The literature search update yielded 1589 titles and abstracts. After duplicates were removed, 1403 titles and abstracts were assessed for eligibility. Of these, 19 articles were selected for full text assessment. In addition to the seven studies originally included in the review by Walter et al. [11], eight new publications met the eligibility criteria and were included [4, 25,26,27,28,29,30,31] (Fig. 1), resulting in a total of 15 studies. Of the eight new studies, three were RCTs [4, 30, 31] and five were observational studies [25,26,27,28,29].

Adapted from Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71

PRISMA 2020 flow diagram for the study selection process. aMEDLINE and EMBASE searches were made simultaneously in the embase.com search engine. bRecords excluded after assessment of title and abstract with the justification of not addressing the research question.

The key characteristics of the included studies are summarized in Table 2 [4, 25,26,27,28,29,30,31,32,33,34,35,36,37,38]. Of the new studies, none assessed SIRT; five studies compared TFD/TPI with REG [25,26,27,28,29], one compared TFD/TPI with a placebo group [30], and two compared REG with a placebo group [4, 31].

The risk of bias was assessed for randomised and non-randomised studies separately. All RCTs presented low risk of bias for both OS and PFS (Supplementary Materials Tables S3 and S4). Two non-randomised studies presented critical risk of bias and the other five presented serious risk of bias (Supplementary Materials Table S5). The main issues in these studies referred to patient selection and lack of information on adjustments for confounders. Although blinding is not mandatory and also not possible for some outcomes, the fact that patients and investigators knew the intervention introduces some level of bias.

Assessment of the certainty of the evidence (GRADE) is shown in Supplementary Materials Tables S6–S9. There was moderate confidence in the effect size estimate for OS in the pairwise comparisons of SIRT versus BSC and of TFD/TPI versus REG. In the pairwise comparisons of TFD/TPI versus BSC and REG versus BSC, there was a high confidence effect estimate for both OS and PFS. The comparison between TFD/TPI and REG provided low certainty in the estimated effect for PFS. There were no available data to compare SIRT and BSC for PFS.

The demographic and clinical characteristics of the study populations are summarized in Supplementary Materials Table S10. The median age ranged from 55 to 68 years and most patients were men. Three of the eight new studies assessed the presence of KRAS mutation [28, 30, 31], detected in half of the population in one study [28] and in approximately one-third of the sample in the other two studies [30, 31]. Two found extrahepatic metastases in more than 70% of the population [26, 31]. Nakashima et al. [27] found three or more metastatic sites including the liver in about one third of the sample (2165 of 7279 patients); Xu et al. [30] also found three or more metastatic sites in about one third of the sample (158 out of 406 patients); however, location was not specified.

Considering all 15 studies, ECOG performance status was not reported in three studies [25, 27, 32]. EHD was an exclusion criterion in one study [33]. Two studies evaluated EHD [32, 34] and four described the number of metastatic sites (including the liver) [26, 31, 32, 34].

The proportion of patients exposed to previous chemotherapy is summarized in Supplementary Materials Table S11. Among the 15 studies, three did not provide information on the number of prior chemotherapy regimens [4, 27, 33], and five did not report the chemotherapeutic agent used [25, 26, 28]. Of the studies providing this information, most reported two or three prior chemotherapy regimens.

Thirteen studies provide HR results. Four of them compared REG with BSC [4, 31, 35, 36], three compared TFD/TPI with BSC [30, 37, 38], three compared SIRT using Y-90 resin microspheres with BSC [32,33,34], and the other three compared TAS with REG [26,27,28].

The fixed effects model had the best fit, with lower residual deviance (DIC < 3), and was therefore chosen for the base-case analysis. Random effects model results were calculated in a sensitivity analysis, with no change in the overall interpretation of the results (Table 3). The results of individual studies can be seen in Supplementary Materials Table S12.

The HR (95% credible interval [95% CrI]) for OS considering SIRT versus BSC was 0.48 (0.27, 0.87); SIRT versus REG was 0.62 (0.32, 1.21); and SIRT versus TFD/TPI was 0.78 (0.4, 1.53). For TFD/TPI versus REG, HR was 0.79 (0.6, 1.06) for OS and 0.97 (0.65, 1.44) for PFS; for TFD/TPI versus BSC, HR was 0.62 (0.46, 0.83) for OS and 0.44 (0.31, 0.62) for PFS; and for REG versus BSC, HR was 0.78 (0.57, 1.05) for OS and 0.46 (0.33, 0.63) for PFS. HR estimates show SIRT using Y-90 resin microspheres as the best intervention from all compared treatments (point estimates of effect size), followed by TFD/TPI, REG, and BSC in the fixed effect model. This is confirmed by surface under the cumulative ranking curve (SUCRA) analysis (which indicates the probability of a treatment being ranked first among comparators in the NMA) (Table 4). Considering the results of both direct and indirect comparisons, any of the active comparators may be considered superior to BSC. Current evidence is unable to demonstrate statistically significant differences between SIRT, TFD/TPI, and REG in terms of OS.

A scenario analysis was performed to exclude studies assessing SIRT plus chemotherapy, which entailed exclusion of the Hendlisz et al. study [33]. In this analysis, the HR for OS, comparing SIRT versus BSC, yielded a HR of 0.37 (95% CrI 0.18, 0.75), indicating a notably more favourable outcome for SIRT using Y-90 resin microspheres.

Eight studies provided data for PFS analyses. Four studies compared REG with BSC [4, 31, 35, 36], three studies compared TFD/TPI with BSC [30, 37, 38], and one study compared TAS with REG [26]. Information regarding SIRT was not sufficient for inclusion in the PFS analysis. HR and SUCRA estimates suggest that TFD/TPI (HR 0.44 [0.33, 0.63]; SUCRA 0.56) has the highest probability of being the best intervention in terms of PFS followed by REG (HR 0.46 [0.33, 0.63]; SUCRA 0.44) in the fixed effects model. The graphic representation of the NMA for OS and PFS are shown in Supplementary Materials Fig. S1. The results of individual studies are shown in Supplementary Materials Table S12.

None of the studies added to the present review reported on SIRT. Therefore, the percentage of patients experiencing grade 3 or higher adverse events remained the same for patients receiving SIRT as those reported by Walter et al. [11]: hand-food skin reaction (4.8% of the sample) [33] and radioembolization-induced liver disease (REILD) (10.3%), which was symptomatically managed [32]. It should be noted that in the study reporting hand-food skin reaction, SIRT was used in combination with 5-FU, which could explain these AEs. Conversely, higher levels of non-haematological AEs such as severe hand-foot skin reaction fatigue, diarrhoea, hypertension, and rash or desquamation are reported with REG compared to TFD/TPI and SIRT using Y-90 resin microspheres [26, 29]. While TFD/TPI was associated with a higher number of haematological events such as febrile neutropoenia (4%) [37] and with one death attributed to septic shock [37], grade 3 or higher drug-related AEs occurred in 57% patients receiving REG treatment compared to 37% patients receiving TFD/TPI treatment in the European Vitale et al. study [29] (Supplementary Materials Table S13), indicating a favourable safety profile for SIRT using Y-90 resin microspheres.