Competing risks of treatment discontinuation of atezolizumab plus bevacizumab in unresectable hepatocellular carcinoma.

Introduction

Hepatocellular carcinoma (HCC) remains a major cause of cancer-related mortality worldwide, with incidence rising in recent years. 1 Systemic therapy for unresectable HCC has advanced rapidly over the past several years, with immune checkpoint inhibitor-based combination regimens now widely incorporated into clinical practice. Representative regimens include atezolizumab (anti–programmed death-ligand 1 [PD-L1]) plus bevacizumab (anti-vascular endothelial growth factor [VEGF]) and tremelimumab (anti–cytotoxic T-lymphocyte-associated protein 4 [CTLA-4]) plus durvalumab (anti–PD-L1), which are widely recommended as the preferred first-line options in major international guidelines. [2] , [3] , [4] , [5] Dual checkpoint blockade with ipilimumab (anti–CTLA-4) plus nivolumab (anti–PD-1) has also emerged as an additional first-line option, supported by phase III evidence and recent regulatory approvals. 6

Atezolizumab plus bevacizumab (Atezo/Bev) has been established as a major first-line systemic therapy for unresectable HCC, demonstrating superior survival and response outcomes compared with sorafenib in the IMbrave150 trial. These benefits have been consistently maintained in updated analyses with extended follow-up. 7 , 8 Moreover, accumulating real-world evidence has corroborated its effectiveness in routine clinical practice, further consolidating its position as a first-line standard of care in HCC. [9] , [10] , [11] , [12] , [13] The clinical benefit of Atezo/Bev generally depends on maintaining treatment exposure; 14 , 15 therefore, avoiding premature discontinuation whenever feasible is an important practical goal. However, in routine care, treatment may be discontinued for diverse reasons, such as progression, adverse event (AE), hepatic decompensation, or response-driven strategy change, each implying a different clinical meaning and next-step strategy.

Despite these advances, most studies on Atezo/Bev have focused primarily on key outcomes such as overall survival/progression-free survival, objective response, and general safety profiles; only few studies have quantified how often each specific reason for discontinuation occurs over time. [7] , [8] , [9] , [10] , [11] , [12] , [13] In real-world practice, discontinuation may be attributed to clinically distinct pathways, including radiographic progression, treatment-related toxicities, hepatic decompensation reflecting limited hepatic reserve, and response-driven strategy changes under sustained disease control. These pathways have fundamentally different implications: discontinuation as a result of progressive disease indicates inadequate tumour control, whereas discontinuation because of AE or hepatic dysfunction reflects treatment-limiting vulnerability and constraints in tolerability, potentially independent of antitumor activity. Moreover, because these causes compete with one another, Kaplan–Meier methods can overestimate cause-specific probabilities; therefore, a competing-risk framework that directly estimates the cumulative incidence function (CIF) is preferable for characterising the frequency and predictors of each discontinuation pathway. 16

Clarifying the specific causes of treatment discontinuation and identifying patient subgroups most likely to discontinue for each reason would enable more tailored upfront decisions regarding treatment initiation, monitoring intensity, toxicity prevention, and timely transition to subsequent strategies. However, robust evidence directly addressing these questions remains limited.

Previous real-world studies have shown that impaired liver function is associated with treatment discontinuation and poor prognosis in patients treated with Atezo/Bev. In particular, hepatic decompensation has been identified as a major determinant of mortality, whereas tumour-related progression is more strongly associated with aggressive tumour biology and elevated alpha-fetoprotein (AFP) levels. [17] , [18] , [19] However, these studies have primarily focused on overall discontinuation or survival outcomes, without comprehensively evaluating the time-varying probability and competing nature of specific discontinuation causes.

Therefore, in this multicentre retrospective study conducted in a large, well-characterised real-world cohort, we aimed to (i) quantify cause-specific discontinuation probabilities for Atezo/Bev using competing-risk CIF; (ii) identify cause-specific discontinuation predictors (progression vs. toxicity); (iii) examine whether the aetiology (HBV vs. HCV vs. alcoholic liver disease (ALD) vs. metabolic dysfunction-associated steatohepatitis [MASH]) modifies the discontinuation patterns, and (iv) develop predictor-based risk stratification (AFP/albumin–bilirubin [ALBI]) that supports pretreatment planning and optimised, risk-adapted monitoring to enable early intervention for AE and hepatic decompensation.

Patients and methods

Study design and patients

In this retrospective multicentre study conducted by the North Japan HCC Study Group consisting of 28 institutes in Northern Japan, consecutive patients with unresectable HCC who were treated with Atezo/Bev between September 2020 and August 2025 were screened. The analytical cohort comprised 695 patients who received treatment. For time-to-discontinuation analyses, 685 patients met this criterion (10 were excluded because of insufficient clinical data).

