Switching patterns of GLP-1 receptor agonists from 2018 to 2025 in the largest private healthcare network in Poland Authors: Krzysztof Łupina, Artur Dziewierz, Jakub Janczura, Zbigniew Siudak Published: 2026 ABSTRACT AbstractAimsTo characterize switching among GLP-1 receptor agonists (GLP-1 RAs) in a large private-sector cohort in Poland and to quantify therapy- and patient-level associations with switching while accounting for switching opportunity and calendar-time dynamics.MethodsWe conducted a retrospective analysis of GLP-1 RA prescription records from the LUX MED network (2018–2025). Switching was defined as any change in agent between consecutive prescriptions. Patients with more than one prescription were included (n = 42,423). The primary analysis used a transition-level discrete-time hazard model in which each prescription-to-prescription interval contributed one observation, and the outcome was switching at that interval. Current-therapy contrasts were reported relative to subcutaneous semaglutide. Sensitivity analyses examined alternative temporal parameterizations and additional adjustment for elapsed time.ResultsOverall, 29.7% of patients switched at least once and 14.3% switched two or more times. In the transition-level analysis, 12,620 patients contributed 27,095 transitions. After adjustment for opportunity and calendar time, liraglutide was associated with substantially lower odds of switching compared with subcutaneous semaglutide (OR, 0.02; 95% CI, 0.01–0.03), whereas oral semaglutide (OR, 1.30; 95% CI, 0.78–2.17) and dulaglutide (OR, 1.70; 95% CI, 0.95–3.04) did not differ significantly. Temporal analyses revealed peaks consistent with episodic substitution and accelerated tirzepatide uptake after market entry. The principal associations remained directionally consistent in sensitivity analyses.ConclusionsSwitching among GLP-1 RAs is common and time-dependent. Time-aware modelling identified therapy-specific switching patterns and pronounced temporal variation; reasons for switching remain unmeasured, and the observed associations should be interpreted as hypothesis-generating.Supplementary InformationThe online version contains supplementary material available at 10.1007/s00592-026-02709-1.Keywords: Glucagon-like peptide-1 receptor agonists, Tirzepatide, Treatment switching, Real-world evidence, Private healthcare, Obesity Introduction The rapid expansion of incretin-based therapies has reshaped the pharmacological management of obesity and type 2 diabetes mellitus [1]. Glucagon-like peptide-1 receptor agonists (GLP-1 RAs) are now central to the treatment of both conditions, providing clinically meaningful improvements in glycemic control and body weight reduction [1–4]. The therapeutic landscape has evolved quickly with the introduction of more potent agents, including semaglutide and the dual glucose-dependent insulinotropic polypeptide (GIP)/GLP-1 receptor agonist tirzepatide, which have demonstrated superior efficacy in randomized controlled trials [1, 5, 6]. Despite robust efficacy in controlled settings, real-world use of GLP-1 RAs is frequently limited by poor long-term persistence [7–10]. In Poland, available population-level data suggest low treatment continuation, which may attenuate expected clinical benefits [6]. Moreover, uptake for weight management among individuals without diabetes remains relatively low, reflecting a complex interplay of access barriers, reimbursement limitations, and patient preferences [3, 11–14]. These observations underscore that treatment trajectories in routine clinical practice often diverge substantially from those observed in trials [1, 3, 7, 10, 12]. As the number of available agents and formulations continues to grow, questions regarding treatment sequencing, persistence, and real-world patterns of use have become increasingly relevant [1]. Within this context, therapeutic switching between GLP-1 RAs represents an important but understudied phenomenon [15]. Switching is conceptually distinct from discontinuation: it reflects a modification of therapy within the incretin class and may arise in relation to treatment experience, formulation preferences, or external prescribing constraints [10, 15, 16]. Prior real-world evidence indicates that switching is common and may serve as a pragmatic alternative to discontinuation or treatment intensification [12, 15]. However, local switching patterns - particularly in settings where prescribing is strongly shaped by formulation availability and market dynamics - remain poorly characterized [17]. In Poland, GLP-1 RA selection may be influenced by the availability of specific formulations (including oral vs. subcutaneous preparations), evolving reimbursement policies, and supply fluctuations [11, 13]. Recent work has described distinct preference patterns for oral versus subcutaneous semaglutide that appear sensitive to reimbursement status and market conditions [13]. Nevertheless, large-scale analyses characterizing GLP-1 RA switching and its temporal dynamics in a private, predominantly nonreimbursed healthcare setting are lacking. We therefore aimed to characterize