Real‐World Luspatercept Evidence in Myelodysplastic Neoplasms: A Systematic Review and Bayesian Meta‐Analysis

ABSTRACT

Luspatercept has emerged as a therapeutic option for patients with lower‐risk myelodysplastic neoplasms (MDS). Nonetheless, the performance of luspatercept outside of controlled clinical trials remains unclear. We aimed to synthesize real‐world evidence (RWE) on the effectiveness and safety of luspatercept. This study was conducted following PRISMA guidelines. We searched for studies up to June 2025 evaluating luspatercept in adult MDS patients. Data on hematologic improvement‐erythroid (HI‐E), transfusion independence (TI at 8, 12, and 16 weeks), adverse events, and overall survival (OS) were extracted. Bayesian random‐effects meta‐analyses were performed using priors derived from pooled clinical trials. Seventeen studies were included: five clinical trials (440 patients) and twelve real‐world cohorts (1821 patients). The pooled estimate for HI‐E was 46.6% (95% CrI: 32.5%–63.9%). For TI, pooled rates at 8, 12, and 16 weeks were 44.7% (95% CrI: 28.6%–61.5%), 38.6% (95% CrI: 21.1%–60.0%), and 30.9% (95% CrI: 10.7%–53.9%), respectively. The subgroups with highest TI and HI‐E were patients with positive SF3B1 status (58.5%, at 8 weeks) and Asian patients (54.8%), respectively. Male sex was associated with lower HI‐E, 8, and 12 weeks TI rates. Hypertension and falls were more frequently reported in RWE. We estimated OS rates of 88.9% at 1 year and 74.4% at 2 years following treatment initiation. Real‐world HI‐E were modestly attenuated, likely reflecting selection, adherence, and monitoring differences. The response seems dependent on disease, geographical, and demographical moderators. Such aspects should be taken into account in the design of future studies and in clinical decisions.

Abbreviations

CrI

Credible interval

ESA

Erythropoietin‐stimulating agents

FDA

Food and Drug Administration

HI‐E

Hematological improvement‐erythroid

MDS

Myelodysplastic neoplasms

RWE

Real‐world evidence

SD

Standard Error

TI

Transfusion independence

1. Introduction

Myelodysplastic neoplasms (MDS) are a heterogenous group of clonal hematopoietic malignancies characterized by ineffective hematopoiesis, peripheral blood cytopenias, and a risk of progression to acute myeloid leukemia (AML) [1]. In low‐risk MDS, the inflammatory bone marrow niche combines with intrinsic defects in epigenetics and splicing to promote enhanced innate immune signaling [2, 3, 4]. A proapoptotic phenotype then emerges, disrupting both early and late hematopoiesis [5, 6]. The main manifestation of this phenomenon is impaired erythropoiesis, with approximately 73%–87% of low‐risk MDS patients eventually becoming transfusion‐dependent [7, 8]. This becomes particularly challenging after the failure of erythropoiesis‐stimulating agents (ESAs). Treatment failure of ESAs is common, particularly after a median of 2 years of treatment, which makes other options crucial to avoid transfusion dependency [9, 10, 11].

Luspatercept (Reblozyl) is a recombinant fusion protein composed of modified activin receptor type IIb that binds select ligands of the transforming growth factor‐beta superfamily [12, 13]. It works mainly by counteracting overactive Smad2/3 signaling, which is involved in impaired erythroid differentiation and prevalent in MDS CD34+ cells [12, 13, 14]. This blockade enhances hematopoietic differentiation and increases late‐stage erythroid precursors in the bone marrow [14, 15]. Initially approved by the Food and Drug Administration (FDA) in 2019 and the European Medicines Agency in 2020 for transfusion dependent MDS patients with ring sideroblasts unresponsive to or ineligible for ESAs, luspatercept received expanded FDA approval in 2023 for first‐line therapy in very low‐to intermediate‐risk MDS patients requiring red blood cell transfusions and without prior ESA treatment [5, 16].

