Treatment benefit of upfront autologous stem cell transplantation for newly diagnosed multiple myeloma: a systematic review and meta-analysis

This study was conducted in compliance with the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) guidelines [9]. Using the search terms “ASCT”, “MM”, “publication year 2012–2022” (which was extended to 2023/03 in revision), “randomized controlled trial”, “comparative study”, and “observational study” in the Embase, Cochrane Library, and PubMed (MEDLINE) databases, potentially eligible studies that met the following criteria were included: (1) subjects were newly diagnosed MM patients; (2) the study compared ASCT with no ASCT; (3) the study reported the post-ASCT/consolidation complete response (CR) rate, PFS or OS; and (4) the publication year was after 2012. The exclusion criteria were as follows: (1) review articles; (2) conference presentations; (3) subjects were refractory MM patients who received salvage ASCT or not; and (4) the study compared early vs. delayed ASCT.

The revised Cochrane risk-of-bias tool for randomized trials (RoB2) [10] and the Cochrane risk of bias tool for nonrandomized studies (ROBINS-I) [11] were adopted to assess the risk of bias in the RCTs and observational studies, respectively. The summary plots were generated using the robvis tool (https://​mcguinlu.​shinyapps.​io/​robvis/​) [12].

Four reviewers (CML, LCC, CYY, and FMT) independently reviewed the articles and abstracted the data. The following data were extracted from each selected article: study design, first author, publication year, number of patients, median or mean age of patients, proportion of patients with International Staging System (ISS) stage III classification, proportion of patients with high-risk cytogenetic features such as 1p deletion, 17p deletion, t(4;14) translocation, t(14;16) translocation, t(14;20) translocation, and 1q gain [13], median follow-up period, proportion of male patients, enrollment periods, treatment regimens, and outcomes including CR, PFS, and OS. If the article merely provided the ages in each subgroup, the median or mean age of the 2 arms was summarized and divided by 2 to denote the age of the overall population. The percentages of PI and IMiD utilization were either documented according to the exact records or estimated by equal division of the kinds of regimens if the real number was not available.

CR encompassed stringent complete response/remission (sCR), which is a more in-depth status of CR, while very good partial response (VGPR) was not counted. Only the post-ASCT/consolidation CR rate accounted for the overall CR rate, while the post-induction CR rate before ASCT/consolidation was excluded from the meta-analysis. For PFS and OS, the between-arm hazard ratio (HR) was extracted from each study if available. For those studies that did not report the HR directly, the HR was evaluated from the Kaplan‒Meier curve and the at-risk table. The value of survival probability in the curve was identified using WebPlotDigitizer software (version 4.5) (https://​automeris.​io/​WebPlotDigitizer​/​), and the estimated HR with its 95% confidence interval (CI) was then measured using a HR calculation spreadsheet created by Tierney et al. [14].

Statistical analysis was performed using Review Manager (RevMan) version 5.4 (The Cochrane Collaboration, Oxford, England) and Stata version 16.0 (College Station, Texas). Odds ratios (ORs) for the CR rate and HRs for PFS and OS were used to quantify the effect size. Based on a random effects model, the Mantel‒Haenszel (M-H) and inverse variance (IV) methods were used to analyze the OR and HR, respectively. The combined OR, HR, and their 95% CIs were illustrated by forest plots with 2 subgroups: RCTs and observational studies. Publication bias was assessed by the funnel plot in RevMan and the Egger’s test in Stata based on a random effects model and the restricted maximum likelihood (REML) method. The trim-and-fill method for adjusting for publication bias and the corresponding contour-enhanced funnel plot were implemented in Stata based on a random effects DerSimonian‒Laird method, which is the mode most comparable with the random effects M-H and IV methods in RevMan. In the contour-enhanced funnel plot, the imputed study located in the area of P > 0.1 represents a real adjustment for publication bias. Otherwise, the imputed study may adjust the plot asymmetry caused by other problems, such as heterogeneity. Heterogeneity between studies was evaluated by Cochran’s Q test and quantified with the I² statistic. An I² less than 25%, between 26–74%, and more than 75% denoted low, moderate, and high heterogeneity, respectively[15]. Univariate meta-regression and bubble plots were further used to depict the factors that possibly accounted for the heterogeneity.

A total of 10,493 articles were identified initially, including 1325 RCTs and 438 observational studies. After excluding the studies that did not meet the inclusion criteria, 5 RCTs and 15 observational studies were included in the meta-analysis to compare the treatment outcomes between upfront ASCT and no upfront ASCT (briefly no ASCT) [16‐35]. Two latest RCTs were added into the meta-analysis in the revised version after extending the enrolled publication term to 2023/03 [36, 37]. Figure 1 and Table 1 showed the research design and the baseline characteristics of the included studies. Among the 7 RCTs, 4 RCTs published during 2014 ~ 2020 were included in the previous versions of the review, whereas 3 RCTs published during 2021 ~ 2023 were newly enrolled [7, 8]. One of the 3 new RCTs (Gay et al., 2021) had a 3-arm design. In this RCT, 2 of the 3 arms (carfilzomib, lenalidomide, dexamethasone [KRD] + ASCT vs. KRD) were extracted for meta-analysis, while the KCD + ASCT arm (C: cyclophosphamide) was excluded due to a lack of comparative KCD alone group (Supplementary Table S1). The latest PI carfilzomib was used in 4 studies [24, 20, 34, 37], and the latest IMiD pomalidomide was mentioned in 1 study [25]. These 2 regimens were not included in the previous reviews [7, 8].

