Matching-adjusted indirect comparison of zanubrutinib and real-world chemotherapy, rituximab and chemoimmunotherapy in relapsed or refractory marginal zone lymphoma

For this MAIC, the zanubrutinib population was defined as the pooled efficacy-evaluable R/R MZL population from the contributing trials. This included 66 patients from the Phase II, single-arm MAGNOLIA trial and 20 patients from BGB-3111-AU-003 [14]. BGB-3111-AU-003 was a broader Phase I/II, single-arm trial in patients with B-cell malignancies and only the R/R MZL subset was included in the present analysis [14]. In both zanubrutinib trials, patients must have had documented failure to achieve at least a partial response or documented disease progression after their last systemic treatment and received at least one prior line of treatment. In MAGNOLIA, patients were also required to have received a prior anti-CD20 regimen. Both studies were multicenter with sites across Europe, North America, Asia and Oceania. Median study follow-up was 28.0 months in MAGNOLIA (data on file) and 35.2 months for patients with R/R MZL in BGB-3111-AU-003. Disease progression was assessed by an independent review committee using the Lugano 2014 classification criteria [16]; progression-free survival (PFS) in both trials was defined as the time from treatment initiation to date of disease progression or death, whichever occurred first. Given that patients recruited into the two trials were broadly similar, individual patient-level data from MAGNOLIA and BGB-3111-AU-003 were pooled for the analysis (hereafter referred to as MAGNOLIA-003).

The HMRN registry represents one of the largest registries in Europe with information on lymphomas and other blood disorders from a population-based patient cohort. The HMRN registry is a collaborative venture initiated in 2004 by researchers at the University of York, York, UK and National Health Service clinicians working across 14 hospitals in the UK.

Between 2005 and 2020, the registry collected medical-record data from 2039 patients with newly diagnosed MZL treated with an anti-CD20-based therapy, 155 of whom went on to receive further treatment [17]. For the purposes of the current analysis, the data were restricted to 90 patients most comparable to MAGNOLIA-003 based on the following inclusion criteria: patients who enrolled in the registry from 2014 onwards (to align with the recruitment period of the zanubrutinib trials), had an Eastern Cooperative Oncology Group performance status score between 0 and 2 at time of entry into the registry, received at least one anti-CD20-based therapy, and received further treatment recommended by European Society for Medical Oncology 2020 guidelines after receipt of prior anti-CD20-based therapy. Patient groups were not further separated into individual CIT or chemotherapy treatments due to insufficient sample sizes. The individual treatment regimens received in this final cohort of 90 patients are listed in Supplementary Table 1. The HMRN comparator cohort included patients treated with chemotherapy, rituximab monotherapy or rituximab-based CIT. These regimens were combined into a single comparator group because they represent established real-world treatment approaches used in R/R MZL and because the number of patients receiving each individual regimen was insufficient to support robust regimen-specific MAICs. The combined comparator therefore reflects routine-practice treatment heterogeneity rather than a single uniform intervention. Given the potential for differences in outcomes across comparator regimens, a sensitivity analysis excluding patients who received chemotherapy-only regimens was conducted to evaluate whether results were consistent when the comparator was restricted to rituximab monotherapy or rituximab-based CIT. PFS was defined as time from treatment initiation to date of disease progression or death due to any cause based on medical-record data.

The present study was a secondary analysis of aggregate registry data shared according to data use agreement between HMRN and the sponsor. The pivotal study that was used in this analysis had the required ethical approval and informed consent. Detailed study design and patient eligibility criteria from MAGNOLIA, BGB-3111-AU-003 and the HMRN registry are available in Supplementary Table 2.

The approach used for the MAIC was in accordance with guidance published by the National Institute for Health and Care Excellence Decision Support Unit Technical Support Document 18 [18]. Because the zanubrutinib trials were single-arm studies and no common comparator arm was available, an unanchored MAIC was performed. This approach assumes that all clinically relevant prognostic factors and treatment-effect modifiers that differ between the trial and external comparator populations are measured and adequately adjusted for. Individual patient data from MAGNOLIA-003 were weighted such that the mean baseline characteristics of greatest prognostic importance matched those in the HMRN registry. This was achieved using propensity scores estimated from a logistic regression model. Candidate variables for the MAIC model were identified based on clinical expert input, published prognostic relevance or clinical plausibility in R/R MZL, and availability across both the zanubrutinib trial datasets and the HMRN registry. Compared with MAGNOLIA-003, fewer patient characteristics were collected in the registry due to the real-world nature of the dataset. The final matching model included variables that were consistently available across datasets and considered relevant to prognosis or treatment context: number of prior lines of therapy, refractory status to last systemic therapy, progression of disease within 24 months of treatment initiation, age and time since diagnosis. Other clinically relevant variables, including MZL subtype, disease burden, lactate dehydrogenase level, disease stage, bone marrow involvement, extranodal involvement, detailed prior treatment regimen and comorbidities, were unavailable or inconsistently reported across datasets and therefore could not be included in the matching model.

