Types of studies

We included only RCTs, without language or date restriction, published in any database, indexed or not. We excluded quasi-experimental studies, observational studies, and reviews.

Types of participants

We considered studies including adults with primary or recurrent, advanced (stages IIIb or IV) or metastatic gallbladder or bile tract cancer. We excluded neuroendocrine, stromal or lymphatic neoplasms.

Types of interventions

We considered eligible interventions using any ACD, including chemotherapy (either monotherapy or in combination), biological/targeted therapy or immunotherapy, whether individual or combined, with or without supportive care. We excluded trials considering only surgery or radiotherapy as intervention.

We considered BSC as the comparator, broadly defined as any supportive treatment administered with the purpose of symptomatic or palliative control. ^{15 , 16} We considered trials that used BSC, supportive care, placebo, or that did not explicitly use ACDs as a comparator. We excluded comparisons that consisted of other treatments with non-palliative intention (e.g. radiotherapy or surgery with curative intention).

Types of outcomes

We considered overall survival (OS) (dichotomous, continuous or time-to-event outcome); quality of life (measured with validated scales); progression-free survival (PFS) (as a dichotomous and a time-to-event outcome); functional status (measured with validated scales); and toxicity (measured as moderate or severe adverse events, according to standardised classification) as critical outcomes. We also considered the following as important outcomes: Symptoms related to the disease (measured with validated scales); admissions to hospital or long-term centres, or number of emergency consultations; and quality of end-of-life care, defined as a composite outcome including any admission to hospital at the end of life (i.e. the last 30 days of life), palliative care provided during the last year, and place of death.

We conducted a sensitive search strategy for the first two stages of the overall evidence synthesis project ²² in MEDLINE (access via PubMed), EMBASE (access via OVID), the Cochrane Database of Systematic Reviews, CENTRAL, and Epistemonikos until December 2019. Table S1 (Additional file 1) provides the search strategy for MEDLINE/PubMed. Using the same search strings, but only focusing on advanced BTCs, we ran an update in MEDLINE/PubMed and CENTRAL. Table S2 (Additional file 1) details this update for MEDLINE/PubMed. We conducted complementary searches in clinicaltrials.gov, and also asked experts in the field for relevant studies. Using a subset of identified RCTs, we conducted a backward and forward citation search using citationchaser. ^{30 – 32}

Two reviewers performed an independent title and abstract screening of the results obtained from the search. A third reviewer solved any disagreement. Afterwards, two reviewers conducted the full-text screening, also with a third author solving any disagreement. We used the Covidence software for the selection process ³³ (although other open access alternative softwares are available, see “Availability of data and materials” section).

Two reviewers extracted data independently from the included studies, using a previously piloted data extraction sheet. We solved disagreements through discussion. We extracted the following data for each included RCT: general characteristics of the study (country, specific study design, setting, objective, funding, conflict of interests, total patients included and randomised); type of experimental intervention (chemotherapy, targeted/biological therapy and/or immunotherapy); type of control intervention (BSC, placebo, other or non-specified); outcomes assessed; and results for each outcome. Two reviewers independently assessed all included studies using the Cochrane ‘risk of bias’ tool. ³⁴ Disagreements were resolved by consensus or by a third reviewer.

For dichotomous outcomes, we estimated treatment effect with risk ratio (RR) and its respective 95% confidence interval. For continuous outcomes, we calculated mean difference (MD) if the outcomes were reported with the same scale across the included studies, or standardised mean difference (SMD) if different scales were used. For time-to-event outcomes, we used hazard ratios (HR). Whenever a study included multiple arms, we only considered comparisons relevant for our reviews. We assessed heterogeneity both visually and using I ² through Software Review Manager 5.4.1, ³⁵ and considered it for purposes of estimation of certainty of evidence.

If there were less than two studies reporting for a specific outcome, or in case they were considerably heterogeneous, we did not perform meta-analysis and only reported results descriptively. ³⁶ If studies were reasonably homogeneous (both clinically and methodologically), we performed meta-analysis, using a random effects model. For each type of intervention in the experimental arm, we grouped all the interventions in the control group (i.e., placebo, supportive care, best supportive care, usual treatments and others) as one unique comparator. We planned to assess publication bias by visually analysing funnel plots only if 10 or more studies provided data for a specific outcome. We conducted subgroup analysis according to line ³⁷ : specific types of chemotherapy, specific types of immunotherapy, specific types of biological/targeted therapy.