Given its recent development and even more recent regulatory approval, evaluations of the effectiveness and safety profiles of luspatercept in real‐world scenarios are urgently needed. Knowledge gaps persist regarding its long‐term effectiveness, its utility across specific patient subgroups, and its potential role in combination with other ESAs [5, 17, 18]. Furthermore, as most evidence on luspatercept is derived from controlled clinical trials, real‐world evidence (RWE) remains underexplored, with variable response rates [19, 20]. To address these gaps, we synthesized all available published data on luspatercept in the treatment of MDS, leveraging Bayesian meta‐analytical methods to update our prior knowledge (clinical trials) with emerging RWE about luspatercept's effectiveness and safety.

2. Methods

2.1. Protocol Registration

This review was conducted and reported according to the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) guidelines (Supporting Information S1: Table S1). The protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO) database under protocol CRD420251030529.

2.2. Search Strategy and Study Selection

We searched the PubMed, Cochrane Central Register of Controlled Trials, and EMBASE databases for studies published up to June 7, 2025. The search strategy is shown in Supporting Information S1: Table S2. Studies that met all the following criteria were considered for inclusion: (i) evaluated adult (> 18 years) patients diagnosed with MDS; (ii) had at least one group treated with luspatercept; (iii) assessed effectiveness or safety; and (iv) included data from either clinical trials or cohort studies. The exclusion criteria were as follows: (i) including patients diagnosed with other hematological malignancies; (ii) studies not clearly fitting in the RWE or clinical trial categories; (iii) studies outside the established timeframe; and (iv) studies with lower evidentiary weights, such as case reports, conference abstracts, and protocols.

2.3. Data Extraction

Two investigators (P.P. and R.V.) screened titles and abstracts, followed by an analysis of full‐text articles selected in the previous step. Discrepancies in inclusion criteria were resolved by subsequent discussion and consensus. Data extracted from eligible studies included: (i) general characteristics; (ii) effectiveness ‐related measures; (iii) safety‐related measures (e.g., proportion of adverse effects). For studies reporting only medians, means were estimated using the Box‐Cox approach via the 'estmeansd' package in R [21].

2.4. Assessment of Quality

The assessment of study quality was conducted independently by two authors (V.F. and R.V.) using the Newcastle–Ottawa Scale (NOS) for nonrandomized studies [22], the ROBINS‐I tool for nonrandomized studies [23], and the ROBINS‐II tool for randomized studies [24].

2.5. Outcomes

The primary outcomes of this meta‐analysis were: (i) red blood cell transfusion independence (TI, ≥ 8, ≥ 12, or ≥ 16 weeks); and (ii) hematologic improvement‐erythroid (HI‐E), based on the International Working Group (IWG) 2000, 2006, or 2018 criteria, as specified per study. All effects were synthesized through proportions and their respective 95% credible interval (CrI). As secondary outcomes, we analyzed safety‐related events, including the frequency of specific adverse effects, the number of grade ≥ 3 adverse events and the number of treatment discontinuations due to complications. Furthermore, we evaluated the overall survival (OS) of patients submitted to luspatercept treatment through Kaplan‐Meier (KM) curves.

2.6. Bayesian Prior Construction

We employed a semi‐informative prior for the overall effect defined as a normal distribution, which is generally ideal [25], on the logit scale with a mean equal to the logit‐transformed pooled success rate from the included clinical trials. These data were derived from pooling in a frequentist random‐effects fashion with generalized linear mixed models through the “metafor” package [26]. Standard deviations reflected our level of confidence and were adjusted so that the 95% prior interval would be approximately 20% beyond the upper and lower bounds of the 95% confidence interval (CI) obtained from the trial estimates. It is necessary to maintain sufficient variance to reflect realistic uncertainty when updating with new evidence, explicitly allowing real‐world data to meaningfully update expectations, given known differences in patient selection, adherence, and monitoring between trials and routine care [27]. The use of trial‐based meta‐analyses to inform priors has precedent in Bayesian evidence synthesis [25, 28]. Between‐study heterogeneity was modeled using a Half‐Cauchy(0, 2.5) prior, an accepted choice for hierarchical models [29]. We also conducted subanalysis based on alternative priors on the main outcomes.

