We conducted a retrospective, propensity score–matched cohort study using de-identified electronic health record (EHR) data from the TriNetX US Collaborative Network. The study period spanned January 1, 2000 through December 31, 2022, with follow-up accrued to the longest available time horizon within the network for each patient. Summary provided as Figure 1.

TriNetX is a federated network of participating health systems that harmonizes encounter-level EHR data (demographics, diagnoses, procedures, medications, vitals, and selected laboratory results) using standard terminologies (e.g., ICD-9/10, CPT, LOINC, RxNorm). Only de-identified, aggregate counts and summary statistics are returned to end users; site-level data quality checks and centralized mapping ensure semantic alignment. Because TriNetX reflects routine clinical documentation, missingness is not imputed and represents real-world data capture. Platform analytics support cohort building, outcome definitions, and propensity score matching with standardized diagnostics (e.g., absolute standardized mean differences [SMDs]) (11).

We identified adults (≥18 y) with a diagnosis of chronic myeloid leukemia (CML) who initiated first-line tyrosine kinase inhibitor (TKI) therapy with imatinib, dasatinib, or nilotinib during the study window.

To reduce immortal-time and reverse-causality biases, we excluded patients who had any study outcome recorded before the exposure anchor (for switchers) or its aligned reference time (for non-switchers): all-cause mortality, heart failure (HF), thromboembolic events, diabetes, cardiovascular complications, all-cause hospitalization, or ICU admission. Patients with missing age or sex were excluded. Where exact dates were coarsened, standard TriNetX anchoring rules were applied.

For switchers, follow-up began immediately after the switch date. For non-switchers, we assigned a pseudo-switch reference time equal to the median switch time observed among switchers within the same TKI stratum (or the individual-level matched switch time in the post-match analysis), and follow-up began immediately after that reference time. Patients were followed to the earliest of outcome, death, disenrollment/no further encounters, or the analytic horizon (3 or 5 years).

Outcomes were ascertained from diagnosis/procedure codes and vital status where available, and evaluated at 3-year and 5-year horizons:

Baseline (pre-anchor) covariates included demographics (age, sex, race), disease duration proxies (time from CML diagnosis to index), comorbidities (e.g., cardiovascular disease, diabetes, hypertension, obesity, chronic kidney disease, chronic lung disease), prior hospital/ICU utilization, and (where available) selected labs/vitals. Covariates were captured during a look-back window up to one day before the exposure anchor (switch or pseudo-switch).

Within each TKI stratum (imatinib, dasatinib, nilotinib), we estimated the propensity to early-switch using logistic regression including all covariates above (main effects). We performed 1:1 greedy nearest-neighbor matching without replacement using a caliper of 0 (12). standard deviations of the logit of the propensity score. Covariate balance was assessed by absolute SMD, with SMD <0 (13). prespecified as acceptable residual imbalance. We report pre- and post-match SMDs and p-values for transparency, while interpreting treatment effect estimates primarily under SMD thresholds.

For matched cohorts, we estimated risk ratios (RRs) with 95% confidence intervals (CIs) for each outcome at 3 and 5 years using log-binomial or Poisson models with robust variance where appropriate; two-sided p < 0 (8). denoted statistical significance. In matched analyses, standard errors were clustered on matched pairs. Descriptive statistics are presented as mean (SD) or % as appropriate. Missing data were analyzed as observed (no imputation), consistent with TriNetX conventions and real-world data guidance. Analyses followed commonly used TriNetX pipelines for outcome analytics and matched cohort estimation.

Cohort counts (pre- and post-match):

After propensity score matching, the analytic cohorts comprised imatinib (early switch ≤3 months, n=349; non-switchers, n=4,579), dasatinib (n=377 vs n=2,849), and nilotinib (n=114 vs n=1,308). Post-match distributions of age, sex, and race were well balanced within each TKI stratum. Baseline characteristics after matching are summarized in Table 1, and full before/after balance panels are presented in Supplementary Tables S1–S3.

Imatinib—At 5 years, all-cause mortality was higher among early switchers compared with non-switchers (24.0% vs 11.9%; RR 2.03, 95% CI 1.43–2.87; p<0.001). Heart failure was higher at 3 years (10.8% vs 5.4%; RR 1.99, 95% CI 1.14–3.48; p=0.015) and 5 years (12.1% vs 7.3%; RR 1.65, 95% CI 1.01–2.71; p=0.046). Hospitalizations (27.9% vs 16.5%; RR 1.69, 95% CI 1.09–2.62; p=0.019) and ICU admissions (11.9% vs 7.0%; RR 1.70, 95% CI 1.04–2.78; p=0.033) were also higher.

Dasatinib—Early switching was associated with higher 5-year mortality (21.1% vs 14.0%; RR 1.51, 95% CI 1.10–2.08; p=0.011). Heart failure at 3 years was higher (12.1% vs 6.7%; RR 1.80, 95% CI 1.09–2.97; p=0.020). ICU utilization was greater at 3 years (17.3% vs 9.2%; RR 1.88, 95% CI 1.28–2.75) and 5 years (19.9% vs 10.3%; RR 1.88, 95% CI 1.27–2.79; p=0.002).

