The present study was performed and reported in adherence to the guidelines outlined in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses statement (19). The protocol for the present systematic review is registered with the International Prospective Register of Systematic Reviews (registration no. CRD42022300207).

A systematic search of the literature was performed across the following electronic databases: PubMed (https://pubmed.ncbi.nlm.nih.gov/), the Cochrane Library (https://www.cochranelibrary.com/) and Embase (https://www.embase.com/), covering all entries from database inception to April 12, 2023. The search strategy employed a combination of key words and Medical Subject Headings terms, including: ‘multiple myeloma’, ‘myeloma’, ‘plasma cell myeloma’, ‘autologous stem cell transplantation’, ‘autologous transplant’, ‘autograft’, ‘stem cell transplant’, ‘hematopoietic stem cell transplantation’, ‘tandem transplant’, ‘double transplant’ and ‘high-dose therapy’. Boolean operators (AND, OR) were applied to combine search terms appropriately. Restrictions on publication type were not applied. In addition to examining published trials, the World Health Organization International Clinical Trials Registry Platform (www.trialsearch.who.int) was also searched for any relevant registered trials. To ensure a broad scope, the review incorporated clinical guidelines from several countries. Moreover, in an effort to minimize publication bias, grey literature sources, including conference abstracts and academic dissertations, were also considered (20).

The present study exclusively included RCTs to minimize selection bias and confounding factors, thereby providing the highest level of evidence on efficacy. Studies were considered eligible for inclusion if they met all of the following criteria: i) RCTs; ii) enrolled patients had a confirmed diagnosis of symptomatic or progressive, previously untreated MM; iii) participants were randomized to undergo either a single HDT/ASCT or a double HDT/ASCT within 6 months after the first ASCT; iv) reported data on response rates, OS, EFS and/or progression-free survival (PFS) and treatment-related mortality (TRM) for both treatment arms; and v) published in the English language. Studies were excluded if they met any of the following criteria: i) Non-randomized study designs (such as observational studies, case series or case reports); ii) studies enrolling patients with relapsed or refractory MM; iii) studies comparing transplantation strategies other than single vs. double HDT/ASCT; iv) studies that did not report at least one of the prespecified outcomes of interest; v) duplicate publications reporting exactly the same patient cohort and identical outcomes without providing any new information (conference abstracts that supply additional subgroup analyses or long-term follow-up data not available in the corresponding full-text articles were considered complementary and were included after confirming no overlap in the extracted patient data); and vi) reviews, editorials, commentaries or guidelines without original data.

A total of two independent reviewers assessed each article against the eligibility criteria. A third reviewer was consulted to arbitrate any discrepancies that arose. Another pair of reviewers were responsible for extracting the following data from the studies that were included: i) General study information (first author, year of publication, study design and sample size); ii) baseline patient characteristics (diagnosis, prior treatments, sex and age); iii) details of the treatment protocols (patient groups, pre-treatment regimens, treatment plans, duration and follow-up period); and iv) outcomes measured for each group. The endpoints of primary interest were OS and PFS. Secondary outcomes encompassed response rate and TRM. Where available, data on adverse events were also extracted to evaluate safety.

For conference abstracts that provided additional subgroup or long-term follow-up data not reported in the corresponding full-text articles, only those unique estimates were extracted. It was verified that the same patient cohort was not counted twice for the same outcome in any meta-analysis.

The potential for bias in the included RCTs was evaluated using the Cochrane Risk of Bias tool, version 2 (RoB 2) (21). A summary of this assessment is presented in Table SI, with detailed domain-level judgments for each included study. Whilst the Grading of Recommendations, Assessment, Development and Evaluations (GRADE) framework is another important system for evaluating evidence certainty, the principal evaluation of methodological quality and bias risk was grounded in the Cochrane RoB 2 tool. This method is specifically tailored and extensively validated for use with RCTs. Consequently, a formal GRADE assessment was not performed for this analysis.