2.7. Data Synthesis

Proportions were extracted or calculated via established methods, each accompanied by 95% CIs. Real‐world studies were then pooled based on Bayesian random‐effects models with logit transformation, using the priors derived from clinical trials. All Bayesian models were fitted using the “brms” package, with 4 chains of 4000 iterations each (1000 warm‐ups) [30]. Results are reported as pooled proportions with corresponding 95% credible intervals (CrIs). For the assessment of heterogeneity, we followed the guidelines by Spiegelhalter (2003) [31], who proposed that values of the between‐study standard deviation (τ) between 0.1 and 0.5 suggest reasonable heterogeneity, 0.5 to 1.0 indicate fairly high heterogeneity, and values above 1.0 imply fairly extreme heterogeneity. Although this guidance was originally proposed for meta‐analyses of odds ratios [31], modeling of proportions on the log‐odds scale implies a similar interpretation of τ, since both are based on the same underlying transformation. All analyses were performed using R version 4.3.2 (R Foundation for Statistical Computing).

2.8. Reconstruction of Time‐to‐Event Data

We reconstructed time‐to‐event data from KM curves reported in the included studies. Coordinates were extracted from each curve using validated digital methods [32], and the corresponding numbers at risk were retrieved from the published risk tables. We then applied the IPDfromKM R package [33] to simulate individual patient‐level data (IPD). The reconstructed datasets were aggregated to generate a unified KM curve. For group comparisons, we used a univariate Cox proportional hazards model, and assessed the proportional hazards assumption through Schoenfeld residuals [34].

2.9. Additional Analyses

Subgroup analyses were conducted whenever appropriate, using comparable data aggregates informed by clinical rationale. The same Bayesian methods and priors described previously were applied. In parallel, Bayesian meta‐regression analyses were conducted by introducing continuous covariates as linear terms individually through the “brms” package [30]. Variables were considered influential if their associated regression coefficients demonstrated a posterior probability of direction (P[D]) of at least 95%, indicating a high probability that the effect was nonzero and directionally consistent [35, 36]. Given the ecological nature of meta‐regression, all results should be viewed as hypothesis‐generating [37]. For publication bias assessment, we employed Peter's test, which has been shown to outperform other funnel plot asymmetry tests in meta‐analyses using the natural logarithm of effect sizes [38].

3. Results

3.1. Search Results

Among the 880 studies identified through the literature search, 535 remained after duplicate removal (Supporting Information S1: Figure S1). From these, 25 were selected for full‐text screening. Eight studies were excluded: three due to overlapping populations (with preference given to those with larger sample sizes), one for not specifically focusing on patients with MDS, and four for presenting long‐term follow‐up data from clinical trials, which were excluded as they do not fall under either of the two categories of interest (clinical trials or RWE, see Supporting Information S1: Table S3). Ultimately, 17 full‐text studies were included in the analysis [9, 18, 19, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52].

3.2. Main Characteristics of the Included Studies

Of the 17 studies included in this meta‐analysis, five were clinical trials (used for prior construction), of which two were Phase 3 randomized clinical trials [9, 50] and three were Phase 2 studies [18, 40, 51], and 12 were real‐world retrospective cohorts [19, 39, 41, 42, 43, 44, 45, 46, 47, 48, 49, 52]. Overall, 2261 MDS patients on luspatercept were included, of which 59.4% were men and 84.5% had prior ESA usage. The mean age was 73.93 ± 8.14 years. No real‐world studies were conducted in the ESA‐naive context, but two made estimates for ESA‐naive groups available [39, 49]. Among patients with international prognostic scoring system (IPSS‐R) data, the majority of patients were of low‐risk (71.4%), while only 1.6% of patients were of high or very high risk. For the molecular alternative (IPSS‐M), 82.5% were of either very low, low, or moderate low risk. Within RWE, seven studies were conducted in the USA (58.3%), but two included centers from Asia (16.7%), and four included European centers (33.3%). Within the clinical trials, all were multicenter. Specific characteristics of real‐world studies can be seen in Table 1, and of clinical trials on Supporting Information S1: Table S4.

TABLE 1.

Main baseline characteristics of the included real‐world studies.

