We registered the protocol for this systematic review and meta-analysis on PROSPERO as recommended (https://www.crd.york.ac.uk/PROSPERO, ID: CRD42022321949).

A systematic literature search was performed using PubMed, Embase, the Cochrane Central Register of Controlled Trials (CENTRAL), and China National Knowledge Infrastructure (CNKI) using the Medical Subject Headings (MeSH) terms. The search covered the period from the inception date of each database until December 21, 2022. There was no restriction on publication status or language, and all the included non-English studies were translated into English. The keywords were as follows: (atezolizumab or bevacizumab or sorafenib or oxaliplatin or lenvatinib or pembrolizumab or nivolumab or camrelizumab or apatinib or tyrosine kinase inhibitor* or TKI or PD-1 or PD-L1 or immune checkpoint inhibitor* or ICI) and (hepatocellular cancer or liver cancer or hepatocellular carcinoma or HCC or liver neoplasms*) and (advanced or unresectable or inoperable).

The inclusion criteria were as follows: 1) studies that evaluated the effects of systemic combination therapies on aHCC, with or without controls; 2) studies that included research subjects with advanced/unresectable HCC; 3) studies that incorporated at least one of available endpoints (overall survival, progression-free survival, or objective response rate). The exclusion criteria were as follows: 1) studies that enrolled patients with cancers that metastasized to the liver, 2) studies with the combination of systemic and topical treatments, and 3) studies that lacked necessary information for data extraction.

Patients with aHCC were eligible, defined as advanced metastatic or unresectable hepatocellular carcinoma (i.e., patients with characteristics such as multifocal and/or infiltrative disease within the liver, vascular invasion, or extrahepatic spread), with the diagnosis confirmed by histologic or cytologic analysis or clinical features. Systemic combination therapies are defined as those systemic combination regimens for aHCC, recommended in the NCCN and ASCO guidelines (8, 9), or the actual combinations used clinically, including “atezolizumab+bevacizumab”, “bevacizumab+erlotinib”, “nivolumab+ipilimumab”, “capecitabine+oxaliplatin”, and “sorafenib+GEMOX”. Comparators were the systemic monotherapies including sorafenib, gemcitabine, or oxaliplatin. There was no restriction on the types of control treatment. The primary endpoints were OS at 6 months or longer, PFS at 6 months or longer, and ORR. OS was defined as the interval between the date of random assignment and the date of death from any cause; PFS was defined as the interval between random assignment and progression or death from any cause; ORR was defined as the percentage of patients who had a confirmed complete or partial response. Those endpoints were assessed according to Response Evaluation Criteria in Solid Tumors (RECIST) 1.1. The second endpoints were treatment-related adverse events (TrAEs) and ≥3 Grade TrAEs, according to the National Cancer Institute (NCI) Common Terminology Criteria for Adverse Events (CTCAE) (version 3.0).

Our systematic searching was conducted according to the search terms we set before and followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines (PRISMA 2020, http://www.prisma-statement.org). The headings and abstracts identified in those databases were reviewed and cross-checked by four investigators (LM, HM, RX, and LD) for the identification of studies that fulfill the eligible criteria. Moreover, reference lists of eligible studies, conference abstracts, and systematic reviews were reviewed to acquire relevant papers as well. For studies using the same data source, the most recent study or the study with the largest sample size was included. Some complicated papers were judged by panel discussion and further arbitrated by the third experienced reviewer (LP).

Prespecified data were independently extracted and double-checked by reviewers (LM, HM, RX, and LD). A standardized data extraction form was designed to manage the necessary items. The following information was extracted from all included publications (papers, abstracts, and reported data of registered trials): title, author, study design, country, sample size, experimental arms, control arms, other demographic characteristics, description of outcomes and adverse events, and their definitions. In particular, a study involving multiple systemic combination therapies was divided into a series of comparisons, each of which contained a group of combination and control regimens.

The characteristics of the included studies were synthesized and presented in a tabular form. Categorical variables were presented as counts (%), and continuous variables were described using means or medians. The HR, relative risk (RR), and median with corresponding 95% CI were pooled by the Mantel–Haenszel fixed-effects model if no evidence of significant heterogeneity existed; otherwise, the DerSimonian–Laird random-effects model was applied. The heterogeneity was assessed by I ² statistics, classified as low (I ² < 25), moderate (25 ≤ I ² < 75), and high heterogeneity (I ² ≥ 75). Generally, I ² statistic >50% was considered significant heterogeneity. The subgroup analysis was conducted by types and specific therapies of systemic combination therapy to detect the effects and heterogeneity of the treatments on aHCC across different combinations. The sensitivity analysis was performed by study design. The publication bias of the included studies was evaluated qualitatively by funnel plot and quantitatively by Egger’s test. These statistical analyses and plots were performed using “meta” and “metamedian” packages from R software, version 3.6.2 (R Foundation for Statistical Computing, Canberra, Austria).