Nilotinib—A numerical increase in 5-year mortality did not reach significance (20.2% vs 11.5%; RR 1.75, 95% CI 0.95–3.23; p=0.074), while ICU admissions were significantly higher at 5 years (19.4% vs 9.3%; RR 2.10, 95% CI 1.07–4.11; p=0.033).

A consolidated summary of absolute risks and relative risks is provided in Table 2. Figure 2 displays a forest plot of risk ratios across outcomes and TKIs. Figure 3 depicts Kaplan–Meier survival curves for the imatinib stratum using survival probabilities that reproduce the 5-year estimates. Figure 4 illustrates cohort assembly.

To improve interpretability of the outcome data, we also summarize the 5-year absolute event rates as grouped bar charts by TKI and early-switch status (Figure 5), while retaining Table 2 to provide complete numerical detail.

In this large, propensity-matched, real-world cohort drawn from the TriNetX US Collaborative Network, early switching of frontline TKI (≤3 months) was associated with higher long-term risks (including mortality and major complications) compared with continued therapy on the initial agent, despite balance on measured covariates after matching (13, 15–18).

These results align with the broader real-world observation that the need to switch early often reflects intolerance or emerging resistance—clinical states that portend inferior molecular trajectories and more complicated courses in practice (19–24).

Together, these data reinforce that “routine” early switching should not be reflexive in the absence of guideline-defined failure or unmanageable toxicity (3, 7, 8, 25).

Our observations are consistent with the SIMPLICITY program and other multicenter registries showing that intolerance is the predominant driver of early TKI change and that patients who remain on their initial TKI are more likely to achieve key cytogenetic/molecular milestones, with similar short-term survival but more favorable complication profiles over time (19, 20, 23, 24).

In SIMPLICITY’s European cohort, non-switchers more frequently attained CCyR and MMR by year 3, supporting the concept that treatment continuity facilitates timely depth of response when clinically feasible (20).

Additional real-world datasets (Italian Network/CML Campus, UK TARGET, Swiss series) further suggest that durable outcomes are achievable across TKI choices provided patients can persist on therapy with adequate monitoring and supportive care (22–24).

Randomized and pooled analyses show that second-/third-generation TKIs induce faster and deeper molecular responses vs. imatinib but have not translated into overall survival advantages in the frontline setting, while increasing specific toxicities (e.g., cardiovascular, pleural, hepatic) (26–28).

These trade-offs imply that an early switch intended to “accelerate” molecular kinetics may expose certain patients to higher adverse-event burdens without proven survival gain, particularly when the trigger is intolerance rather than bona fide resistance; our findings are consistent with this risk–benefit tension in routine care (26–28).

From a kinetic standpoint, second-generation TKIs achieve early and deep molecular responses more quickly than imatinib, but long-term randomized and pooled data show little or no overall survival advantage when milestone-driven switching is applied (23, 26). At the same time, generic imatinib is substantially less expensive than patented 2GTKIs, and health-economic analyses consistently find that “imatinib-first” strategies remain the most cost-effective frontline option in many health systems, particularly for older or comorbid patients (1, 2, 10, 26, 27). Nilotinib and other 2GTKIs also carry a time-dependent risk of arterial occlusive and other vascular events that continues to accrue with prolonged exposure, often beyond 5–8 years of therapy (26). Because our median follow-up was approximately 5 years, our study likely underestimates the full lifetime burden of nilotinib-associated vascular toxicity and the extent to which this toxicity might offset any earlier molecular advantages of switching.

Patients who switch early likely represent a biologically and clinically enriched risk subset—those with higher comorbidity burden, frailty, pharmacogenomic variability, or early suboptimal response (including occult kinase domain mutations or adverse ELTS/Sokal risk)—all factors associated with inferior outcomes independent of the specific switch decision (3, 7, 29–33).

In routine practice, baseline TKI selection itself encodes aspects of clinical risk: older patients, those with substantial cardiovascular comorbidity, and those with presumed low- or intermediate-risk CML are more often started on imatinib, whereas younger patients and those perceived to have higher-risk disease or a greater need for rapid cytoreduction are preferentially offered 2GTKIs (1, 2, 5, 8). In our TriNetX cohorts, this pattern was reflected in the unmatched characteristics, with imatinib recipients being older and more comorbid, and early switchers tending to have a heavier comorbidity burden than non-switchers within each TKI stratum. Our propensity-score models incorporated age, sex, comorbid conditions (including cardiovascular disease), and available baseline laboratory parameters to minimize these imbalances, but residual confounding by unmeasured disease-specific risk (e.g., Sokal or ELTS scores) cannot be excluded.

Moreover, rapid exposure to multiple TKIs in succession can compound organ-specific toxicities, drug–drug interactions, and monitoring complexity, increasing the probability of treatment interruptions and cumulative harm that may offset any theoretical efficacy gains from a more potent agent (26, 27, 31, 32).