A table was constructed to summarize the characteristics of the included studies, which helped identify the data available for synthesis. For comparing categorical outcomes (such as response rate and TRM) between intervention groups, the risk ratio (RR) was employed. Estimates of the RR were calculated using Review Manager 5.3 software (The Cochrane Collaboration). Pairwise meta-analyses were performed, and the results are displayed as forest plots. A random-effects model was applied for all meta-analyses to incorporate potential heterogeneity arising from variations in study populations, protocols and clinical settings across the included trials. Heterogeneity was assessed using the Q test and the I² statistic for descriptive purposes.

For time-to-event data (OS, EFS and PFS), hazard ratio (HR) was used for inter-group comparisons. HR estimates were derived employing the inverse variance method in Stata 16.0 (StataCorp LP). In cases where HRs and their confidence intervals (CIs) were not directly reported in the publications, they were approximated from published Kaplan-Meier curves using established methodological approaches (22).

A network meta-analysis was performed to assess the relative effects of different treatment strategies (23), utilizing a mixed-effects model in R software, version 4.2.1 (https://www.R-project.org/). This methodology enabled the simultaneous comparison of multiple interventions, even in the absence of direct head-to-head trials, thus yielding a hierarchical ranking of treatment efficacy and improving the clinical relevance of the results. Inconsistency within the network was also assessed using Cochran's Q statistic and a design-by-treatment interaction model (24). The effects of each treatment regimen relative to single HDT/ASCT alone are presented in a forest plot with 95% CIs.

Publication bias was assessed using Egger's test for PFS and OS, and Harbord's test for overall response rate and treatment-related mortality (25,26). Additional methods for detecting publication bias have been described by Jin et al (27). For the network meta-analysis of OS, a funnel plot was generated.

The process of study identification is detailed in Fig. 1. Initially, 1,101 records were identified. Following a review of titles and abstracts, 18 records were selected for a full-text assessment based on the inclusion criteria. A total of 8 studies were considered potentially relevant. In the final selection, 6 full-text articles and 2 conference abstracts met the eligibility criteria. The 6 full-text articles (11,12,14,17,28,29) provided data for the conventional pairwise meta-analysis, encompassing 2,173 unique patients from the arms directly comparing double vs. single ASCT. The two conference abstracts (30,31) did not introduce additional patient cohorts; instead, they supplied supplementary subgroup analyses and extended follow-up data derived from the same trial populations already represented in the included full-text articles. No double counting of patients occurred in any meta-analysis. Among the 8 studies, 5 categorized patients into different risk groups according to adverse prognostic factors, and 3 were incorporated into the subgroup meta-analysis focusing specifically on high-risk patients.

The key features of the included studies are summarized in Table I. The median duration of follow-up across these studies varied from 38–134 months. The patient population in each treatment arm ranged between 76–501 individuals. All studies provided a direct comparison of single vs. double ASCT in patients with MM.

The study by Attal et al (11) involved previously untreated patients classified with Durie-Salmon stage I, II or III myeloma. All participants received 3–4 cycles of vincristine, doxorubicin, dexamethasone (VAD) induction chemotherapy prior to transplantation. The conditioning regimen for the single ASCT arm consisted of melphalan (140 mg/m²) combined with total-body irradiation. For the double transplant group, the first procedure used melphalan (140 mg/m²) alone, and the second procedure employed the same regimen as the single-transplant group. Maintenance therapy with interferon-α (IFN-α) was administered to all patients. Cavo et al (12) (the Bologna 96 trial) enrolled previously untreated patients with symptomatic MM. All participants received VAD induction therapy. The conditioning regimen for the initial transplant was melphalan (200 mg/m²). Patients assigned to the double transplantation group underwent a second transplant 3–6 months later, which utilized melphalan (120 mg/m²) combined with busulfan (12 mg/kg). IFN-α served as the maintenance therapy. This study also performed an analysis of outcomes specifically for patients who failed to achieve at least a near-CR after a single transplant. Both Mai et al (14) (the GMMG-HD2 trial) and Stadtmauer et al (17) (the BMT CTN 0702 trial) used high-dose melphalan (200 mg/m²) as the conditioning regimen. However, these trials did not include subgroup analyses focused on high-risk patients. Cavo et al (28) (the EMN02/HO95 MM trial) also reported outcomes for participants with high-risk cytogenetic profiles. Straka et al (29) analyzed pooled data from two clinical trials (NCT00416273 and NCT00416208), proposing that double ASCT may extend EFS in newly diagnosed MM. Finally, two conference abstracts (30,31) also provided data on EFS and OS for the overall patient population and for subsets with adverse prognostic factors.

