Study outcomes

Our primary outcome of interest was the overall survival (OS). Secondary outcomes included (1) time to radiological progression, defined whether as progression-free survival or time to progression at any site; (2) disease control rate, defined as the sum of complete response, partial response and stable disease; (3) severe/significant adverse events, defined as the proportion of patients who developed at least one grade ≥3 adverse event according to Common Terminology Criteria for Adverse Events [16]; and (4) incidence of gastro-intestinal ulcers of any severity.

Search strategy and selection criteria

We searched Medline, Embase, Web of Science and Cochrane CENTRAL from inception up to April 14, 2020. Our search strategy was developed in collaboration with an information specialist and is available online (S1 Text). Conference abstracts obtained through our search strategy were considered. We hand searched the National Institutes of Health clinicaltrials.gov. No language restriction was applied.

We included RCTs evaluating Y90-TARE (monotherapy or in combination) in non-surgical HCC patients compared to any treatment or intervention, placebo, sham-intervention, or no intervention. RCTs reporting at least one of our outcomes of interest were considered for inclusion.

Two reviewers (SL, AB) independently reviewed all trial titles and abstracts to determine eligibility. When pertinent, the full publication was assessed independently by the same reviewers to determine final inclusion. When duplicate populations from same trials were reported, the data from the largest cohort of patients was included in the final analysis. Discrepancies or uncertainties at any point were resolved by consensus. The agreement between the two reviewers was measured using the quadratic weighted κ statistic [17].

Assessment of risk of bias

We assessed the risk of bias of the selected trials using Cochrane’s revised tool for Risk of Bias in randomized trials (RoB 2) [18]. Two reviewers (SL, AB) independently completed the assessment of bias according to the “Template for completion” (available on riskofbias.org). The RoB 2 tool assesses five domains of bias: randomization process, deviations from the intended interventions, missing outcome data, measurement of the outcome, and selection of the reported results. We classified specifically all reported outcomes from each trial as low risk, some concerns, or high risk of bias.

Quality of evidence

Two reviewers (SL, AB) independently applied the Grades of Recommendations Assessment, Development and Evaluation framework (GRADE) to assess external validity by evaluating the quality of evidence for all outcomes reported in the systematic review [19].

Data extraction and statistical analyses

The same reviewers independently (SL, AB) extracted trial identification, name, year of publication, funding sources, disclosures, country of origin, number and location of participating centers, dates of conduction, inclusion/exclusion criteria, population, details about the interventions, including monotherapy or combination, and patient characteristics from included trials.

Trial data relevant to time-to-event outcomes (OS, progression-free survival, and time to progression) was analyzed using hazard ratios (HR) and extracted on a suggested data collection form [20]. We requested unpublished data from the authors when necessary. When the HR was not available, we extrapolated approximate individual patient data from published Kaplan-Meier curves with a digitizer software (DigitizeIt, Germany, available from digitizeit.de) to obtain the corresponding HR. The algorithm used is associated with a mean absolute error of 0.017 (95% CI 0.002–1.222), so that the true HR would be between 1.475–1.525 for an extracted HR of 1.5, therefore representing no relevant systematic error [21]. For time-to-event outcomes, the (log)HR was used to summarize the results. The summary estimates were presented as a HR with its corresponding 95% CI using the inverse-variance method. For dichotomous outcomes, contingency (2x2) tables were constructed and number of events were recorded on the basis of the intervention received. The summary estimates were presented as a risk ratio (RR) with its corresponding 95% CI using the Mantel-Haenszel method. If the numerator cells contained values of zero, we added 1 to the numerator and denominator cells to calculate the RR. Statistical heterogeneity between trials was assessed according to I² with the following thresholds: (a) 0–40%: might not be important; (b) 30–60%: may represent moderate heterogeneity; (c) 50–90%: may represent substantial heterogeneity; and (d) 75–100%: considerable heterogeneity. We used the DerSimonian and Laird random-effects model [22] which accounts for within-trial and between-trial variability.

Statistical analyses were performed with Review Manager (version 5.3) and with R Software (version 4.0.0). We explored potential sources of heterogeneity with pre-specified subgroup analyses namely: (1) risk of bias; (2) type of comparator; (3) intervention in monotherapy or combination; (4) disease stage; (5) proportions of early and intermediate HCC (BCLC stages A and B); (6) proportions of advanced HCC (BCLC stage C); (7) portal vein invasion or tumor thrombosis; and (8) type of Y90 microsphere. We planned to visually assess publication bias using funnel plots, but the “rule of thumb” to include of a minimum of 10 trials was not fulfilled for any of the outcomes.

Risk of bias

For our primary outcome (OS), four trials were considered at low risk of bias (Fig 2, Table A in S1 Appendix) [23,24,29,30]. For two trials [25,27], the analytical plan was not described a priori and were therefore attributed as some concerns for risk bias. Two trials were considered at high risk of bias [26,28]. In one trial [26], seven participants were withdrawn after randomisation in the sorafenib group compared to one in the Y90-TARE which this trial also had missing outcome data. In the other trial [28], 47.2% of the participants did not receive the intervention they were allocated to or sustained major protocol deviations. A summary of the risk of bias assessments for the secondary outcomes is available in the online supplementary materials (S1 Appendix).