Study author	Study year	Number of patients on luspatercept	Country	Prior ESA usage (%)	Number of males (%)	Mean age ± SD (years)	Transfusion dependency at enrolement (%)
Andritsos	2025	871	USA	84.7	501 (57.5)	74.7 ± 6.4	67.6
Consagra	2025	331	Italy/USA	95.8	211 (63.7)	74.1 ± 8.6	93.7
Farrukh	2022	39	USA	95.0	30 (76.9)	74.0 ± 9.2	79.5
Mukherjee	2024	253	USA	87.0	133 (52.6)	72.5 ± 9.6	91.7
Heyrman	2024	77	Belgium	70.1	51 (66.2)	78.4 ± 7.8	90.9
Jonasova	2024	54	Czech	83.3	33 (61.1)	74.4 ± 8.8	100.0
Koons	2023	17	USA	94.1	13 (76.5)	79.9 ± 2.1	100.0
Komrokji	2023	28	USA	85.7	19 (67.9)	72.4 ± 11.0	60.7
Mariani	2025	40	Italy	100.0	25 (62.5)	70.2 ± 6.8	100.0
Kometas	2025	37	USA	92.0	18 (48.6)	74.5 ± 9.9	81.1
Zhang	2025	60	China	100.0	38 (63.3)	65.9 ± 9.9	51.7
Liang	2025	14	China	42.9	6 (42.9)	59.8 ± 7.1	78.6

Abbreviations: ESA, erythropoietin stimulating agents; SD, standard deviation.

3.3. TI Outcome

For 8 and 12 weeks TI, results from clinical trials informed prior construction, while only one trial contributed data for 16‐week TI (Supporting Information S1: Table S5). The estimated 8 weeks TI rate was 44.7% (95% CrI: 28.6%–61.5%, Figure 1A), with extreme heterogeneity (τ = 1.21; 95% CrI: 0.70–2.06) and a 35.7% probability of being lower than the trial‐derived estimate. For 12 weeks TI, the pooled rate was 38.6% (95% CrI: 21.1%–60.0%, Figure 1B), again with extreme heterogeneity (τ = 1.24; 95% CrI: 0.50–2.73) and a 76.95% chance of falling below the trial‐based estimate. At 16 weeks, TI was achieved in 30.9% of patients (95% CrI: 10.7%–53.9%, Figure 1C), with similarly high heterogeneity (τ = 1.18; 95% CrI: 0.33–3.02) and a 14.45% chance of being below the trial estimate. Results utilizing less informative priors can be seen in Supporting Information S1: Table S6. The results of subgroup analyses can be seen in Supporting Information S1: Table S7 Among the covariates tested in the meta‐regression (Table 2), the percentage of male patients was the only adequate moderator, showing strong evidence of association with TI at 8 weeks (P[D] = 99.8%, Figure 2A) and 12 weeks (P[D] = 98.8%, Figure 2B).

FIGURE 1.

Bayesian meta‐analytical pooling of 8‐week transfusion independence [A], 12‐week transfusion independence [B], 16‐week transfusion independence [C], and hematological improvement [D]. Estimates shown from individual studies were submitted to Bayesian shrinking and, therefore, are not the exact estimates extracted from these studies. The distribution at the right shows the effect distribution of the pooled effect, with the black line representing the median (estimated pooled effect), and the two dotted lines representing the first and third quartiles (95% credible interval). CrI, credible interval.

TABLE 2.

Results of the meta‐regression analysis.

Covariate	Probability of direction (%)
	HI‐E	8‐week TI	12‐week TI	16‐week TI
Male percentage within the study	99.2*	98.2*	98.8*	88.2
Percentage of transfusion dependent patients in the study	65.3	74.3	68.5	93.0
Total sample size of the study	87.7	88.0	75.8	69.7
Percentage of prior ESA usage within the study	81.2	81.5	85.3	70.1
Mean age within the study	82.0	85.0	85.2	65.4
Percentage of low‐risk patients in the study (IPSS‐R)	73.1	75.6	51.0	65.5
Percentage of low‐risk patients in the study (IPSS‐M)	77.6	77.3	71.7	75.2

Note: The * indicates a posterior probability exceeding 95%, the threshold chosen for an influential effect.