As seen in Supplementary Table S1, the RCTs were all open-label and multicenter, with a similar randomization process stratified by age, ISS, or other factors. The drop-out rates of these RCTs after randomization were predominantly higher in the no ASCT group than in the ASCT group, contributing to a possible attrition bias. In the RoB2 assessment (Fig. 2A and C), there were some concerns of bias due to deviations from the intended intervention in all 7 RCTs because of the inevitable open-label design. Five RCTs with obvious distinct drop-out rates between the ASCT and no ASCT groups raised some concerns of bias due to missing outcome data. Overall, the risk of bias in all 7 RCTs remained low. In contrast, the risk of bias in 6 observational studies was serious because they did not utilize common methodologies for bias control (Supplementary Table S2, Fig. 2B and D). Four observational studies were considered to have a moderate risk of bias because they utilized only one methodology to control for bias. The other 5 observational studies were regarded to have a low risk of bias because they executed a prospective intention-to-treat (ITT) cohort design or adopted multiple methods to control for bias and confounders. Regarding the study by Lemieux et al., multivariate regression adjustment was only applied for PFS, not for OS, increasing the bias risk of OS from moderate to serious. The PFS advantage of ASCT compared to no ASCT was strengthened after adjustment in the study by Lemieux et al., while the PFS and OS benefits of ASCT were attenuated after adjustment in the studies by Belotti et al. and Biran et al., respectively (Supplementary Table S3). For those observational studies that provided the adjusted HRs, the adjusted HRs were used for pooling outcomes. Otherwise, the unadjusted HRs and the HRs estimated from the Kaplan‒Meier curves were adopted for the following meta-analysis. Those unadjusted/estimated HRs were excluded from the sensitivity analysis to verify the robustness of the major outcomes.

As shown in Supplementary Table S3, the survival outcomes of 5 observational studies were transformed into HRs and added to the meta-analysis. Regarding CR, the M-H random method reported a significantly higher CR rate in the ASCT group (OR 1.24, 95% CI 1.02 ~ 1.51) (Fig. 3A). An I² of 44% indicated a moderate probability of heterogeneity, which was more likely to be attributed to the RCT by Yong et al. with a contrasting tendency. This non-inferiority trial only randomized the patients with at least partial responses to the induction chemotherapy, which may be profitable to the no ASCT group. The nonsignificant Egger’s test implied a low probability of publication bias. Though the sensitivity analysis that excluded 2 studies with a high risk of bias insignificantly favored ASCT, the other 2 kinds of sensitivity analyses that excluded the RCT by Yong et al. significantly favored ASCT over no ASCT for the CR rate (Supplementary Fig. S1A-C). The trim-and-fill method imputed 1 more study and still demonstrated a preference for ASCT for the CR rate (Supplementary Table S5 and Supplementary Fig. S2A). Regarding PFS, the IV random method depicted a significantly better PFS in the ASCT group (HR 0.53, 95% CI 0.46 ~ 0.62) (Fig. 3B). An I² of 81% and a significant Egger’s test (P = 0.03) implied a high possibility of heterogeneity and publication bias. Similar results could be seen in the sensitivity analysis, from which 3 studies with a high risk of bias were excluded (Supplementary Fig. S1D). The significant benefit of ASCT for PFS was sustained even after trim-and-fill imputation for publication bias adjustment (Supplementary Table S5 and Supplementary Fig. S2B). Regarding OS, although an overall significantly better prognosis was observed in the ASCT group (HR 0.58, 95% CI 0.50 ~ 0.69), there was no significant benefit for the ASCT group in the RCTs alone (Fig. 3C). An I² of 79% and a nonsignificant Egger’s test delineated a high level of heterogeneity but a low risk of publication bias. The sensitivity analysis that excluded 7 studies with a high risk of bias revealed a comparable result (Supplementary Fig. S1E). The trim-and-fill method also demonstrated a sustained advantage of ASCT for OS after imputation (Supplementary Table S5 and Supplementary Fig. S2C).

Seven possible characteristics of the enrolled patients contributing to the heterogeneity were identified after data analysis, including the number of patients, median or mean age, percentage of patients with ISS stage III, percentage of patients with high-risk genetic features, median follow-up duration, percentage of males, and percentage of enrolled periods after 2012 (Table 1). As shown in Fig. 4, meta-regression revealed that increased high-risk genetic features were significantly related to a better PFS and OS benefit of ASCT. Older median or mean age, increased percentage of patients with ISS stage III, decreased median follow-up duration, and decreased percentage of males were significantly associated with an improved OS with ASCT. Regarding the treatment regimens of PI, IMiD, combined PI/IMiD (at induction and consolidation phase), and triplet (a 3 or more combination of traditional agents/novel agents/steroids at induction phase), the meta-regression demonstrated that increased percentage of PI and combined PI/IMiD utilization was significantly associated with a lower PFS and OS benefit of ASCT, respectively (Fig. 5).