In the MAIC model comparing zanubrutinib versus chemotherapy, rituximab or CIT, five characteristics of prognostic importance were included: number of prior lines of therapy (1, 2, >2), refractory to last prior systemic therapy (yes vs no), progression of disease within 24 months of treatment initiation, mean age (<65 vs ≥65 years) and median time since diagnosis. Weights to balance these characteristics were applied to the individual patient data from MAGNOLIA-003 prior to estimating relative treatment effects. Sensitivity analyses using the leave-one-out method, whereby one covariate was removed from the model at a time, were performed to assess the impact of each covariate on the results. As patients receiving non-rituximab-based chemotherapy have a generally less favorable prognosis compared with a rituximab monotherapy or rituximab-based chemotherapy regimen [17], an additional sensitivity analysis was performed to estimate the comparative efficacy of zanubrutinib versus rituximab only or CIT, whereby patients receiving chemotherapy alone were removed from the HMRN registry cohort.

The degree of overlap in the distribution of covariates between the zanubrutinib population and the comparison population was expressed in terms of the effective sample size (ESS) [18]. A Cox proportional hazards model was applied to weighted zanubrutinib individual patient data and registry data with overall survival (OS) and PFS results expressed as hazard ratios (HRs) with 95% CIs. Assessment of the proportional hazards assumption was performed using log-cumulative hazard plots and Schoenfeld residual tests [19,20]. All analyses were performed using R version 4.0.2 and using the ‘survival’ package to compare time-to-event outcomes.

Prior to matching, compared with the HMRN registry cohort, the MAGNOLIA-003 population was younger (mean age 68 vs 73 years), more heavily pretreated (55.8% vs 21.1% with two or more prior lines of therapy), and had a longer time since diagnosis (65.1% vs 50.0% with greater than or equal to median time since diagnosis). MAGNOLIA-003 also had a higher proportion of patients who were refractory to their last systemic therapy (30.1% vs 25.6%) and a lower proportion who had progressed within 24 months of treatment initiation (44.7% vs 51.1%). After applying weights, convergence was achieved using the full set of covariates, resulting in key prognostic characteristics being balanced (Table 1). The sample size of MAGNOLIA-003 was reduced by 56% after matching, resulting in an ESS of 38. The reduced ESS reflects the degree of overlap between the trial and registry populations after matching and should be considered when interpreting the precision of the adjusted estimates. In the sensitivity analysis of MAGNOLIA-003 versus the HMRN registry population receiving rituximab-based therapies (rituximab only or CIT), the ESS increased to 40, a 53% reduction from the sample size of 86 (Supplementary Table 3).

In the unadjusted comparison and MAIC, zanubrutinib was associated with significantly increased PFS (unadjusted HR: 0.47 [95% CI: 0.29–0.76]; adjusted HR: 0.30 [95% CI: 0.15–0.63]; Figure 1 & Table 2) and OS (unadjusted HR: 0.34 [95% CI: 0.19–0.61]; adjusted HR: 0.23 [95% CI: 0.10–0.50]; Figure 2 & Table 2) relative to chemotherapy, rituximab and CIT combined. Outcomes were generally consistent across the leave-one-out analyses (Table 2). Results of these analyses suggest that the number of prior lines of therapy had the largest impact on prognosis. In the unadjusted comparison and MAIC for the comparison with therapies containing rituximab, zanubrutinib was associated with significantly increased PFS (unadjusted HR: 0.51 [95% CI: 0.31–0.83]; adjusted HR: 0.28 [95% CI: 0.14–0.57]; Supplementary Figure 1) and OS (unadjusted HR: 0.37 [95% CI: 0.20–0.68]; adjusted HR: 0.23 [95% CI: 0.10–0.49]; Supplementary Figure 2) relative to rituximab or CIT.