Abbreviations: ESA, erythropoietin stimulating agents; HI‐E, hematological improvement; IPSS‐M, molecular international prognostic scoring system; IPSS‐R, revised international prognostic scoring system; TI, transfusion independence.

FIGURE 2.

Meta‐regression analyses derived from the Bayesian model for male percentage's effect on transfusion independence for 8 weeks [A], transfusion independence for 12 weeks [B], and hematological improvement [C]. In these models, the odds ratio per 1% increase in male percentage reflects the association between the proportion of male patients in each study and the average treatment response. Panel [D] displays adverse events reported in at least two clinical trials or two real‐world studies. Bars reflect the percentage of patients affected per design type. Absence of a bar on either side indicates that the event was not reported in at least two studies of that design. CrI, credible interval; HI‐E, hematological improvement; OR, odds ratio; TI, transfusion independence.

3.4. HI‐E Outcome

Five clinical trials reported results on HI‐E, which were used to define the priors for the Bayesian model (Supporting Information S1: Table S5). In the Bayesian analysis, a HI‐E percentage of 46.6% (95% CrI 32.5%–63.9%, Figure 1D) was observed across 595 patients from the seven studies included in this analysis, with fairly high heterogeneity (τ = 0.99, 95% CrI 0.45–2.00). Based on this pooling, the probability of the effect being lower than the effect estimated by the clinical trial frequentist pooling was 98.6%. Results utilizing alternative priors can be seen in Supporting Information S1: Table S6. Results from subgroup analyses can be seen in Supporting Information S1: Table S8. In meta‐regression analysis (Table 2), only male percentage emerged as an adequate moderator across studies (P[D] = 99.2%, Figure 2C). The prior and posterior estimates for each prior in TI and HI‐E analysis can be seen in Supporting Information S1: Figures S2 and S3.

3.5. AE Outcome

Considering the unstandardized reporting in AEs (i.e., some studies only report the most frequent adverse effects), we chose a descriptive rather than meta‐analytical approach. AE data were available from all five clinical trials [5, 18, 40, 50, 51] and six real‐world studies [42, 43, 44, 46, 48, 52]. In the trials, 81 grade 3 or higher events were reported, corresponding to 19.8 events per 100 patients, along with 25 treatment discontinuations due to AEs (6.1 per 100 patients). The most commonly reported AE in clinical trials was back pain (19.0 per 100 patients; Supporting Information S1: Table S9). Among the RWE, only one reported the number of grade 3 or higher AEs, at 3.3 per 100 patients, with 7.4 treatment discontinuations per 100 patients. The most frequent AE in real‐world data was severe hypertension (27.0 per 100 patients; Supporting Information S1: Table S10). The AEs assessed in at least two studies in either category (trial or RWE) can be seen in Figure 2D.

3.6. Overall Survival Outcome

Only three real‐world studies [20, 41, 44] reported OS using KM curves (Figure 3A). A total of 428 patients were included, with estimated survival rates of 88.9% (95% CI, 85.9%–92.1%) at 1 year and 74.4% (95% CI, 70.0%–79.2%) at 2 years following treatment initiation (Figure 3B). Notably, when comparing the only Asian cohort with the two other studies, comprising exclusively American or mixed American and European populations, a hazard ratio of 0.19 (95% CI, 0.05–0.78) was observed (Figure 3C). All survival analyses were truncated at 24 months, as the study by Zhang and colleagues [20] only reported data up to approximately 25 months post‐treatment initiation.

FIGURE 3.

Kaplan‐Meier survival estimates from reconstructed individual‐patient data from the real‐world studies that reported plots [A], the pooled data [B], and stratified by continent [C].

3.7. Bias Assessment

Of the two non‐randomized clinical trials assessed, both were classified as having serious concerns due to the lack of adjustment for confounders (Supporting Information S1: Table S11). In regards to the randomized studies, MEDALIST was considered of low risk of bias, while COMMANDS had some concerns due to its open‐label design (Supporting Information S1: Table S12). Furthermore, seven real‐world studies were considered of good quality, although five graded as of poor quality (Supporting Information S1: Table S13). Although all of the poolings were fairly heterogeneous, Peter's test did not point towards significant asymmetry in funnel plot analysis for HI‐E, 8 weeks TI, 12 weeks TI, or 16 weeks TI (p‐value = 0.88, 0.22, 0.47, and 0.30, respectively). Funnel plots can be seen in Supporting Information S1: Figure S4 for each outcome.