The results of the risk of bias assessment are presented in Table SI, with detailed domain-level judgments for each included study. Whilst all studies described their randomization procedures, a high risk-of-bias related to deviations from the intended interventions was identified. This was primarily due to the inherent difficulty, if not impossibility, of blinding both participants and healthcare providers to the transplantation assignment. In certain instances, participants did not adhere to the study protocol, partly due to their awareness of the assigned intervention and partly due to disease progression. All studies reported performing intention-to-treat analyses. The two conference abstracts (30,31) lacked detailed information on random sequence generation and the handling of missing data. Overall, the risk-of-bias for all included studies was judged to entail ‘some concerns’. A detailed account of the methodological quality assessment is provided in Table SI.

Fig. 2 shows response rates and TRM and Fig. 3 presents survival outcomes for high-risk patients.

Data on response rates were available from 5 studies (11,12,14,17,28). The meta-analysis of response rates [achievement of at least very good partial response (VGPR)] included 4 studies that provided sufficient data for pooling, with results shown in Fig. 2A. Cavo et al (12) reported a significant increase in the CR/near-CR rate from ~33% with single ASCT to ~47% with double ASCT (P=0.008). Mai et al (14) also reported a significant rise in CR rates from the first to the second HDT/ASCT procedure (P=0.04). By contrast, Attal et al (11) reported no statistically significant difference in the CR/VGPR rates between the single and double transplant groups (42 vs. 50%; P=0.10). Similarly, Stadtmauer et al (17) reported no significant advantage for achieving a response of at least a VGPR with double vs. single HDT/ASCT (P=0.37). The pooled RR for the ORR (defined as at least partial response) showed no significant difference between the treatment groups (RR, 1.03; P=0.42; data not shown). However, using a random-effects model, the combined RR for achieving a response of at least a VGPR was 1.17 (95% CI, 1.03–1.33; P=0.02), demonstrating a statistically significant advantage for the double HDT/ASCT approach (Fig. 2A). The analysis indicated no significant heterogeneity among the studies (I²=14.3%; P=0.321).

A total of 3 studies provided data on TRM (11,12,17). The corresponding forest plot is presented in Fig. 2B. No heterogeneity was detected among these studies (I²=0%; P=0.626). Furthermore, the results indicated no statistically significant difference in the risk of TRM between the double and single HDT/ASCT groups (RR, 1.54; 95% CI, 0.79–3.01).

For this specific analysis, patients were classified into a high-risk category if they possessed at least one of the following adverse prognostic factors: i) Failure to achieve at least a near CR or VGPR following induction therapy or the first transplant; ii) high-risk cytogenetic abnormalities, as defined by the original studies [such as the presence of t(4;14), t(14;16) or del(17p)]; and iii) International Staging System (ISS) stage II or III disease (32).

A total of 4 studies reported on PFS in patients characterized by the aforementioned adverse factors (12,28,30,31). However, only 2 studies (28,30) provided extractable HR estimates with CIs suitable for quantitative pooling. Cavo et al (12) reported subgroup findings without providing usable HRs and Rocchi et al (31) presented median PFS values without HRs. To avoid potential double-counting of patients from overlapping cohorts in the conference abstracts, the most complete and non-duplicative data were prioritized. Therefore, data from Cavo et al (28) and Cavo et al (30) were pooled in the meta-analysis (Fig. 3A). Cavo et al (30) reported that double HDT/ASCT was associated with a significantly longer PFS for patients possessing one or two adverse factors (HR, 0.70; P=0.006). Rocchi et al (31) reported a significant PFS benefit from double ASCT in patients with high-risk cytogenetics (median PFS, 36 months for double ASCT vs. 20 months for single ASCT; P=0.032). The other 2 studies reported findings consistent with these results. The pooled HR was 0.58 (95% CI, 0.43–0.80; P=0.001; Fig. 3A), indicating that double HDT/ASCT significantly extended PFS in patients presenting with at least one adverse prognostic factor.