For descriptive reporting, follow-up time was calculated as the interval from Atezo/Bev initiation to the last observation date (last observation date - initiation date) and summarised using the median and interquartile range. Time to treatment discontinuation (TTD) was defined as the interval from initiation to the end date, where the end date was the recorded discontinuation date when available; thus, TTD = end date - initiation date. For patients who remained on treatment, TTD was right-censored at the last observation date.

Aetiology classification was based on the predefined indicators for HBV, HCV, ALD, and MASH in the dataset. Single aetiology was defined when liver disease was attributable to one aetiology only. Patients with mixed/other aetiologies were retained for overall estimates, but were excluded from aetiology-adjusted multivariable Cox models. Aetiology (HBV, HCV, ALD, and MASH) was assigned according to predefined aetiology indicators recorded in the multicentre dataset (aetiology flags/diagnoses). For ALD and MASH, granular metabolic risk factors ( e.g. diabetes, dyslipidaemia, hypertension, medications, and waist circumference) and quantitative alcohol intake (g/day and duration) were not uniformly captured across centres and were partially missing; therefore, centralised re-adjudication using standardised criteria was not performed. Baseline variceal status was assessed using available gastroscopy or endoscopic documentation before Atezo/Bev initiation. High-risk varices were defined as F2 or larger varices, and patients with a prior history of variceal treatment, including endoscopic variceal ligation (EVL) or endoscopic injection sclerotherapy (EIS), were also included in this category. Bevacizumab withholding (BEV-skip) was defined as temporary or permanent omission of bevacizumab during Atezo/Bev treatment, irrespective of the reason for withholding. Patients were classified as IMbrave150 criteria-compatible or criteria-incompatible according to the key eligibility criteria requirements defined in the IMbrave150 trial.

The primary endpoints were (i) TTD, (ii) distribution (proportions) of causes of discontinuation, and (iii) risk factors associated with discontinuation, including cause-specific predictors of each discontinuation cause. Discontinuation reasons were adjudicated as mutually exclusive competing events as follows: (a) progressive disease (PD)-driven discontinuation, defined as discontinuation because of radiographic progression or clinical worsening; (b) AE-driven discontinuation, defined as discontinuation primarily attributed to treatment-related AE, as judged by the treating physician, provided that these events did not result in hepatic decompensation; (c) hepatic dysfunction-driven discontinuation, defined as discontinuation because of hepatic decompensation, including the development of refractory ascites/pleural effusion and/or overt hepatic encephalopathy and/or variceal haemorrhage, in the absence of radiological PD, irrespective of the underlying mechanism; (d) response/strategy change, defined as discontinuation because of disease control ( e.g. complete response [CR]), including planned drug-free intervals and conversion to locoregional therapies ( e.g. resection or ablation); and (e) other causes, defined as reasons not captured above ( e.g. patient preference or progression of another malignancy), adjudicated as ‘Other’. When PD co-occurred with other reasons, discontinuation was classified as PD. Radiological tumour assessments were performed using contrast-enhanced computed tomography/magnetic resonance imaging in routine practice and evaluated according to RECIST v1.1. AE were graded according to Common Terminology Criteria for Adverse Events criteria (v5.0) at each participating centre.

Baseline candidate predictors included age, sex, Eastern Cooperative Oncology Group performance status (ECOG PS), tumour burden/stage variables (Barcelona Clinic Liver Cancer [BCLC] stage modelled as C vs. A/B, macrovascular invasion [MVI], extrahepatic metastasis [EM], maximum intrahepatic tumour diameter, and AFP), and hepatic reserve assessed by the ALBI score (modelled as a continuous variable).

This study adhered to the ethical principles outlined in the Declaration of Helsinki. The Iwate Medical University Hospital ethics committee approved the study protocol (approval number: MH2020-155). Participation in the study was contingent upon patients providing written informed consent to participate. Participants were also included under an institutional review board-approved opt-out procedure for individuals who had previously given consent but had not signed a study-specific consent form and had not declined participation.

Statistical analysis

TTD was summarised as median with IQR and imaged using box-and-whisker plots. Differences in TTD across discontinuation causes (PD, AE, hepatic dysfunction, and response/strategy) were assessed using the Kruskal–Wallis test, followed by a two-sided Mann–Whitney U test with Holm adjustment when appropriate.

