First-Line Immune Checkpoint Inhibitor-Based Sequential Therapies for Advanced Hepatocellular Carcinoma: Rationale for Future Trials

Hepatocellular carcinoma (HCC) is the fourth-leading cause of cancer deaths worldwide [1]. Curative treatment options, such as liver transplantation, resection or ablation, exist for Barcelona Clinic Liver Cancer (BCLC) stage 0/A (early-stage) HCC, while patients with locally advanced disease may be candidates for liver-directed therapies, including transarterial chemoembolization and transarterial radioembolization [2]. Unfortunately, more than half of all HCC cases are diagnosed at a stage with no potentially curative treatment options. Since 2008, the oral tyrosine-kinase inhibitor (TKI) Sorafenib (SORA), has been recommended as the standard first-line systemic therapy for patients with BCLC stage C (advanced) HCC and those with BCLC stage B (intermediate) HCC who are unfit for, or fail to respond to, loco-regional therapies [3].

Recently, several newer systemic therapy options have shown efficacy in the first- and second-line settings. In the first-line setting, combination therapy with atezolizumab (ATEZO) plus bevacizumab (BEVA) has exhibited an impressive improvement in overall survival (OS) and progression-free survival (PFS) compared with SORA and it now represents the standard of care [4]. In the second-line setting, multiple TKIs have shown survival benefit when compared with placebo [5-7], although a definitive standard of care has not yet been defined [8].

Taken individually, these randomized controlled trials (RCTs) offer a number of effective choices for first- and second-line systemic treatments. Ideally, they should be combined in a rational sequence to offer the best net health benefit for patients with advanced HCC. However, RCTs were designed to maximize the effectiveness of each new drug under evaluation, and the best sequential systemic treatment remains debated [8-10]. Unfortunately, sequential trials in oncology are difficult to perform and real-world data of sequential therapies for advanced HCC are lacking to date. Moreover, all approved second-line therapies in HCC have been assessed in patients who received SORA as first-line treatment, and data on second-line treatment after first-line ATEZO + BEVA are lacking. Nevertheless, the American Society of Clinical Oncology (ASCO) and the European Society for Medical Oncology (ESMO) guideline recommendations recently suggested the use of SORA or lenvatinib (LENVA) as second-line treatment options after ATEZO + BEVA failure [9, 10]. Given the lack of direct comparative effectiveness data for sequential therapy options, we developed a simulation model based on published phase 3 RCTs to identify the optimal strategy for sequencing systemic agents after first-line immune checkpoint inhibitor (ICI)-based combination treatments.

Data from phase 3 RCTs of systemic agents with proven survival benefit were extracted. For ATEZO + BEVA, data were extracted from the IMbrave 150 trial [4]. Data were extracted from 8 phase III RCTs [3, 4, 11-16] for SORA and pooled by a random-effects model. Data were extracted from REFLECT [16], RESORCE [5], and CELESTIAL [6] trials for LENVA, Regorafenib (REGO), and Cabozantinib (CABO), respectively. For Ramucirumab (RAMU), data were extracted from a pooled meta-analysis of independent patient data [17] from REACH [18] and REACH-2 [7] trials. Characteristics of the included RCTs are showed in online supplementary Tables 1 and 2 (for all online suppl. material, see www.karger.com/doi/10.1159/000520278).

To examine the extent to which differences in OS with second-line TKI therapy (including SORA and LENVA) could be explained by differences in study- and patient-level covariates, the following explanatory variables were included in a meta-regression model: year of study publication (only for SORA trials), type of drug, sex, ethnicity, Eastern Cooperative Oncology Group (ECOG) Performance Status, etiology of liver disease, Child-Pugh class, Albumin-Bilirubin (ALBI) grade, BCLC stage, alpha-fetoprotein (AFP), macrovascular invasion and extrahepatic spread. Data on ALBI grade were extracted from post hoc exploratory analyses when available [19-21]. Other variables of interest, including degree of liver involvement and degree of portal vein invasion were not uniformly reported across trials, so could not be included in meta-regression analysis. Median OS times were used as the effect size, weighted by standard errors. In order to reduce the risk of ecological bias, only univariate meta-regression was performed. Restricted maximum likelihood was used for the model estimation (R package nlme). Data used in the meta-regression model are reported in online supplementary Table 2.