We assessed the certainty of the evidence for each primary outcome according to GRADE guidance. ³⁸ We created Summary of Findings (SoF) tables, classifying the certainty of evidence for each outcome as high, moderate, low, or very low. We also report the main findings of the SoF table in plain language, according to their specific assessment of the certainty of evidence. ³⁹

The included studies were conducted in Sweden, ⁷ Japan, ⁴⁰ India, ^{6 , 42} Korea, ⁴¹ Belgium, ⁴³ and UK. ⁴⁵ One trial was conducted in various countries, including France, Italy, Spain, UK, US and South Korea. ⁴⁴ RCTs included patients with advanced cholangiocarcinoma, ^{7 , 41 , 44} gallbladder cancer, ^{6 , 42} or both. ^{40 , 43 , 45} One study included exclusively patients with IDH-1 mutation. ⁴⁴ Two trials also included patients with pancreatic adenocarcinoma. ^{7 , 40} Four RCTs provided the intervention as first line therapy, ^{7 , 40 – 42} while three studies specified that patients were previously treated. ^{43 – 45} One RCT admitted patients with previous treatment if it was completed at least six months before enrolment. ⁶ None of the included studies assessed immunotherapy. Table 1 provides the general characteristics of the included studies

One study had an overall low risk of bias. ⁴³ Most of the included studies had an unclear risk of selection bias due to unclear reports about the random sequence generation ^{6 , 7 , 40 , 44} or allocation concealment. ^{6 , 7 , 40 , 41} Several RCTs had high risk of performance bias, ^{6 , 7 , 40 – 42 , 45} and high ^{40 , 42 , 45} or unclear ^{6 , 7} risk of detection bias. None of the studies was considered to be at high risk of attrition bias, with only two being unclear in this domain. ^{40 , 42} Reporting bias was considered at high risk for three studies, ^{40 , 41 , 44} and unclear for two. ^{7 , 45} Table 2 provides the summary of the risk of bias assessment of the included studies.

Three RCTs provided data for this comparison. ^{42 – 44} One RCT compared erlotinib as first line therapy versus BSC in patients with advanced gallbladder cancer. ⁴² The other two used biological/targeted therapies in previously treated patients, comparing its effects to placebo. One used regorafenib in patients with advanced cholangiocarcinoma or gallbladder cancer, ⁴³ and the other assessed the effects of ivosidenib in patients with IDH-1 mutant cholangiocarcinoma. ⁴⁴

Our analysis showed that the use of biological/targeted therapies in patients with advanced BTCs may result in little to no difference in OS (MD: 1.66 months higher; 95%CI, 0.65 months lower to 3.96 months higher; low certainty of evidence). This outcome was informed by the three RCTs, ^{42 – 44} with no significant heterogeneity observed between the studies. The two RCTs assessing previously treated patients ^{43 , 44} also showed that biological/targeted therapies probably result in little to no difference in PFS (MD: 1.44 months higher; 95%CI, 0.83 months higher to 2.05 months higher; moderate certainty of evidence), and suggest that they increase toxicity (RR: 1.38; 95%CI, 0.99 to 1.93; low certainty of evidence). The evidence is very uncertain about the effects of biological/targeted therapies on quality of life. There were no studies reporting data for the other prespecified outcomes. Table 3 shows the SoF and figure S1 (Additional file 2) provides the meta-analyses for the comparison of biological/targeted therapies versus BSC.

Six RCTs provided data for this comparison, ^{6 , 7 , 40 – 42 , 45} including patients with advanced gallbladder ^{6 , 40 , 42 , 45} and bile duct ^{7 , 40 , 41 , 45} cancer. Two studies also included patients with pancreatic cancer. ^{7 , 40} Four RCTs studied ACDs as first line therapy, ^{7 , 40 – 42} and one study exclusively enrolled patients previously treated. ⁴⁵ Other RCT considered ACDs as first line therapy or more. ⁶ Most of the studies described their comparator as BSC, ^{6 , 7 , 40 , 42 , 45} with one study describing only photodynamic therapy as a cointervention, with no further description of the control group. ⁴¹

Our meta-analysis of five RCTs ^{6 , 7 , 41 , 42 , 45} showed that the evidence is very uncertain about the absolute effects of chemotherapy on OS (MD: 3.28 months higher; 95%CI, 0.16 months higher to 6.39 months higher; very low certainty of evidence). The findings derived from a meta-analysis of two RCTs ^{6 , 41} providing data about the absolute effects on PFS was also very uncertain (MD: 4.27 months higher; 95%CI, 0.25 months higher to 8.78 months higher; very low certainty of evidence). Only one study provided clear data for toxicity, ⁴⁵ suggesting that chemotherapy increases adverse events compared to BSC (RR: 1.33; 95%CI, 1.03 to 1.72; low certainty of evidence). Two studies assessed quality of life using validated scales. Two studies ^{7 , 41} applied the EORTC QLQ-C30 global health status subscale, ranging from 0 to 100 points ⁷ and the Digestive Disease Questionnaire-15 (DDQ-15) ranging from 1 to 5 points ⁴¹ to assess QoL. Overall, the effect was a non-significant trend in favour of the intervention group (SMD: 0.08; 95%CI, -0.65 to 0.81; very low certainty). Another study also measured quality of life, but the full results have not been published yet. ⁴⁵ One study ⁷ reported the change in Karnofsky scale from baseline to last follow-up, showing a decrease of 5.19 points from randomisation to last follow-up in the experimental group, and a decrease of 16.06 points in the control group (MD: 10.87). Table 4 shows the SoF and figure S2 (Additional file 2) provides the meta-analyses for the comparison of chemotherapy versus BSC