GLP-1 RA switching trajectories in a large cohort from the LUX MED network, a major private healthcare provider in Poland. Using prescription sequences spanning 2018 through 2025, we quantified switching frequency, described therapy-to-therapy transitions, and evaluated associations between current therapy, patient characteristics, and switching while explicitly accounting for switching opportunity and calendar time. This approach provides a real-world description of how incretin-based therapies are used in clinical practice and how switching patterns vary over time, including after the market entry of newer agents. The study was designed to characterize switching structure and temporal patterns rather than to determine causal drivers of switching or optimal therapeutic sequencing. Methods This retrospective cohort study used anonymized electronic health record (EHR) data from LUX MED, a large private healthcare provider in Poland, spanning January 1, 2018, through December 31, 2025 [8, 11, 13, 18]. Data were extracted from the enterprise EHR system, which covers approximately 300 outpatient clinics nationwide. We assembled a utilization cohort of adults (aged ≥ 18 years) with at least two recorded GLP-1 RA prescriptions, enabling observation of changes between consecutive prescriptions. Switching was assessed from prescription sequences reconstructed using drug-name fields and their corresponding service dates. Because linked pharmacy dispensing data were unavailable and continuous treatment exposure could not be directly observed, the prescription-to-prescription interval was treated as the primary observable decision unit for switching analyses. GLP-1 RA exposure included all formulations available in Poland during the study period: injectable agents (semaglutide [Ozempic/Wegovy], liraglutide [Saxenda/Victoza], dulaglutide [Trulicity], and tirzepatide [Mounjaro]) and oral semaglutide (Rybelsus). Demographic characteristics (age and sex) were determined at the index date (first recorded GLP-1 RA prescription). Comorbidities defined by International Classification of Diseases, Tenth Revision (ICD-10) codes were treated as binary indicators and included type 1 diabetes mellitus (E10), type 2 diabetes mellitus (E11), arterial hypertension (I10), coronary artery disease (I25), atrial fibrillation or flutter (I48), chronic heart failure (I50), obstructive sleep apnea (G47.3), dyslipidemia (E78), chronic kidney disease (N18), and acute pancreatitis (K85); these variables serve as clinical proxies and do not capture disease severity, duration, or treatment indication. Because routine EHR coding may involve overlap between E10 and E11, these diagnoses were not treated as definitive phenotypic classifications but as recorded comorbidity indicators within the source database. The study was designed as a utilization and switching analysis rather than an effectiveness study. Longitudinal efficacy endpoints (e.g., weight change, glycated hemoglobin trajectories, or symptom measures) were not extracted, because outpatient diagnostic reporting is incompletely structured and sporadic, and adherence metrics such as the proportion of days covered could not be calculated without linked pharmacy dispensing data. In addition, the spacing between prescriptions is irregular in routine care and may partly reflect the observation process itself; therefore, switching analyses were interpreted as conditional on recorded follow-up within the network rather than as fully observed continuous-time treatment trajectories. The study protocol (No. 62/2024) was approved by the Bioethics Committee at Jan Kochanowski University in Kielce, Poland. The study was conducted in accordance with the Declaration of Helsinki, and the requirement for informed consent was waived owing to the retrospective design and use of anonymized data. Statistical methods Continuous variables were assessed for distributional characteristics and are presented as mean (SD) or median (IQR), as appropriate. Between-group comparisons used the Welch t test for approximately normally distributed variables and the Mann–Whitney U test otherwise. Categorical variables are reported as n (%) and compared using the chi-square test or Fisher exact test when expected cell counts were less than 5. Because treatment exposure in this database was observed through prescription sequences rather than through dispensing-confirmed continuous use, the primary inferential analysis used a transition-level discrete-time hazard framework in which each prescription-to-prescription interval (Rx_t→Rx_t + 1) contributed one observation, and the dependent variable was switching (yes/no) at that interval. This time scale was selected because the next recorded prescription represents the most reliable observable clinical decision point at which switching can be identified. Current therapy (the “from” agent) was modeled as a categorical exposure, with subcutaneous semaglutide as the reference. The primary model adjusted for switching opportunity using the transition number and for temporal market dynamics using calendar time, modeled with month-of-year