Efficacy of CARVYKTI in CARTITUDE-4 versus other conventional treatment regimens for lenalidomide-refractory multiple myeloma using inverse probability of treatment weighting

IPD were available from all data sources. The CARTITUDE-4 cohort consisted of patients who were randomized to the cilta-cel cohort (all patients underwent apheresis). Comparison of the cilta cel cohort was done versus each of the comparator cohorts, which consisted of all eligible patients who had received one of the following comparator treatments: DVd, Vd, DKd, Kd, or Pd. Patients from the daratumumab clinical trials were included in the comparator cohorts if they met at baseline the following key inclusion criteria from the CARTITUDE-4 trial: have received one to three prior LOTs including a PI and an IMiD, were refractory to lenalidomide and Eastern Cooperative Oncology Group (ECOG) status 0–1. As none of the patients in the daratumumab clinical trials had prior treatment with anti-CD38, patients with prior exposure to anti-CD38 therapies in the cilta-cel cohort were excluded. The index date was defined as the date of randomization within each trial.

Comparisons of outcomes between studies may be prone to bias due to confounding if not properly adjusted [11]. IPTW using average treatment effect in the treated (ATT) weights derived from propensity scores, was used to emulate hypothetical comparative trials in which patients were randomized to cilta-cel or one of the comparator regimens. Prognostic baseline characteristics for adjustment were identified and ranked in order of importance prior to the present analysis, based on input from independent clinical experts [12,13]. The following five factors were identified as most prognostic and clinically relevant: refractory status, cytogenetic risk [14], International Staging System (ISS), presence of plasmacytomas/extramedullary disease and time to progression in prior line. However, due to the high proportion of patients with unknown cytogenetics in the daratumumab clinical trials (ranging from 25% to 60%), cytogenetic risk could not be included. The remaining four factors constitute the base case adjustment set. A full list of the prognostic factors and their availability is shown in Supplementary Table 2.

Outcomes of interest included overall response rate (ORR), rate of very good partial response or better (≥VGPR), rate of complete response or better (≥CR) and progression-free survival (PFS). ORR was defined as the proportion of patients who achieved partial response (PR) or better (stringent CR [sCR], CR, VGPR, or PR). ≥VGPR was defined as the proportion of patients who achieved a VGPR, CR, or sCR based on International Myeloma Working Group (IMWG) consensus criteria. ≥CR was defined as the proportion of patients who achieved a CR or sCR based on IMWG consensus criteria. PFS was defined as the time from index date to the date of progression or death, whichever occurred first.

IPTW with the base case covariates was used to generate adjusted comparisons between the cilta-cel cohort versus DVd, Vd, DKd, Kd and Pd [15]. Propensity scores were estimated using a logistic regression model that predicted assignment in the cilta-cel cohort as a function of key prognostic factors. The estimated propensity scores, p, were used to derive ATT weights, where patients in the comparator arms were assigned a weight of p/(1-p), while patients from the cilta-cel cohort were kept as observed (assigned a weight of one) [15]. Consequently, patients in the comparator arms with more similar baseline characteristics to the cilta-cel cohort received larger weights while less similar patients were down weighted to balance the two groups. To reflect the original sample size, a rescaled version of these weights was used (see details in Appendix B).

Relative efficacy was assessed for both the unadjusted (i.e., cilta-cel versus comparative treatment prior to IPTW) and for the adjusted (i.e., with IPTW) comparisons for all outcomes. For response outcomes (i.e., binary outcomes) (ORR, ≥VGPR and ≥CR), logistic regression and Poisson regression models [16] (with weights applied for the adjusted comparison) were used to estimate odds ratios (ORs) and rate ratios (RRs), respectively. For PFS (i.e., time-to-event outcome), a Cox proportional hazards model (with weights applied for the adjusted comparison) was used to derive hazard ratios (HRs). Corresponding 95% confidence intervals (CIs) were calculated, where the variance was estimated using the robust sandwich variance estimator [17,18]. Appropriateness of the proportional hazards assumption for PFS was assessed based on visual inspection of the log-cumulative hazard plot, visual inspection of the Schoenfeld residuals plot and performance of the Grambsch–Therneau test [19] (with a p-value less than 0.05 considered to indicate a violation of the assumption). In case of violations of the proportional hazards assumption, time-dependent HRs were estimated. Furthermore, the restricted mean survival time (RMST) difference [20], restricted at the minimum of the largest observed event time in each of the two arms, was additionally reported.

Variables with a proportion of missing values less than 25% were imputed using multiple imputation with chained equation. Statistical analyses and graphical visualization were conducted using SAS 9.4 (SAS Institute, NC, USA) and R version 4.0.3 (R Foundation for Statistical Computing, Vienna, Austria).