Discontinuation causes were mutually exclusive; therefore, we adopted a competing-risk framework. Cause-specific CIFs were estimated using the Aalen–Johansen estimator, and 1-year CIFs were reported with CIF curves. Where applicable, between-group differences in CIFs were evaluated using Gray’s test.

Baseline covariates included age, sex, ECOG performance status, BCLC stage (C vs. A/B), MVI, EM, maximum intrahepatic tumour diameter (largest diameter of the largest intrahepatic lesion), log 10 (AFP+1), and ALBI score (modelled as a continuous variable). In aetiology-adjusted analyses, models were restricted to the single-aetiology cohort (HBV, HCV, ALD, or MASH), excluding mixed/other aetiologies. Missing baseline covariates (ECOG performance status and maximum intrahepatic tumour diameter) were handled using multiple imputation by chained equations with 20 imputations (m = 20; maxit = 10; seed = 20251227). Cause-specific Cox proportional hazards models were fitted for each discontinuation cause (PD-driven, AE-driven, and hepatic dysfunction-driven) within each imputed dataset, treating other discontinuation causes as competing events, and estimates were pooled using Rubin’s rules. Results are presented as hazard ratios (HRs) with 95% CI. As an exploratory sensitivity analysis, baseline portal hypertension-related factors, including ascites, hepatic encephalopathy, and variceal status, were additionally evaluated. High-risk varices were defined as F2 or larger varices, and patients with a prior history of variceal treatment were included in this category.

Post-discontinuation survival (PDS) was defined as the interval between treatment discontinuation and death from any cause or last follow-up. The PDS was estimated using the Kaplan–Meier method and compared across discontinuation causes using the log-rank test.

For clinical risk stratification and visualisation, AFP was dichotomised at 400 ng/ml and modified ALBI (mALBI) grade as 1–2a vs. 2b–3, and 1-year CIFs and 1-year status counts were summarised across the four resulting strata (2×2). All tests were two-sided, and p <0.05 was considered statistically significant. Analyses were performed using R (R Foundation for Statistical Computing, Vienna, Austria), EZR (Saitama Medical Center, Jichi Medical University, Saitama, Japan), and Prism 10.1.1 (GraphPad Software, Inc., La Jolla, CA, USA).

Results

Patient cohort, baseline characteristics, and follow-up

A total of 695 patients treated with Atezo/Bev were identified. For competing-risk analyses of TTD, 685 patients with valid date intervals and sufficient clinical information were included. Baseline characteristics of these 685 patients are summarised in Table 1 . The median age was 72 years (IQR 67–77), and 555 (81.0%) were men; the median BMI was 23.95 kg/m 2 (IQR 21.51–26.67). Aetiologies were not mutually exclusive: 131 (19.1%) patients were diagnosed with HBV, 170 (24.8%) with HCV, 217 (31.7%) with ALD, and 138 (20.4%) with MASH. Table 1 Baseline patient characteristics. Table 1 Factor Overall (N = 685) Age, years 72.0 (67–77) Sex, n (%), male/female 555 (81.0)/130 (19.0) Single aetiology, n HBV 113 HCV 159 ALD 184 MASH 126 Others, mixed aetiology 103 Child–Pugh score 5 (5–6) Child–Pugh A/B or C 615 (89.8)/70 (10.2) mALBI grade, n (%) 1 218 (31.8) 2a 191 (27.9) 2b 263 (38.4) 3 13 (1.9) BCLC stage, n (%) A 25 (3.6) B 267 (39.0) C 393 (57.4) Macrovascular invasion 127 (18.5) Maximum intrahepatic tumour diameter (cm) 4.90 (2.50–8.50) Median (range) time from treatment initiation to data cut-off (days) 433 (IQR 239–736) Plt, ×10 4 /μl 16.00 (11.70–21.30) PT-INR 1.05 (0.97–1.14) AST, IU/L 39.00 (27.00–60.00) ALT, IU/L 29.00 (19.00–46.00) T.Bil, mg/dl 0.80 (0.60–1.10) Alb, g/dl 3.70 (3.30–4.00) AFP, ng/ml 46.30 (5.10–1,450.00) IMbrave150 in/out 385/300 Best response, n (%) (missing n = 5) CR 27 (4.0) PR 188 (28.4) SD 306 (45.0) PD 142 (20.9) NE 17 (2.5) ORR, n (%) 32.4% (215/663) DCR, n (%) 78.6% (521/663) Data are presented as numbers, percentages, or median (IQR). IMbrave150 in/out: patients were classified as compatible or incompatible with the key eligibility criteria from the IMbrave150 trial). Continuous variables are summarised using IQR. AFP, alpha-fetoprotein; Alb, albumin; ALT, alanine aminotransferase; AST, aspartate aminotransferase; BCLC, Barcelona Clinic Liver Cancer; CR, complete response; DCR, disease control rate; FIB-4, fibrosis-4 index; mALBI, modified albumin–bilirubin; NA, not available; NBNC, non-HBV non-HCV; ORR, objective response rate; PD, progressive disease; Plt, platelet count; PR, partial response; PT, prothrombin time; SD, stable disease; T.Bil, total bilirubin; TNM, tumour–node–metastasis; Vp, portal venous invasion.