fixed effects and a linear time trend. Baseline covariates included age, sex, body mass index (BMI) at the first prescription, type 2 diabetes mellitus, hypertension, dyslipidemia, chronic kidney disease, heart failure, sleep apnea, coronary artery disease, and atrial fibrillation or flutter. Inference used patient-clustered robust standard errors to account for within-patient correlation across multiple observed transitions. Because prescription intervals are irregularly spaced in routine care, the transition-level model was interpreted as estimating switching propensity at the next observed prescription, conditional on continued follow-up within the network, rather than a fully observed continuous-time treatment hazard. Several sensitivity analyses were performed to assess robustness. First, the elapsed time since the index prescription (in days) was added to the transition-level model. Second, calendar time was alternatively modeled as a restricted cubic spline over calendar month to allow for nonlinear temporal dynamics. Third, an additional transition-level sensitivity analysis incorporated both the elapsed time since index therapy and the interval length between consecutive prescriptions to further assess the influence of irregular spacing. Fourth, a Cox proportional-hazards model was fitted for the time to the first observed switch, with censoring at the last recorded GLP-1 RA prescription date. Because the transition-level model is anchored on current therapy and evaluates switching at each observed prescription interval, whereas the Cox model is anchored on index therapy and evaluates the time to the first observed switch, direct numerical concordance between these estimates was not expected — particularly for late-entry therapies with shorter observation windows. All results are reported as odds ratios (OR) or hazard ratios (HR) with 95% confidence intervals (CI). All tests were two-sided, and statistical significance was set at an alpha level of 0.05. Analyses were performed in Python version 3.11.2 (pandas 2.2.3, NumPy 1.24.0, SciPy 1.14.1, statsmodels 0.14.3). Results The study population comprised 54,260 patients with at least one recorded GLP-1 RA prescription. Among the 42,423 patients with more than one recorded prescription (switch-observable cohort), 29.7% experienced at least one switch and 14.3% had two or more switches. The distribution of index therapies is shown in Table 1. The median time from the first prescription to the first observed switch was 307 days (148–574). First-switch transitions were dominated by moves toward semaglutide and tirzepatide, particularly from liraglutide and dulaglutide (Fig. 1, Supplementary Figure S1). The first-switch matrix provides structural context for the transition-level model by illustrating the dominant therapy-to-therapy pathways within the observed switching network. Table 1Baseline characteristics of patients stratified by GLP-1 receptor agonist switching statusVariableTotal (N = 42,423)No switch (N = 29,803)Switch (N = 12,620)P value Demographics Age, years47.9 (12.9)47.5 (13.1)48.6 (12.2)< 0.001Age ≥ 65 years5266 (12.4)3771 (12.7)1495 (11.8)0.022Female sex29,556 (69.7)20,332 (68.2)9224 (73.1)< 0.001 Anthropometrics BMI, kg/m²33.6 (5.8)33.4 (5.8)33.9 (5.7)< 0.001BMI ≥ 30 kg/m²27,671 (65.2)19,539 (65.6)8,132 (64.4)0.027 Prescription patterns Index year2023 (2022–2024)2024 (2023–2025)2022 (2022–2023)< 0.001Total prescription count6 (3–11)4 (3–8)11 (7–19)< 0.001 Comorbidities Type 1 diabetes mellitus*2230 (5.3)1412 (4.7)818 (6.5)< 0.001Type 2 diabetes mellitus*14,223 (33.5)9442 (31.7)4781 (37.9)< 0.001Dyslipidemia22,656 (53.4)15,267 (51.2)7389 (58.5)< 0.001Obstructive sleep apnea2562 (6.0)1642 (5.5)920 (7.3)< 0.001Optic nerve disorders28 (0.1)14 (0.0)14 (0.1)0.033Arterial hypertension21,971 (51.8)14,832 (49.8)7139 (56.6)< 0.001Coronary artery disease2076 (4.9)1,458 (4.9)618 (4.9)> 0.999Atrial fibrillation1227 (2.9)834 (2.8)393 (3.1)0.081Chronic heart failure1163 (2.7)793 (2.7)370 (2.9)0.126Acute pancreatitis184 (0.4)114 (0.4)70 (0.6)0.017Chronic kidney disease1331 (3.1)877 (2.9)454 (3.6)< 0.001 Index GLP-1 RA class < 0.001Liraglutide or dulaglutide17,556 (41.4)10,490 (35.2)7066 (56.0)Semaglutide21,641 (51.0)16,376 (54.9)5265 (41.7)Tirzepatide3226 (7.6)2937 (9.9)289 (2.3)Data are presented as mean (standard deviation), median (interquartile range), or n (%), as appropriateBMI, body mass index; GLP-1 RA, glucagon-like peptide-1 receptor agonist*ICD-10 diabetes codes were treated as recorded diagnostic indicators and were not mutually exclusive; therefore, some patients carried both E10 and E11 codes Baseline characteristics of patients stratified by GLP-1 receptor agonist switching status Data are presented as mean (standard deviation), median (interquartile range), or n (%), as appropriate BMI, body mass index; GLP-1 RA, glucagon-like peptide-1 receptor agonist *ICD-10 diabetes codes were treated as recorded diagnostic indicators and were not mutually exclusive; therefore, some patients carried both E10 and E11 codes Transition matrix of first observed GLP-1 receptor agonist switch (counts and row-normalized