4. Discussion

RWE plays an increasingly important role in evaluating the effectiveness, safety, and utilization of new therapies in broader and more heterogeneous patient populations than those typically represented in clinical trials [53, 54]. This is especially important in MDS, which is the most heterogeneous bone marrow cancer [55]. As a result, regulatory agencies have progressively incorporated RWE to support drug approvals [56, 57, 58]. Given the pronounced phenotypic heterogeneity of MDS [59, 60, 61], extrapolating clinical trial findings to the general patient population is challenging and often insufficient. Moreover, as luspatercept moves into later stages of evaluation, such as in the ongoing ELEMENT‐MDS trial involving non–transfusion‐dependent MDS patients [62], the pool of eligible participants is expected to narrow. As such, RWE may play an important role in complementing traditional clinical trial data and informing treatment strategies in MDS.

Real‐world studies are more susceptible to selection bias and often reflect lower adherence to guideline‐recommended therapies compared to clinical trials [63]. These and other factors, such as differences in monitoring intensity and patient comorbidities, can lead to different effect estimates in real‐world settings [64]. In our analysis, incorporating RWE led to a downward shift in the estimated HI‐E response, with a 98.6% probability that the real‐world effect is lower than the estimate derived from clinical trials. There are various possible explanations for the downward shift, but we believe that the high number of cases with SF3B1 mutation in the clinical trials (mainly the MEDALIST and COMMANDS trials) can explain part of this divergence. Of note, our subanalysis of HI‐E in patients with SF3B1 mutations displayed a response in 52% of patients (Supporting Information S1: Table S9).

These findings are relevant not only for adjusting physician and patient expectations but also for highlighting potential barriers to effective luspatercept use, including suboptimal adherence and inappropriate prescribing. Although luspatercept's administration every 21 days represents an improvement over the more frequent multiple dosing per week required for ESAs, it is well‐documented that adherence to subcutaneous therapies tends to be poor in real‐life scenarios, particularly in comparison to trials [56, 65]. For instance, in the context of subcutaneous immunotherapy, only 34%–50% of patients complete an adequate course of treatment [66, 67, 68]. Additionally, Asian studies reported higher rates of HI‐E, OS, and TI compared to those from Europe and the United States. This aligns with previous evidence suggesting that Asian patients with MDS often experience a milder disease course, with lower rates of progression to AML and longer survival [69, 70]. This apparent regional difference should be interpreted cautiously, as only two Asian cohorts were available. It is therefore more appropriate to consider this observation as hypothesis‐generating, potentially influenced by population‐specific genetic profiles or healthcare system differences.

Estimates for TI in different timeframes did not have a high probability of diverging from trial data. This may be due to lower reporting bias, as transfusions are generally recorded in central registries or electronic health records [71, 72], not relying on lab interpretations or borderline cases as much as HI‐E. Moreover, all clinical trials assessed HI‐E during weeks 1–24 of treatment, whereas only four real‐world studies used the same timeframe. This suggests that some RWE may have included shorter or inconsistent periods of assessment. Indeed, in our subgroup analysis restricted to real‐world studies that explicitly evaluated HI‐E during weeks 1–24, the estimated effects aligned more closely with those from clinical trials (Supporting Information S1: Table S9).

We found that HI‐E, 8, and 12 weeks TI had posterior estimates greatly influenced by the male percentage within studies, with a greater percentage translating into lower estimates (Figure 2A–C). This aligns with evidence indicating that male MDS patients tend to have a higher burden of somatic mutations and reduced overall and progression‐free survival [73, 74, 75, 76]. In the recently developed IPSS‐M, males are associated with a significantly worse prognosis [77]. Although male sex has not yet been explicitly investigated as a predictor of luspatercept response, we hypothesize that the observed effect may reflect the greater biological propensity among males to progress to advanced disease stages, such as high‐risk MDS or secondary AML [76]. While this finding is hypothesis‐generating [37], it warrants further investigation. Interestingly, women are known to respond better to ESAs in MDS and myelofibrosis [78, 79]. Although individual patient‐level data were unavailable, our synthesis highlights SF3B1 mutation and sex as consistent effect modifiers of luspatercept effectiveness. Other potentially relevant determinants, such as IPSS‐M risk category, transfusion dependence, prior ESA exposure, and age, could not be quantitatively evaluated due to incomplete reporting in most real‐world cohorts.