Fig 2

Risk of bias of the primary outcome, overall survival.

Weighted graph showing the risk of bias according to each domain. Green = low risk. Yellow = some concerns. Red = high risk.

Quality of the evidence

According to the GRADE framework, we rated the quality of evidence for overall survival (Table A in S2 Appendix). We rated the quality of evidence for the secondary outcomes as very low except for grade ≥3 adverse event as low (S2 Appendix).

Primary outcome

The OS was reported in all eight trials. The corresponding HR was reported in three trials [23,28,30], obtained through correspondence with the author in one trial [24], and extrapolated from reconstructed approximate individual patient data from published Kaplan-Meier curves in two trials [27,29]. We could not use the data from two trials in our analyses as they either solely provided the median survival [26] or the survival rates at six and 12 months [25], respectively. Y90-TARE was not associated with differences in overall survival (HR 0.99 [95% CI 0.81–1.21], six trials, I² = 77.6%, Fig 3) compared to standard of care. Our results were comparable in our subgroup analyses (Table 2).

Fig 3

Forest plot of time-to-event outcomes.

Cumulative (log)HR estimates with their 95% confidence intervals in the random-effects model for (A) overall survival, (B) progression-free survival, and (C) time to progression.

Table 2

Predefined subgroup analyses for the primary outcome.

	Number of trialsa	Population		Random-effects model		Heterogeneity
		Y90-TARE	Standard treatment	Pooled HR	95% CI	I2 (%)
Risk of bias (according to RoB 2)
Low	4[23, 24, 29, 30]	475	457	0.95	0.69–1.30	69%
Some concerns	1[27]	12	12	0.85	0.31–2.32	N/A
High	1[28]	216	208	1.01	0.81–1.25	N/A
Type of comparator
TACE	3[24, 27, 29]	68	69	0.63	0.38–1.04	19%
Sorafenib	3[23, 28, 30]	635	608	1.09	0.97–1.24	0%
Co-treatment
Systemic	1[28]	216	208	1.01	0.81–1.25	N/A
None	5[23, 24, 27, 29]	487	469	0.95	0.71–1.26	59%
Disease stage
Mostly BCLC A-B	4[23, 24, 27, 29]	250	247	0.81	0.49–1.35	65%
Mostly BCLC C	2[28, 30]	453	430	1.08	0.93–1.26	0%
Proportion of BCLC stages A and B
<33%	2[28, 30]	453	430	1.08	0.93–1.26	0%
33–66%	1[23]	182	178	1.12	0.90–1.40	N/A
>66%	3[24, 27, 29]	68	69	0.63	0.38–1.04	19%
Proportion of BCLC stage C
<33%	3[24, 27, 29]	68	69	0.63	0.38–1.04	19%
33–66%	1[23]	182	178	1.12	0.90–1.40	N/A
>66%	2[28, 30]	453	430	1.08	0.93–1.26	0%
Portal vein invasion or tumor thrombosis
Majority	1[30]	237	222	1.15	0.94–1.41	N/A
Minority	2[23, 28]	398	386	1.06	0.91–1.24	0%
None	3[24, 27, 29]	68	69	0.63	0.38–1.04	19%
Type of Y90 microsphere
Resin	4[23, 27, 28, 30]	647	620	1.09	0.96–1.23	0%
Glass	2[24, 29]	56	57	0.60	0.30–1.23	48%

Y90-TARE = Yttrium-90 transarterial radioembolization. HR = hazard ratio. CI = confidence interval. N/A = not applicable.

^aThe SIRTACE (Kolligs 2015) and YES-P (Mazzaferro 2019) trials had partially reported data and are therefore not included in the effect estimate and subgroup analyses for OS.

Secondary outcomes

Strength and limitations

Our study has several strengths. Our study population is exhaustive and, as compared to other systematic reviews, we expressed time-to-event outcomes using HRs [39]. Median survival times was the most reported measurement of the included trials in our meta-analysis, but is not an optimal pooled estimate for survival data [40]. However, underreporting of relevant information concerning survival analysis, notably the HR, was observed in retrieved trials in our systematic review [41]. We managed this limitation in two of the included trials [27,29] by reconstructing individual patient data from published Kaplan-Meier curves using a proposed algorithm [21], enabling us to present a cumulative pooled HRs of time-to-event outcomes, which represents the most up-to-date and relevant effect estimates. Also, the comprehensiveness of the search strategy renders unlikely the omission of pertinent trials.

Our systematic review also has limitations. First, despite being more exhaustive than previous work, we included a limited number of trials and most of the largest trials were designed to compare Y90-TARE to sorafenib. Moreover, neither the HR nor a Kaplan-Meier curve were provided in two small trials, precluding the extrapolation of missing data on survival and thus reducing our sample size. Most subgroup analyses were also not very robust considering the limited number of trials. For our primary outcome, two of the trials were considered at high risk of bias and two with some concerns on the risk of bias. In addition, for all secondary outcomes except adverse events, most trials were at high risk of bias or of some concerns. The quality of the evidence was thus affected by these important limitations. Also, reporting bias could not be excluded as the few numbers of included RCTs precluded the reliable assessment of the funnel plot. Finally, as combination therapies are increasingly being used in contemporary HCC treatments, our meta-analysis did not aim to evaluate optimal sequencing of therapy.