S8).