This analysis has several limitations that should be accounted for. First, Bayesian analyses are highly dependent on prior specification, and some subjective choices in prior definition may affect results [80], although this was explored in our subanalyses. Second, our results were mostly burdened by high heterogeneity, although this is expected in meta‐analyses of proportions and largely tackled by the Bayesian model [81, 82]. Third, data on specific subgroups are still sparse, and much of our subgroup analyses are based on a low number of studies. Finally, we were not able to perform an evaluation of which oncogenic mutations influence Luspatercept response in MDS.

In summary, by formally integrating real‐world evidence with clinical trial data, this study moves beyond descriptive pooling to quantify how effectiveness estimates for luspatercept evolve when routine clinical practice is incorporated into inference. This approach allows estimation of the probability and direction of attenuation relative to trial expectations, identifies patient‐ and study‐level factors that may modify treatment response, and characterizes the uncertainty surrounding transfusion independence and hematologic improvement across heterogeneous MDS populations. As the use of luspatercept expands, continued incorporation of real‐world data will be essential to refine expectations, inform hypothesis generation, and contextualize trial results in everyday practice. Despite limitations inherent to observational data and Bayesian modeling, this synthesis demonstrates how real‐world evidence can meaningfully update trial‐based knowledge and support more nuanced interpretation of treatment outcomes in MDS.

5. Critical View

Luspatercept was approved for the treatment of anemia in lower‐risk, transfusion‐dependent myelodysplastic neoplasms primarily on the basis of randomized phase III clinical trials. However, myelodysplastic neoplasms represent a biologically and clinically heterogeneous group of disorders, and outcomes observed in routine practice may differ from those reported in tightly controlled trial settings. A review of the published literature reveals that available real‐world studies of luspatercept are frequently limited by small sample sizes, heterogeneous patient populations, and inconsistent reporting of key disease characteristics, including prior erythropoiesis‐stimulating agent exposure, transfusion burden, and molecular features such as SF3B1 mutation status. These limitations have hindered the ability to contextualize trial‐derived expectations and to interpret variability in reported outcomes across real‐world cohorts.

This review distinguishes itself from existing narrative and real‐world summaries by applying a Bayesian evidence synthesis framework that formally integrates clinical trial data with observational real‐world evidence. By using trial results as informative priors and allowing real‐world data to update inference, we move beyond descriptive pooling to quantify how effectiveness estimates evolve when applied in routine clinical practice. Across five clinical trials and twelve retrospective cohorts comprising 2261 patients, we derived pooled estimates for hematologic improvement‐erythroid and transfusion independence, evaluated differences between trial and real‐world settings, and explored potential modifiers of response in a hypothesis‐generating manner. This approach clarifies the probability and direction of attenuation relative to trial‐based expectations and highlights patient‐ and study‐level factors, such as SF3B1 mutation status, sex, and geographic region, that may contribute to heterogeneity in observed outcomes.

Taken together, the available evidence supports the effectiveness of luspatercept in achieving transfusion independence and hematologic improvement in real‐world settings, while suggesting that some outcomes (particularly hematologic improvement‐erythroid) may be less pronounced than those reported in clinical trials. Rather than proposing changes to established treatment paradigms, this synthesis provides a structured framework for interpreting trial results in the context of routine care and for understanding sources of variability across patient populations. As real‐world datasets continue to expand, approaches that explicitly integrate observational evidence with trial data may help refine expectations, identify gaps in evidence, and inform the design of future prospective studies aimed at underrepresented subgroups within MDS.

Funding

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting Information S1

Data Availability Statement

The data is available upon reasonable request to the corresponding author.