Regarding the tumour stage and burden, BCLC stage was A in 25 (3.6%), B in 267 (39.0%), and C in 393 (57.4%) patients. MVI was present in 127 (18.5%) and EM in 256 (37.4%) patients. The median maximum intrahepatic tumour diameter was 4.9 cm (IQR 2.5–8.5) and the median number of intrahepatic tumours was 2 (IQR 1–5).

Baseline hepatic reserve was modestly preserved, with a median Child–Pugh score of 5 (IQR 5–6) and a median ALBI score of -2.39 (IQR -2.69 to -2.05). mALBI grades were 1 in 218 (31.8%), 2a in 191 (27.9%), 2b in 263 (38.4%), and 3 in 13 (1.9%) patients. Although most patients had preserved liver function, patients with Child–Pugh class B/C were also included (n = 70, 10.2%). Accordingly, 409 (59.7%) patients were categorised as mALBI 1–2a, and 276 (40.3%) as mALBI 2b–3. Baseline AFP was 46.3 ng/ml (IQR 5.1–1450.0). The median follow-up from treatment initiation to the last observation date was 433 days (IQR 239–736).

Among single-aetiology patients (n = 582), there were 113 HBV, 159 HCV, 184 ALD, and 126 MASH cases; the remaining 103 patients had mixed/other aetiologies. Among the 685 patients, 385 were classified as IMbrave150 criteria-compatible and 300 as criteria-incompatible. The major reasons for criteria-incompatibility were non-first-line use of Atezo/Bev (n = 181) or missing treatment-line information (n = 22), Child–Pugh class B/C (n = 70), and platelet count <75,000/μl (n = 47), with some patients meeting multiple criteria. Baseline gastroscopy or endoscopic documentation of variceal status was available in 634 of 685 patients (92.6%). High-risk varices or prior variceal treatment before Atezo/Bev was identified in 42 of 634 evaluable patients (6.6%), and prior EVL or EIS was documented in 15 patients. Non-selective beta-blockers were rarely prescribed in this cohort as their use for prophylaxis of variceal bleeding is not an approved indication in Japan; consequently, variceal prophylaxis in Japanese clinical practice is mainly managed through endoscopic interventions.

Overall competing risks of treatment discontinuation

In the overall cohort (n = 685), Kaplan–Meier analysis revealed a median overall survival of 666 days (21.9 months). According to RECIST v1.1, the objective response rate was 32.4% (215/663), and the disease control rate was 78.6% (521/663). Among the 685 patients, discontinuation was attributed to PD (n = 334, 48.8%), AE (n = 125, 18.2%), hepatic dysfunction (n = 48, 7.0%), response/strategy change (n = 44, 6.4%), and other causes (n = 35, 5.1%); 99 patients (14.5%) were censored. PD co-occurred with another recorded discontinuation reason in 17 patients (5.1% of PD-classified discontinuations; 2.5% of the overall cohort). The median TTD was 155 days (IQR 68–310) for PD, 105 days (IQR 49–275) for AE, 144 days (IQR 83–276) for hepatic dysfunction, and 296 days (IQR 191–474) for response/strategy-related discontinuation (Kruskal–Wallis p = 4.0×10 -5 ). Notably, response/strategy discontinuation tended to occur later, whereas AE discontinuation occurred earlier in the treatment course ( Fig. 1 ). At 1 year, the CIF was 0.394 for PD-driven discontinuation, 0.153 for AE-driven discontinuation, 0.056 for hepatic dysfunction-driven discontinuation, 0.042 for response/strategy change, and 0.041 for other causes ( Fig. 1 ). Fig. 1 Overall patterns and timing of Atezo/Bev discontinuation. (A) Cause-specific CIFs for treatment discontinuation estimated with the Aalen–Johansen method, accounting for competing risks. PD-driven, AE-driven, hepatic dysfunction-driven, and response/strategy-driven discontinuation are shown (other reasons were treated as competing events but are not displayed). (B) The box-and-whisker plot shows TTD by the discontinuation cause; p values were calculated using the Kruskal–Wallis test. Boxes indicate the IQR with the median line; whiskers indicate 1.5× IQR and points denote outliers. AE, adverse event; Atezo/Bev, atezolizumab plus bevacizumab; CIF, cumulative incidence function; PD, progressive disease; TTD, time to treatment discontinuation.