OS data for patients with adverse factors were available from 4 studies (12,28,30,31). Cavo et al (30) reported that patients with two adverse factors who received double HDT/ASCT had a significantly longer OS compared with those planned for single HDT/ASCT (HR, 0.32; P=0.001). The OS advantage was particularly notable for patients who failed to achieve a CR after induction and who also had high-risk cytogenetics (HR, 0.22; P=0.001) or ISS stage III MM (HR, 0.42; P=0.033). Rocchi et al (31) reported an OS benefit with double HDT/ASCT in the subgroup with high-risk cytogenetics (10-year OS rates, 51% for double ASCT vs. 34% for single ASCT; P=0.004). A subgroup meta-analysis for OS was performed using the 3 studies that provided extractable HRs with CIs for patients with adverse prognostic factors (12,28,30). Rocchi et al (31) reported 10-year OS rates but did not provide an HR suitable for pooling and was therefore excluded from the quantitative synthesis. The pooled analysis demonstrated no significant heterogeneity (I²=4.7%; P=0.350). Moreover, the combined HR revealed a statistically significant OS benefit favoring double ASCT for patients with at least one adverse prognostic factor (HR, 0.70; 95% CI, 0.54–0.90; P=0.006; Fig. 3B).

For the pairwise meta-analyses, each outcome included <10 studies; therefore, funnel plots were not generated. Instead, quantitative tests were used to assess publication bias. Egger's test detected no significant bias for PFS (P=0.437) or OS (P=0.725). Harbord's test showed no significant bias for overall response rate (P=0.947) or treatment-related mortality (P=0.470). For the network meta-analysis of overall survival, a funnel plot is presented in Fig. 4; the symmetrical distribution of points suggests a low risk of publication bias across the network.

A network meta-analysis incorporating 6 studies evaluated PFS across three strategies: Double HDT/ASCT, single HDT/ASCT augmented with consolidation therapy and single HDT/ASCT alone (11,12,14,17,28,29). The network structure for these comparisons is illustrated in Fig. 5 (PFS network) and Fig. 6 (OS network). Fig. S1, Fig. S2, Fig. S3 provide network meta-analysis forest plots for PFS with different references. When double HDT/ASCT was used as the reference (Fig. S1), single HDT/ASCT alone (HR, 1.2; 95% CI, 1.1–1.4) and single HDT/ASCT with consolidation therapy (HR, 1.0; 95% CI, 0.8–1.4) did not show a PFS benefit. With single HDT/ASCT alone as the reference (Fig. S2), double HDT/ASCT demonstrated a significant PFS benefit (HR, 0.83; 95% CI, 0.72–0.94), whereas single HDT/ASCT with consolidation therapy did not (HR, 0.84; 95% CI, 0.6–1.1). When single HDT/ASCT with consolidation therapy was used as the reference (Fig. S3), no significant difference was observed for double HDT/ASCT (HR, 0.99; 95% CI, 0.7–1.3) or single HDT/ASCT alone (HR, 1.2; 95% CI, 0.9–1.6).

Another network meta-analysis, which included all studies, evaluated OS among the same set of treatment strategies (11,12,14,17,28–31). The contribution of direct and indirect evidence to the OS network estimates is illustrated in Fig. 7, and the forest plots for all pairwise OS comparisons are consolidated in Fig. 8. Assessment for potential publication bias within the OS network is presented in Fig. 4, and loop-specific inconsistency is shown in Fig. 9. Neither double HDT/ASCT (HR, 0.86; 95% CI, 0.71–1.10) nor single HDT/ASCT with consolidation therapy (HR, 0.76; 95% CI, 0.40–1.40) showed a statistically significant OS benefit compared with single HDT/ASCT alone (Fig. 8B). Furthermore, no significant difference in OS was demonstrated between double HDT/ASCT and single HDT/ASCT supplemented with consolidation therapy (HR, 1.10; 95% CI, 0.60–2.00; Fig. 8C). The funnel plot (Fig. 4) demonstrates symmetry around the vertical line, with points evenly distributed, suggesting a low risk of publication bias across the studies included in the network meta-analysis. The 95% CI of the odds ratios (1.00–3.07) from Fig. 9 included 1, indicating no statistically significant inconsistency between the direct and indirect evidence within the network. This supports the consistency and validity of the network meta-analysis results.