percentages). Heatmap depicting the first observed switch between GLP-1 receptor agonist agents among patients with more than one prescription record. Rows indicate the origin agent, and columns indicate the destination agent at the first observed switch. Each off-diagonal cell displays both the absolute number of patients and the row-normalized percentage for that transition, thereby showing both transition volume and the relative distribution of first-switch destinations within each source therapy. Color intensity reflects the absolute transition count. Percentages sum to 100% within each row (excluding the diagonal no-switch cell). Only the first observed transition per patient is included; subsequent switches are not represented in this matrix Baseline characteristics of the switch-observable cohort are summarized in Table 1. Compared with the no-switch group (n = 29,803), patients who switched (n = 12,620) were slightly older and more frequently female. Cardiometabolic comorbidities—including obesity, type 2 diabetes mellitus, hypertension, dyslipidemia, chronic kidney disease, and obstructive sleep apnea—were more prevalent among patients who switched. BMI at the first prescription was modestly higher in switchers (33.9 [5.7] vs. 33.4 [5.8] kg/m²; P < 0.001). Switchers also had substantially greater treatment exposure, reflected by a higher total prescription count (11 [7–19] vs. 4 [3–8]; P < 0.001) and an earlier index year (2022 [2022–2023] vs. 2024 [2023–2025]; P < 0.001), consistent with longer available observation time among earlier initiators. The distribution of index therapy also differed between groups (P < 0.001): switching was most frequent among initiators of liraglutide or dulaglutide and least frequent among tirzepatide initiators. Multiple switching (≥ 2 switches) was observed in 6074 of 42,423 patients (14.3%), representing 48.1% of those with at least one switch. Patients with multiple switches had substantially higher prescription counts and earlier index years, with modest differences in age and body mass index (Supplementary Table S1). In the transition-level analysis, 12,620 patients contributed 27,095 prescription-to-prescription transitions. After adjustment for switching opportunity (transition number) and calendar time (month-of-year fixed effects and a linear time trend), with patient-clustered robust standard errors and subcutaneous semaglutide as the reference, current therapy remained strongly associated with switching at the subsequent transition (Fig. 2, Supplementary Table S2). Compared with subcutaneous semaglutide, liraglutide was associated with substantially lower odds of switching (OR, 0.02; 95% CI, 0.01–0.03; P < 0.001). Dulaglutide showed a nonsignificant trend toward higher switching odds (OR, 1.70; 95% CI, 0.95–3.04; P = 0.07), whereas oral semaglutide did not differ significantly (OR, 1.30; 95% CI, 0.78–2.17; P = 0.31). Calendar time was independently associated with switching (OR, 1.04 per month; 95% CI, 1.04–1.05; P < 0.001), indicating substantial temporal variation beyond patient-level characteristics. Among baseline covariates, older age was associated with higher switching odds (OR, 1.17 per 10 years; 95% CI, 1.06–1.29; P = 0.002), whereas coronary artery disease was associated with lower odds (OR, 0.50; 95% CI, 0.29–0.87; P = 0.01). Sex (OR, 0.84; 95% CI, 0.67–1.04; P = 0.11), BMI (OR, 1.00 per kg/m²; 95% CI, 0.99–1.02; P = 0.70), type 2 diabetes mellitus (OR, 1.19; 95% CI, 0.96–1.47; P = 0.11), and other comorbidities were not significantly associated with switching. Estimates for tirzepatide as current therapy were not estimable owing to quasi-complete separation (very few or no nonswitching transitions in this subgroup). This non-estimability precluded comparative interpretation of current tirzepatide within the transition-level model; therefore, this exposure was retained for descriptive completeness but was not used to support inferences regarding relative switching stability. Transition-level associations with switching between consecutive GLP-1 RA prescriptions. Forest plot showing odds ratios (ORs) and 95% confidence intervals from a transition-level (discrete-time) logistic regression in which each consecutive prescription interval (Rxt → Rxt + 1) contributes one observation. The outcome is switching (change in agent) at each transition. Subcutaneous semaglutide serves as the reference current therapy. The model adjusts for transition number, calendar time (month-of-year fixed effects and a linear time trend), and comorbidities. Standard errors are cluster-robust by patient to account for within-patient correlation across multiple transitions. ORs greater than 1 indicate higher odds of switching at the next prescription transition. Sensitivity analyses using alternative temporal parameterizations yielded directionally similar therapy contrasts Sensitivity analyses yielded directionally consistent results. In the extended transition-level model, the time since the index prescription was independently associated with switching (OR, 1.04 per 30 days; 95% CI, 1.03–1.05; P < 0.001), whereas the principal