A semi-Markov model was developed to analyze the effectiveness of different sequential systemic treatments for advanced HCC over a lifetime horizon (Fig. 1). The model allowed the transition hazard to increase over time and it was performed by estimating the scale and shape parameter of a Weibull distribution. Natural mortality rate at age of population was included in the model for a 65-year-old hypothetical male patient. Given the very short life expectancy, we did not discount risks and benefits. The sequential semi-Markov model was built by using an ad hoc routine written with R statistical software. We simulated the clinical course of a hypothetical patient cohort with advanced HCC and similar characteristics as those enrolled in included phase 3 RCTs (online supplementary Table 1).

Data describing OS and PFS treatment benefits were derived from Kaplan-Meier curves from the above-mentioned trials. Specifically, we extracted data from the PFS curve of the IMbrave 150 trial [4] for first-line ATEZO + BEVA. PFS was assumed as the main endpoint for first-line ATEZO + BEVA, because progression is the event leading to treatment discontinuation and transition to second-line therapy [22]. OS curves of patients treated with second-line therapies were extracted from all 8 phase III RCTs [3, 4, 11-16] for SORA and from REFLECT [16], RESORCE [5], and CELESTIAL [6], for LENVA, REGO, and CABO, respectively. For RAMU, data were extracted from a pooled meta-analysis of independent patient data [17] including patients with AFP higher than 400 ng/mL from REACH and REACH-2 trials [7, 18].

Individual patient survival data were reconstructed by using an algorithm proposed by Guyot et al. [23]. This algorithm provides a list of patients with predicted survival times and a predicted event of interest (i.e., alive or dead; progression or no progression) by using digitalized data on survival probabilities, time, and total number of patients and events. Engauge Digitizer software Version 12 was used to extract data from the curves. Each reconstructed survival curve was inspected for accuracy and was compared with originally published curves. For safety measures, we extracted the number of grade ≥3 adverse events (severe adverse events, SAEs) from each included trial.

The model simulated transitions among the following 3 health states, with a cycle length of 1 month: (1) advanced HCC eligible for first-line systemic therapy; (2) first progression; and (3) death (Fig. 1). The health states were mutually exclusive (i.e., a patient could only experience a single health state at any given time). After first-line therapy, patients could experience a response and continue first-line ICI-based therapy, experience progression, and switch to second-line treatment, or death. Patients on second-line therapy could experience treatment response or death. Therefore, patients remained on second-line therapy if they experienced second progression.

Survival estimates for sequential settings considered the proportion of patients who did not receive second-line therapy due to death. Transitions between health states were based on calculated transition probabilities from PFS and OS data extracted from published studies. Simulation of 2000 pseudo-random patients was performed to obtain overall simulated survival times and median times throughout the 3 states of the disease.

Life-year gained (LYG) of the sequential therapies was the main health outcome. We calculated median OS of sequential therapies from the Markov model. To provide long-term survival information, a milestone survival analysis at 24 months was performed. Accordingly, the number needed to treat (NNT) was calculated for each sequential strategy.

The rate of SAEs was calculated for each sequential treatment, taking into account transition probabilities from first- to second-line treatments, by using a weighted mean of the number of patients transited in each disease state (initial treatment and cancer progression). To obtain a net health benefit measure [24], clinical benefit and safety were combined to calculate an innovative measure, that is the incremental safety-effectiveness ratio (ISER). The latter was defined as the difference in the rate of SAEs between 2 sequential treatments, divided by their difference in effectiveness, measured in LYG. This unit of measure expresses the incremental percentage of SAEs for each LYG.