Several sensitivity analyses were conducted to assess the robustness of results. The first sensitivity analysis was an IPTW using average treatment effect in the control (ATC) weights, where patients in the cilta-cel cohort were reweighted to reflect each of the comparator cohorts, whereas the comparator cohorts were kept as observed. While the primary ATT-based analyses estimate comparative efficacy in the cilta-cel population, ATC weighting provides comparative estimates in the patient cohort for the comparator. When the sample size of the external cohort is small, ATT weights may become unstable if adding several covariates in the PS model; therefore, ATC weights are a valid alternative [21]. As sample size for some of the comparator cohorts in the current analyses was limited, IPTW-ATC allowed for the adjustment of number of prior lines, years since diagnosis, age and hemoglobin levels, in addition to the covariates included in the base case. Finally, multivariable regression models were used as an alternative option to IPTW to explore the impact of further adding prior transplant, ECOG status, MM type, creatinine clearance, gender and race as covariates.

The comparative analysis consisted of 155 patients in the cilta-cel cohort of CARTITUDE-4 that had no prior exposure to anti-CD38 therapies, 44 patients treated with DVd, 46 patients treated with Vd, 98 patients treated with DKd, 46 patients treated with Kd, and 92 patients treated with Pd that fulfilled the CARTITUDE-4 eligibility criteria across the CANDOR, CASTOR and APOLLO studies (Figure 1).

Multiple imputation was required for the following missing variables in CANDOR: time to progression on prior line (missing 4.1% for DKd and 4.4% for Kd), years since MM diagnosis (missing 5.1% for DKd and 2.2% for Kd) and hemoglobin (missing 1% for DKd) (Supplementary Table 3). Imputation was not necessary for CARTITUDE-4, CASTOR and APOLLO, as there were no other missing values.

The cilta-cel cohort had a higher proportion of patients with ISS stage I, longer time to progression in prior line and a higher presence of plasmacytomas / extramedullary disease (Table 1). After IPTW, key baseline covariates were well balanced across the cohorts (Supplementary Tables 4–8).

The observed ORR was 89.7% in the cilta-cel cohort and ranged from 42.4% to 79.6% among patients treated with comparator regimens (Figure 2). After adjustment, patients in the cilta-cel cohort were 1.2 (vs DKd) to 2.0 (vs Pd) times more likely than patients treated with comparator regimens to achieve ORR (Figure 3). The observed ≥VGPR rates were 85.2% in the cilta-cel cohort and ranged from 14.1% to 64.3% in patients treated with comparator regimens (Figure 2). After adjustment, patients in the cilta-cel cohort were 1.5 (vs DKd) to 5.1 (vs Vd) times more likely to achieve ≥VGPR than patients treated with comparator regimens, and all comparisons were statistically significant (Figure 3). The observed ≥CR rates were 78.1% in the cilta-cel cohort and ranged from 2.2% to 26.5% in patients treated with comparator regimens (Figure 2). After adjustment, patients in the cilta-cel cohort were 2.9 (vs DKd) to 38.8 (vs Pd) times more likely to achieve ≥CR than patients treated with comparator regimens, all results were statistically significant (all p < 0.01) (Figure 3). Cilta-cel also significantly reduced the risk of disease progression or death by 49% (DKd: HR: 0.51 [95% CI: 0.30, 0.88; p = 0.015]) to 89% (Vd: HR: 0.11 [95% CI: 0.07, 0.17; p < 0.0001]). The median PFS for the cilta-cel and DKd arms was not reached whereas the median PFS post-adjustment was 9.79 months for DVd, 6.21 months for Vd, 11.09 months for Kd and 8.34 months for Pd (Figure 4). The assessment of proportional hazards indicated the proportional hazards held for comparisons versus all regimens, except Vd. Nevertheless, the overall HR represents the average HR across the entire treatment period, and remains a meaningful measure of average benefit [22]. Time-dependent hazard ratios showed an increasing benefit over time, with HR beyond 3 months ranging from 0.05 to 0.43 (Supplementary Table 9). Finally, the RMST difference after approximately 18 months ranged from 2.9 months (DKd) to 8.1 months (Vd), indicating that cilta-cel also significantly increased the average PFS time over all comparator treatments (Supplementary Figure 1).

Results from the IPTW-ATC analyses, including base case covariates and an additional four variables, were consistent with the main analysis (IPTW-ATT) with smaller standardized mean differences (Supplementary Tables 4–8), with cilta-cel demonstrating statistically significant improvement over comparator treatments for all outcomes. The multivariable regression models including all 14 available baseline covariates also showed consistency with the main analysis, with cilta-cel demonstrating improvement over comparator treatments for all outcomes (Supplementary Figures 2–5).