therapy associations remained qualitatively unchanged. In a sensitivity analysis in which the linear calendar-time trend was replaced with a restricted cubic spline over calendar month, the principal therapy associations again remained qualitatively unchanged. In an additional sensitivity analysis incorporating both the elapsed time since index therapy and the interval length between consecutive prescriptions, longer prescription gaps were independently associated with switching, whereas the principal therapy associations remained directionally unchanged (Supplementary Table S3). As a complementary survival-analytic sensitivity analysis, we fitted a Cox proportional-hazards model for the time to the first observed switch (censoring at the last recorded GLP-1 RA prescription date; 12,264 events among 12,620 patients). Relative to subcutaneous semaglutide, initiation with liraglutide (HR, 0.71; 95% CI, 0.67–0.74; P < 0.001) and dulaglutide (HR, 0.72; 95% CI, 0.68–0.76; P < 0.001) was associated with a lower hazard of first switch, whereas oral semaglutide was associated with a higher hazard (HR, 1.31; 95% CI, 1.24–1.39; P < 0.001). Initiation with tirzepatide was also associated with a higher hazard of first switch (HR, 2.38; 95% CI, 2.11–2.68; P < 0.001), consistent with strong time- and market-entry effects on observed switching. In this model, older age (HR, 0.94 per 10 years; 95% CI, 0.93–0.96; P < 0.001) and type 2 diabetes mellitus (HR, 0.84; 95% CI, 0.81–0.88; P < 0.001) were associated with a lower hazard of first switch, whereas sex and BMI were not significantly associated (Supplementary Table S4). Directionally similar therapy contrasts were also observed in stratified analyses of patients with and without type 2 diabetes, as well as after exclusion of patients with any E10 code. Switching patterns differed according to type 2 diabetes status and BMI. Among patients without type 2 diabetes, the most frequent first switches were liraglutide to subcutaneous semaglutide (n = 1,549) and liraglutide to tirzepatide (n = 891), followed by subcutaneous semaglutide to oral semaglutide (n = 697). Among patients with type 2 diabetes, the most common transitions were dulaglutide to subcutaneous semaglutide (n = 888) and subcutaneous semaglutide to oral semaglutide (n = 611), followed by subcutaneous semaglutide to dulaglutide (n = 529) and oral semaglutide to subcutaneous semaglutide (n = 514); class-level first-switch distributions differed between type 2 diabetes strata (P < 0.001). When stratified by BMI, liraglutide to subcutaneous semaglutide was the most frequent first switch in both the BMI less than 30 kg/m² and BMI 30 kg/m² or greater groups (n = 436 and n = 1,242, respectively), whereas the BMI 30 kg/m² or greater group showed relatively more liraglutide-to-tirzepatide and dulaglutide-to-subcutaneous-semaglutide transitions; class-level distributions differed modestly between BMI strata (P = 0.014). Among patients with two or more switches, motif analysis demonstrated a predominance of oscillatory, bidirectional trajectories. At the class level (Fig. 3A), the most frequent motif was semaglutide to liraglutide/dulaglutide to semaglutide (16.4%), followed by liraglutide/dulaglutide to semaglutide to liraglutide/dulaglutide (12.9%) and a four-state extension of the same pattern (9.5%), indicating frequent cycling rather than progressive escalation. At the formulation-aware level (Fig. 3B), subcutaneous semaglutide acted as a central hub, with common motifs including subcutaneous semaglutide to dulaglutide to subcutaneous semaglutide (4.5%) and subcutaneous semaglutide to tirzepatide to subcutaneous semaglutide (4.0%), alongside recurrent oral semaglutide intermediates, consistent with periodic trials of alternative formulations between injectable agents. The motif analysis provides structural context for the regression framework by showing that repeated switching was frequently oscillatory rather than unidirectional. Most frequent GLP-1 receptor agonist switching trajectories among patients with multiple switches (≥ 2): class-level and formulation-aware motifs. Horizontal bar charts depicting the 10 most frequent switching trajectories (motifs) among patients with two or more observed switches in the > 1 prescription cohort. Individual treatment histories were represented as ordered sequences of agents and collapsed by removing consecutive duplicates (e.g., A–A–B simplified to A–B); trajectories were truncated to a maximum of five states, with “…” denoting sequences extending beyond this limit. Bar length represents the absolute number of patients with a given motif; percentages indicate the proportion of all multi-switchers accounted for by each motif. Panel A presents class-level trajectories using the following abbreviations: S semaglutide, L/D liraglutide or dulaglutide, T tirzepatide. Panel B presents formulation-aware trajectories using: S(sc) semaglutide subcutaneous, S(po) semaglutide oral, L liraglutide, D dulaglutide, T tirzepatide Calendar-time analyses of transition rates revealed clear temporal variability (Fig. 4A). Overall switching intensity