Two-way sensitivity analysis for ISER was performed by varying the numerator (percentage difference in SAEs) and the denominator (LYG). The variation range for the denominator was chosen to be as large as the maximal value observed, while a ±5% variation was chosen for the numerator. Analogous to the willingness-to-pay threshold of a cost-effectiveness analysis, a willingness-to-risk threshold value must be established for ISER. Different willingness-to-risk thresholds for ISER, and their visual effect in terms of estimated areas, across the potential range of SAEs and LYG, have been calculated. All the analyses were performed using R (R core team, 2020).

Meta-regression analysis showed that SORA OS increased in RCTs published after 2018, so pooled OS of SORA was stratified before and after 2018 (see online suppl. Table 3). SORA after 2018 and LENVA were the only drugs significantly associated with OS improvement compared to SORA before 2018. Pooled OS from the 2 SORA control arms of IMBrave 150 [4] and REFLECT [16] RCTs were included in the Markov model as post-2018 trials. Among patient-level covariates, ALBI grade 2, BCLC C, and AFP >400 ng/mL were significantly associated with worse OS.

Relevant outcomes and model parameters from the RCTs are described in Table 1. The base-case efficacy and safety of the treatment sequences after ATEZO + BEVA, ordered according to 24-month OS, are reported in Table 2. First-line ATEZO + BEVA followed by second-line LENVA was the most effective treatment with median OS of 24 months, LYG of 0.50 year, 24-month OS of 49.3% and NNT of 5. First-line ATEZO + BEVA followed by second-line SORA was the second most effective treatment with median OS of 23 months, LYG of 0.42 year, 24-month OS of 46.6%, and NNT of 6. Figure 2 shows the simulated survival curves of ATEZO + BEVA followed by LENVA or SORA.

The safety profile of treatment sequences is reported in Table 2. First-line ATEZO + BEVA followed by second-line LENVA was associated with 67.8% SAEs, while first-line ATEZO + BEVA followed by SORA was associated with 63.0% SAEs, resulting the safest sequence. First-line ATEZO + BEVA followed by second-line LENVA-dominated sequential strategies, including REGO or CABO for both effectiveness and safety.

A two-way sensitivity analysis, varying SAE and LYG values, was performed to evaluate which therapy would be preferred at different willingness-to-risk thresholds. In Figure 3, at a willingness-to-risk threshold of 10% of SAEs for LYG, first-line ATEZO + BEVA followed by second-line SORA was favored in 72% of cases, while at a threshold of 30% of SAEs for LYG, first-line ATEZO + BEVA followed by second-line LENVA was favored in 69% of cases.

The best sequential systemic treatment for patients with advanced HCC after first-line ATEZO + BEVA failure still remains debated. Recently, ASCO, ESMO, and the European Association for the Study of Liver societies suggested the use of any TKI (including SORA and LENVA) as second-line therapies after ATEZO + BEVA [9, 10, 25]. Our analysis provided evidence that SORA and LENVA were the 2 TKIs significantly associated with OS improvement. In this line, our sequential model suggests that first-line ATEZO-BEVA followed by second-line LENVA or SORA could be optimal sequential systemic strategies in terms of efficacy, with a median OS of about 2 years, as well as safety. To the best of our knowledge, this is the first model providing comparisons among different second-line TKIs after first-line ICI-based treatment. Waiting for the validation of the key assumptions of our model by RCTs and large prospective real-world data, our simulation model provides further evidence supporting ASCO, ESMO, and the European Association for the Study of Liver recommendations, suggesting that the sequences with first-line ATEZO + BEVA followed by second-line TKIs (LENVA or SORA) may represent the optimal sequences. In this line, we are awaiting for the upcoming results from the phase III trial of the ICI-TKI-based combination LENVA plus Pembrolizumab [26]. Regarding the safety profile, our study showed that the use of second-line SORA after ICI-based first-line was the best sequence in terms of tolerability. It is important to note that the results of our meta-regression, showing that the benefit of SORA significantly increased over time, are plausible and supported by available real-world and clinical trial data [27-29]. The improvement in the management of SORA side effects, tailoring SORA dosage to optimize risk/benefit ratio, and finally the availability of second-line treatment options could explain these results.