The Cochrane Collaboration Tools were applied to assess the risk of bias in the included studies. In the tools, the Newcastle–Ottawa Scale (NOS) and version 2 of the Cochrane tool (RoB 2) were used to assess the risk of bias in cohort studies and randomized trials, respectively, as recommended by the Agency for Healthcare Research (19, 20). An Excel tool developed by the agency was applied to implement RoB 2 (https://www.riskofbias.info/welcome/rob-2-0-tool/current-version-of-rob-2). A cohort study was awarded a maximum of nine stars in three sections, and study quality was judged as follows: high risk of bias = 0–3; moderate risk of bias = 4–6; low risk of bias ≥ 7. RoB 2 contains five domains, and each of them was assessed as low risk of bias, some concerns (moderate risk of bias), or a high risk of bias. The judgment principles of the overall bias were as follows: low risk of bias if all domains were labeled as low risk, and high risk of bias if any of the domains were labeled as high risk; otherwise, the assessment result was some concerns. Two reviewers (LM and HM) assessed the risk of bias for each study independently.

In total, 7,249 potential records (papers, abstracts, and registered trials) were screened from databases and other relevant sources. Ultimately, 88 eligible studies (13, 15, 21–107) fulfilling the criteria were included after removing duplicates and reviewing the papers in detail ( Figure 1 ). Of those studies, Dhooge 2012 was conducted in both Child-Pugh A and B cohorts; Hitron 2014 and Jiang 2019 contained two kinds of combination regimens, and IMbrave150 included global and China cohorts. Thus, 92 records were incorporated into the analyses. The baseline characteristics of the studies are described and summarized in Table 1 . In total, 9,748 patients with aHCC from all around the world were included in our systematic review. There were 47 double-arm studies with 7,431 patients and 45 single-arm studies with 2,317 patients. The mean or median age of the total population was 60 years. Most studies were conducted in China (48, 52.2%), and the USA (18, 19.6%). The systemic regimens were the pairwise combinations of targeted therapy, ICI therapy, chemotherapy, and other therapies. Of those, the number and proportions of chemotherapy plus chemotherapy, targeted therapy plus chemotherapy, targeted therapy plus ICI therapy, and targeted plus targeted therapies were 25 (27.2%), 23 (25.0%), 20 (21.7%), and 16 (17.4%), respectively. Subsequently, four specific therapies of those combinations with study numbers greater than 3 were analyzed to probe heterogeneity: Atezo plus Beva, gemcitabine (Gemc) plus oxaliplatin (Oxal), erlotinib (Erlo) plus Beva, and sorafenib (Sora) plus gemcitabine and oxaliplatin (GEMOX). The majority of monotherapies was sorafenib (18, 38.3%), followed by gemcitabine (7, 14.9%). All studies included were trials (73, 79.3%) and cohorts (19, 20.7%).

The safety profile of systematic combination therapy was also evaluated in this meta-analysis, including the overall TrAEs in 31 two-arm studies and ≥3 Grade TrAEs in 20 two-arm studies. The incidence rate of TrAEs among those combination interventions was comparable (RR: 1.00 [95% CI: 0.98–1.02]; I ^{2 =} 73%; Figure S6 ). For ≥3 Grade TrAEs, the pooled result estimated by the fixed-effects model indicated that the combinations had an increased RR of 1.25 [1.15–1.36], compared to the monotherapy ( Figure S7 ). Moderate heterogeneity was detected across those interventions (I ^{2 =} 25%). In the subgroup analysis, the incidence rates of ≥3 Grade TrAEs in the chemotherapy plus chemotherapy, targeted therapy plus ICI therapy, and targeted plus targeted were significantly higher, compared to the monotherapy, with RR values of 1.19 [95% CI: 1.01–1.39], 1.26 [0.90–1.75], and 1.38 [1.16–1.64], respectively. However, it was not significant in targeted therapy plus chemotherapy (1.08 [0.93–1.25]).

For quality assessment, the risk of bias in most studies was high or moderate, which should be attributed to nearly half of the studies with a single arm. However, the quality of double-arm studies was generally acceptable, of which the proportion with low or moderate risk of bias was 89.4%. Funnel plots for the effects of systematic combination therapies on OS, PFS, and ORR were asymmetrical. Moreover, the results of Egger’s test indicated that publication bias was detected (OS, p = 0.084; PFS, p = 0.04; ORR, p = 0.002; ≥3 Grade TrAEs, p = 0.092; Figure S8 ).