displayed distinct peaks, and subcutaneous-to-oral semaglutide transitions showed intermittent increases (Fig. 4B), a pattern compatible with episodic substitution away from injectable semaglutide during discrete time windows. In parallel, transitions terminating in tirzepatide increased during specific periods after market introduction (Fig. 4C), suggesting accelerated uptake during time-limited intervals. Denominators were displayed alongside rates, and a 3-month rolling mean was used to distinguish sustained shifts from isolated monthly spikes. Temporal signals of GLP-1 receptor agonist switching and availability: monthly transition-based rates with three-month rolling mean. Panels A–C display monthly switching signals calculated at the prescription-to-prescription transition level among patients with more than one prescription record. Solid lines represent the monthly rate per 100 observed transitions; dotted lines represent the three-month rolling mean, included to highlight sustained temporal trends and differentiate them from isolated monthly variation. Grey bars (right y-axis) denote the monthly number of transitions (denominator), providing context for the stability of individual rate estimates. Panel A shows overall switching intensity - defined as any class or formulation change relative to no change - per 100 transitions. Panel B shows the rate of semaglutide s.c. → semaglutide oral transitions per 100 transitions originating from semaglutide s.c., used as a proxy for episodic substitution away from the injectable formulation compatible with periods of reduced subcutaneous availability. Panel C shows the rate of transitions terminating in tirzepatide per 100 transitions, used as a proxy for adoption dynamics following market introduction and/or improved access Discussion In this large real-world analysis from a private healthcare network in Poland, switching among GLP-1 receptor agonists was frequent and temporally dynamic. Because only approximately one-third of patients had type 2 diabetes mellitus, the observed treatment trajectories likely reflect obesity management and weight-focused prescribing in addition to glycemic intensification. Consistent with prior real-world reports [12, 15, 16], we observed substantial heterogeneity in switching patterns across agents, including prominent transitions toward semaglutide and tirzepatide and a sizeable subgroup of patients with repeated switching. A contemporaneous US claims-based analysis by Xie et al. of 126,984 adults with overweight or obesity without diabetes reported a 12-month switching frequency of 20.6%, with liraglutide-to-semaglutide as the most common first transition and injectable semaglutide serving as a central switching hub [17]. Our findings extend this picture in several directions: the longer observation window captures the tirzepatide market-entry period more completely, the mixed-indication cohort reflects a broader prescribing population, and the transition-level framework characterizes switching propensity conditional on observed opportunities rather than treating switching as a single binary outcome within a fixed follow-up window. A central methodological challenge in switching research is that switching is inherently time-dependent and opportunity-driven. Patient-level “ever-switched” outcomes are strongly influenced by observation intensity: patients with more recorded prescriptions and earlier index years have greater opportunity to switch. In the patient-level models, cumulative prescription count dominated the association with switching, supporting the concern that such models largely capture opportunity rather than clinical determinants. For this reason, we prioritized a transition-level (discrete-time hazard) framework that models the probability of switching at each prescription interval, conditions on observed opportunities, and adjusts for calendar time. In this time-aware approach, current therapy remained strongly associated with switching at the next interval compared with subcutaneous semaglutide, whereas most baseline comorbidities showed smaller or nonsignificant associations. These findings should be interpreted as associations with switching propensity at a given prescription interval, not as causal effects of therapies or patient characteristics. The particularly large inverse contrast for liraglutide should not be interpreted as evidence of superior treatment stability or comparative effectiveness; rather, it likely reflects a combination of residual confounding, structural imbalance in observed prescription sequences, and the limitations of routine EHR-based transition data. The transition-level model and the Cox model capture complementary aspects of switching and therefore answer different questions. The transition-level model estimates switching propensity at the next observed prescription interval, conditional on current therapy and continued follow-up, whereas the Cox model estimates the time from index therapy to the first observed switch. For late-entry agents - particularly tirzepatide - these estimands are differentially affected by market-entry