Sequential trials in oncology are difficult to perform and they are practically unfeasible in most settings, such as advanced HCC. Furthermore, large prospective real-world data of sequential therapies in advanced HCC are time-consuming and lacking to date. Our results address an area of urgent clinical need, in which direct comparative effectiveness data are lacking. Our simulation model could be a useful tool to predict the outcome of different systemic treatment sequences. In this line, simulation models starting from locoregional treatments are needed to provide a more accurate assessment of treatment benefit across the patient-journey.

Simulation models are subject to several limitations, mainly due to lack of individual data and absence of validation on key assumptions. For example, we did not have data regarding degree of liver involvement or portal vein invasion for meta-regression models. Therefore, our sequential model should be considered as a hypothesis generating tool, and real-world data should be constantly generated to confirm all assumptions. In this line, Yoo et al. [30] demonstrated the second-line treatment with LENVA or SORA had comparable efficacy and manageable toxicities in patients with advanced HCC after ATEZO + BEVA failure. We believe that simulation models could provide an anticipated evaluation useful to design and power future trials. For instance, our forecasts on the 2 optimal sequences ATEZO + BEVA followed by LENVA or SORA represent the control group of the ongoing trial (IMBrave251, NCT04770896), comparing ATEZO + LENVA versus LENVA alone or ATEZO + SORA versus SORA alone in ATEZO + BEVA failure patients [31]. Moreover, it is important to consider that previously published estimates of sequential strategies are biased by “per-protocol analyses,” and they did not consider patients who died before second-line treatment [32, 33], with a bias of potential overestimate of the efficacy of the proposed sequences. Finally, a strength of our model was the use of PFS as the main endpoint for first-line treatment, since it is likely that, in real-life, treatment sequencing will be considered upon tumor progression [22, 34, 35].

Assessment of the benefit of ICIs raises several issues considering the unconventional pattern of radiological response. Novel tools assessing radiological response (i.e., immuneRECIST or radiomic tools) or innovative measures of net health benefit assessment may be able to capture the long-term benefit observed with ICIs [36]. According to ASCO statements [24], net health benefit is defined as the balance between clinical benefit and SAEs, and it is used to assess the difference between 2 different therapeutic strategies [37]. In our model, we used an innovative measure combining efficacy and safety, called ISER (incremental safety-effectiveness ratio). Given the range of uncertainty on the net health benefit assessed by ISER between ATEZO-BEVA followed by LENVA or SORA, the best treatment strategy should be carefully shared with the patients and agreed with the policy makers, taking also into account cost-effectiveness ratio. In this line, the limited survival benefit observed in RCTs, particularly in the second-line setting, underlines the importance of including both quality of life and the patient-reported outcomes. When we evaluated the net health benefit according to ISER, we observed that ATEZO + BEVA followed by LENVA was clearly favored only in patients with high willingness-to-risk threshold (e.g., fit patients with fighter attitude who dismiss the impact of SAEs), while ATEZO + BEVA followed by SORA was favored in patients with low willingness-to-risk threshold (e.g., frail patients with fatalist attitude in whom a higher risk of SAEs is relevant). It is to note that TKIs can have continuous low grade (grade 2) toxicity that could not be entirely captured by ISER.