NCCN and ELN endorse molecular response milestones at 3 and 6 months, with action thresholds that distinguish “warning” from “failure,” emphasizing that management of mild–moderate toxicities should prioritize dose adjustment or brief interruption before switching whenever feasible (3, 8, 25, 34–36).

Our real-world signal—worse long-term outcomes among early switchers—supports these frameworks by cautioning against pre-emptive switching before the 3-month assessment unless toxicity is severe and non-reversible, or clear resistance biology is present (3, 7, 8, 25, 29, 30, 36, 37).

We defined early switching as any change within ≤3 months of initiation, a pragmatic window broader than guideline triggers that anchor decision-making to the 3-month molecular check (with earlier changes generally reserved for severe intolerance) (3, 25, 29).

Consequently, our early-switch cohort is inherently heterogeneous (intolerance, adherence issues, pharmacokinetics, early suboptimal response), which likely contributes to the adverse outcomes observed and should be considered when extrapolating these findings to guideline-directed switching after formal milestone assessment (3, 25, 38).

Asciminib has demonstrated higher 48-week MMR rates with favorable tolerability vs. investigator-selected TKIs in frontline CML, suggesting that safer first-line options may reduce intolerance-driven switching and its downstream risks observed in legacy practice patterns (39, 40).

As the therapeutic portfolio evolves, the calculus around early modification may shift—yet the core principles of milestone-guided decisions, toxicity mitigation before switching, and individualized agent selection by comorbidity/mutation profile remain central (3, 7, 8, 29, 30, 37, 39).

Strengths include large scale, multi-health-system capture, prespecified outcomes, exclusion of patients with outcome events prior to the switch index, and propensity matching to reduce measured confounding (13, 15–18).

Nonetheless, residual confounding by indication is unavoidable (e.g., undocumented reasons for switching, adherence, dose intensity), molecular landmarks are incompletely captured, misclassification is possible in EHR-derived data, and informative censoring/competing risks may bias effect estimates even after matching (13, 15–18, 38).

Time-varying confounding (e.g., evolving toxicity, disease kinetics) complicates causal interpretation; future target-trial emulations with marginal structural models and richer molecular/PRO inputs could further strengthen inferences (13, 15–18, 38).

A further limitation of this study is that established prognostic scores such as Sokal and ELTS could not be calculated within the TriNetX US Collaborative Network, as several key variables required for their computation are unavailable or incompletely captured. As a result, we were unable to stratify “early switchers” and “non–early switchers” by these risk categories, which may have limited our ability to fully account for baseline disease risk and to explain some of the observed differences in outcomes between groups (4, 33). Although individual-level data needed to compute Sokal, EUTOS, or ELTS scores (e.g., spleen size, peripheral blast percentage, detailed cytogenetics) were not consistently available in TriNetX, several components of these scores—such as age and platelet counts—were indirectly accounted for through our matching strategy (4, 33). Nonetheless, systematic differences in baseline CML risk between frontline imatinib and 2GTKIs, or between early-switchers and non-switchers within a given TKI, may partially contribute to the outcome differences observed.

Finally, mortality in this analysis reflects all-cause death as captured in the aggregated electronic health record rather than adjudicated CML-related mortality. For many patients with chronic-phase CML on long-term TKI therapy, deaths are now more frequently due to competing comorbidities—particularly cardiovascular disease, secondary malignancies, and serious infections—than to leukemic progression itself (1, 2, 5, 31). We could not reliably assign a cause of death or directly link individual deaths to specific toxicities such as arterial occlusive events or pleural effusions, which limits mechanistic interpretation of the modest mortality differences between early-switch and non-switch cohorts.

For most patients, an initial strategy of optimizing the current TKI—addressing adherence, drug interactions, supportive care, and dose modification—before switching aligns with both our real-world signal and contemporary guidance (3, 7, 8, 25, 36, 37). The 2025 European LeukemiaNet recommendations similarly move away from a purely milestone-driven mandate to switch TKIs at the first sign of suboptimal response and instead emphasize individualized decisions that integrate disease risk, comorbidities, toxicity, cost, and patient preferences (34). Our observation that patients who remain on their initial TKI when early response goals are met have outcomes at least as good as, and often better than, early switchers provides real-world support for this more nuanced, patient-centered approach.

Switching should be prioritized for guideline-defined failure, persistent and severe intolerance despite optimization, or resistance with actionable mutations, with agent choice tailored to comorbid risks (e.g., vascular risk with nilotinib/ponatinib, pleural-pulmonary risk with dasatinib) and patient goals (e.g., TFR vs. durable disease control) (3, 7, 26, 29, 30, 37). Consistent with this guidance, judicious dose reduction or supportive management of chronic low-grade adverse events may often be preferable to reflexive TKI substitution, particularly in older or comorbid patients in whom imatinib remains highly effective and cost-effective (1, 2, 10, 26, 27, 34).

Ultimately, disciplined milestone-based management and judicious switching can minimize avoidable harm while preserving long-term survival gains achieved in the TKI era (3, 7, 8, 26).