timing, censoring, and incomplete capture of prescriptions outside the network. We therefore interpret the Cox analysis as a robustness check for time structure rather than as an alternative primary comparison of therapy stability. The descriptive structure of switching provides additional context. First-switch matrices and Sankey diagrams showed that transitions frequently moved from older incretin therapies (liraglutide and dulaglutide) toward semaglutide and, subsequently, toward tirzepatide. Among patients with multiple switches, we observed a predominance of bidirectional, “return-to-origin” motifs, with cycling between semaglutide and liraglutide/dulaglutide classes accounting for a large share of trajectories [18–20]. Notably, Xie et al. reported that 92.7% of switchers in their 12-month window switched only once [17], whereas in our longer observation period nearly half of patients with any switch experienced two or more switches, and oscillatory “return-to-origin” motifs were prominent. This contrast likely reflects both the extended follow-up in the present analysis and the structural features of prescription-sequence data captured across multiple calendar years, and it suggests that the cyclic nature of GLP-1 RA use becomes apparent only with sufficiently long observation. Such oscillatory patterns are consistent with nonlinear decision pathways in routine care, in which therapy changes may reflect periodic reassessment, tolerability trade-offs, formulation preferences, or intermittent external constraints rather than a unidirectional escalation strategy [1, 18–20]. These oscillatory trajectories may also reflect broader clinical uncertainty about whether incretin-based therapy should be maintained continuously or used episodically, a question that remains unresolved even as the therapeutic armamentarium expands [1]. Figures 1 and 3 are intended as structural complements to the regression framework, illustrating the dominant first-switch pathways and recurrent oscillatory motifs within which the transition-level associations arise, rather than as independent inferential analyses. Importantly, these trajectory analyses are descriptive and do not establish mechanisms. Temporal analyses further indicated that switching propensity varied over calendar time, with distinct peaks in overall switching intensity and in specific transitions (e.g., subcutaneous-to-oral semaglutide and transitions terminating in tirzepatide) [15, 20–22]. These observations are consistent with emerging evidence that rapid therapeutic innovation is associated with dynamic real-world treatment patterns in obesity care [17], with incumbent agents frequently displaced as newer, more efficacious formulations become available. Importantly, these findings were robust to a more flexible nonlinear specification of calendar time, modeled as a restricted cubic spline, which suggests that the principal therapy associations were not primarily driven by misspecification of a linear temporal trend. These time-limited patterns are compatible with episodes of substitution and adoption following market entry. However, because the dataset does not capture dispensing, individual-level reimbursement or cost exposure, supply disruptions, or patient and clinician preferences, we cannot attribute these peaks to specific system-level drivers [23]. We therefore frame cost, availability, and tolerability as plausible hypotheses consistent with the observed time structure rather than as empirical conclusions [24–26]. Notably, supply shortages driven by rapidly increasing demand for GLP-1 RAs have been recognized as a growing concern at the class level [1, 27–29], lending plausibility to the hypothesis that at least some of the observed temporal clustering reflects access-related rather than clinically motivated switching. Several findings also warrant cautious interpretation in light of calendar-time and market-entry effects. Tirzepatide entered the market later and had shorter observable exposure windows within the study period. In the transition-level model, estimates for tirzepatide as current therapy were not estimable owing to quasi-complete separation (very few or no nonswitching transitions in that subgroup), which likely reflects a combination of short observation windows and data-capture characteristics rather than intrinsic treatment stability or instability. Cox models of time to first switch provide a familiar survival framework but are similarly sensitive to differential market exposure and incomplete capture of medication use outside the private network; accordingly, these were treated as sensitivity analyses rather than the primary inferential approach. Limitations and clinical implications This study has several limitations inherent to retrospective analyses of outpatient EHR prescription sequences. First, we did not capture reasons for switching (e.g., tolerability, perceived effectiveness, patient preference, cost sharing, reimbursement changes, or supply constraints). Accordingly, observed temporal clustering and therapy-specific switching patterns should be interpreted as associations compatible with