Like all the simulation models, in the present study there are several limitations. First, there is a lack of published data concerning second-line treatments after ATEZO + BEVA. All available second-line therapies were approved in trials including SORA-experienced patients, while no second-line therapies were evaluated after ATEZO + BEVA. As a proxy measure of treatment failure, about 20% of patients in the IMBrave 150 trial [4] received subsequent TKI treatment after ATEZO + BEVA, although the reasons for the transition to second-line therapy were unknown. Second, we were unable to account for some differences in patient inclusion and exclusion criteria across trials. Notably, the REFLECT trial excluded patients with >50% liver involvement or main portal vein invasion. The inclusion of a more favorable population may bias results to favor LENVA [16]. Furthermore, some patients with progression after ATEZO + BEVA may not be candidates for second-line LENVA if they present the previously mentioned exclusion criteria. Moreover, REFLECT trial also selected a population with a lower risk of developing cardiovascular events during TKI treatments [16]. Third, we incorporated data from RCTs of both first- and second-line TKIs to simulate the benefit of sequences after first-line ATEZO + BEVA. Therefore, the estimate of SORA and LENVA OS in RCTs may be overestimated compared to REGO, CABO, and RAMU OS, given the use of different subsequent therapies in some patients after SORA or LENVA discontinuation. Fourth, the sequences including REGO as second-line therapy could have overestimated safety because RESORCE trial included only SORA-tolerant patients and SAEs may be higher in an unselected population [5]. Fifth, we were unable to calculate quality-adjusted life-years, since utilities were not available in the majority of the RCTs examined. Accordingly, we assessed LYG, and not quality-adjusted life-years, as the primary measure of efficacy. Sixth, we lacked data on potentially relevant effect modifiers such as microscopic vascular invasion, histological grading, gene profiling [38]. However, our meta-regression showed that BCLC stage C, ALBI grade 2, and AFP higher than 400 ng/mL were significantly associated with worse OS, confirming their prognostic usefulness. Seventh, the number of grade 3 and 4 SAEs was not separately reported in RCTs, hampering the evaluation of the degree and severity of SAEs in the calculation of ISER. Finally, due to the lack of individual data, our model cannot consider dropouts from first to second-line because of AEs or liver decompensation, which are well-demonstrated drivers of death in patients with HCC [39-42]. Therefore, time to liver decompensation [42] and time to treatment discontinuation due to AEs should be collected and reported in future trials. In conclusion, waiting for the results from RCTs and prospective real-world data evaluating first-line ICI-based combination followed by second-line mono- or combination TKI-based regimens [31], ATEZO + BEVA followed by LENVA or SORA could be considered optimal sequential options for advanced HCC.

The paper is exempt from Ethical Committee approval according to decision of Ethic Committee of Palermo and written consent was not required as only data extracted from published randomized controlled trials were used.

Giuseppe Cabibbo participated in advisory board for Bayer, Eisai, and Ipsen. Maria Reig is a consultant at Bayer-Shering Pharma, BMS, Roche, Ipsen, AstraZeneca, Lilly, and BTG; attended paid conferences at Bayer-Shering Pharma, BMS, Gilead, and Lilly; received research grants from Bayer-Shering Pharma and Ipsen. Ciro Celsa received speaker fees from Eisai. Amit Singal served on advisory boards or as consultant for Genentech, Bayer, Eisai, Exelixis, Bristol Myers Squibb, and AstraZeneca. Jordi Bruix is a consultant at AbbVie, Adaptimmune, Arqule, Astra-Medimmune, Basilea, Bayer-Shering Pharma, Bio-Alliance, BMS, BTG-Biocompatibles, Eisai, Gilead, Incyte, Ipsen, Kowa, Lilly, MSD, Nerviano, Novartis, Polaris, Quirem, Roche, Sirtex, Sanofi, and Terumo; received research grants from Bayer and BTG; received educational grants from Bayer and BTG; attended paid conferences at Bayer, BTG, Astra-Zeneca, and Ipsen; gave paid talks at Bayer-Shering Pharma, BTG-Biocompatibles, Eisai, Terumo, Sirtex, and Ipsen. Calogero Cammà participated in the advisory board for Bayer, MSD/Merck, Ipsen, and Eisai. The other authors have no disclosure to declare.

A.G.S. received Grant support from the National Institute of Health (NIH) R01 MD12565. J.B. received Grant support from Instituto de Salud Carlos III (PI18/00768), AECC (PI044031), and WCR (AICR) 16-0026. CIBERehd is funded by the Instituto de Salud Carlos III.

All the authors take full responsibility for the study design, data analysis and interpretation, and preparation of the manuscript. All authors were involved in planning the analysis and drafting the manuscript. All authors approved the final draft of the manuscript.

All data generated or analyzed during this study are included in this article and/or its online supplementary material files. Further inquiries can be directed to the corresponding author.

Giuseppe Cabibbo and Maria Reig equally contributed to this work.