changes in access or preferences rather than as evidence of causal mechanisms. Second, switching is intrinsically time-dependent and opportunity-driven. Although patient-level “ever switched” models are highly sensitive to differential observation intensity, we mitigated this concern by prioritizing transition-level discrete-time hazard analyses that condition on each observed switching opportunity, adjust for calendar time, and use patient-clustered robust inference. Nevertheless, residual time-related bias cannot be fully excluded, particularly because prescriptions recorded in a private network may not capture all medication use if subsequent care occurs outside the system. Third, prescription entries reflect clinician orders rather than confirmed dispensing; therefore, we cannot confirm medication acquisition, adherence, or continuous exposure, and apparent switches may partly reflect treatment gaps, intermittent use, or reinitiation rather than deliberate within-class substitution. Relatedly, censoring at the last recorded prescription may misclassify follow-up if patients continued therapy outside the network. In addition, the recorded prescription-sequence structure may incompletely capture silent continuation of unchanged therapy, which may amplify apparent imbalances between switching and non-switching transitions for some agents. The observation process itself may also be informative, as patients with more frequent visits or prescriptions may differ systematically with respect to disease complexity, access, or treatment behavior. Fourth, comorbidities were derived from ICD-10 codes and treated as binary proxies that do not capture disease severity, duration, or treatment indication; thus, residual clinical confounding is likely. A small proportion of patients (5.3%) carried a type 1 diabetes mellitus diagnosis, likely reflecting off-label prescribing and/or coding imprecision; these patients were retained to reflect real-world prescribing patterns, but their inclusion should be considered when interpreting comorbidity-related findings. Finally, later-market therapies (particularly tirzepatide) have shorter observable exposure windows, which can influence both descriptive switching rates and time-to-event analyses despite calendar-time adjustment; for this reason, we avoided efficacy-based interpretation and reported instances of nonestimability transparently when data structure limited inference. Likewise, the large inverse contrast observed for liraglutide should be understood as a non-causal association within the recorded transition structure, not as evidence of superior treatment stability or comparative effectiveness. From a clinical perspective, these findings underscore that switching among GLP-1 RAs in routine care is frequent and temporally dynamic, and that treatment stability may be influenced by a combination of patient factors and the broader prescribing environment. Consistent with recent US evidence [17], the frequent and often bidirectional switching patterns observed in our cohort are compatible with switching as a form of active therapy management rather than treatment disengagement. Reframing within-class switching as part of long-term therapeutic optimization - rather than as an early signal of treatment failure or abandonment - may better align clinical expectations with how GLP-1 RAs are used in routine practice, particularly in settings of rapid market evolution and formulation diversification. From a practical standpoint, clinicians should anticipate potential treatment interruptions or substitutions during periods of rapid market change or fluctuating availability and should consider proactive counseling, monitoring, and follow-up planning when initiating or modifying GLP-1 RA therapy. At the same time, the present data do not allow determination of whether individual switches were clinically appropriate, access-driven, or preference-driven. Future studies linking electronic health record prescriptions with dispensing data, reimbursement information, and patient-reported reasons for switching would be needed to distinguish clinical drivers from access- and system-related factors and to quantify their relative contributions. Supplementary Information Below is the link to the electronic supplementary material. Supplementary file1 (DOCX 413 KB) Declarations Conflict of interest The authors declare that they have no conflict of interest. Ethical approval Study has been approved by the appropriate ethics committee and has therefore been performed in accordance with the ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments. Declaration of generative AI use The authors declare that generative artificial intelligence (Claude Opus 4.6, Anthropic) was used during the preparation of this manuscript for language editing and stylistic improvement. Following its use, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Change history 7/28/2026 The original online version of this article was revised: due to affiliation 4 was incorrectly published in the article. References Associated Data Supplementary Materials Supplementary file1 (DOCX 413 KB)