Aetiology-specific cumulative incidence (single-aetiology cohort)

Among 685 patients, 582 had a single aetiology: HBV (n = 113), HCV (n = 159), ALD (n = 184), and MASH (n = 126). The remaining 103 patients had mixed or other aetiologies. By aetiology, the 1-year CIF for PD-driven discontinuation was 0.459 (HBV), 0.410 (HCV), 0.350 (ALD), and 0.402 (MASH) ( Fig. 2 A); 1-year CIF for AE-driven discontinuation was 0.125 (HBV), 0.118 (HCV), 0.200 (ALD), and 0.162 (MASH) ( Fig. 2 B); 1-year CIF for hepatic dysfunction-driven discontinuation was 0.018 (HBV), 0.080 (HCV), 0.044 (ALD), and 0.065 (MASH) ( Fig. 2 C). Gray’s test comparing viral (HBV/HCV) vs. non-viral (ALD/MASH) aetiologies yielded p = 0.081 for PD-driven discontinuation, p = 0.020 for AE-driven discontinuation, and p = 0.396 for hepatic dysfunction-driven discontinuation ( Fig. 2 ). Baseline AFP levels according to disease aetiology are summarised in Table S1 . Although AFP levels tended to be numerically higher in HBV-related HCC, they remained independently associated with PD-driven discontinuation after adjustment for aetiology. Fig. 2 Cause-specific cumulative incidence of treatment discontinuation. CIFs of (A) PD-driven, (B) AE-driven, and (C) hepatic dysfunction-driven discontinuation stratified by aetiology (HBV, HCV, ALD, and MASH; ‘Other’ category is not displayed). Gray’s test was used to compare CIFs across aetiology strata and between viral and non-viral aetiologies: (A) across aetiology p = 0.312, viral vs. non-viral p = 0.081; (B) across aetiology p = 0.124, viral vs. non-viral p = 0.020; (C) across aetiology p = 0.478, viral vs. non-viral p = 0.396. ALD, alcohol-associated liver disease; AE, adverse event; CIF, cumulative incidence function; MASH, metabolic dysfunction-associated steatohepatitis; PD, progressive disease.

Bevacizumab withholding according to discontinuation cause

BEV-skip was observed in 76 patients with PD-driven discontinuation (22.8%), 35 with AE-driven discontinuation (28.0%), 16 with hepatic dysfunction-driven discontinuation (33.3%), eight with response/strategy-related discontinuation (18.2%), and 11 with other causes (31.4%). The frequency of BEV-skip did not significantly differ across discontinuation causes ( p = 0.272; Table S3 ).

Cause-specific predictors of discontinuation

In multiple-imputation cause-specific Cox models (m = 20), PD-driven discontinuation was independently associated with a higher baseline AFP (HR 1.18 per log 10 (AFP+1), 95% CI 1.09–1.28; p = 3.95×10 -5 ) ( Fig. 3 A). AE-driven discontinuation was independently associated with older age (HR 1.03 per year, 95% CI 1.01–1.06; p = 0.0066), ECOG PS ≥1 (HR 1.82, 95% CI 1.21–2.74; p = 0.0039), and worse baseline liver function (ALBI per 0.1 increase: HR 1.07, 95% CI 1.03–1.12; p = 7.43×10 -4 ) ( Fig. 3 B). Hepatic dysfunction-driven discontinuation was strongly associated with older age (HR 1.07/year, 95% CI 1.03–1.12; p = 4.28×10 -4 ) and worse baseline liver function (ALBI per 0.1 increase: HR 1.29, 95% CI 1.21–1.39; p = 8.65×10 -13 ) ( Fig. 3 C). Fig. 3 Cause-specific Cox models. Models were fitted to identify baseline factors associated with (A) PD-driven, (B) AE-driven, and (C) hepatic dysfunction-driven discontinuation. For each panel, the event of interest was the corresponding discontinuation cause, and other discontinuation causes were treated as competing events (censored at the time of occurrence). Missing baseline covariates were handled using multiple imputation (m = 20), and estimates were pooled using Rubin’s rules. HCV, ALD, and MASH were analysed relative to HBV as the reference category. AE, adverse event; AFP, alpha-fetoprotein; ALBI, albumin–bilirubin; ALD, alcohol-associated liver disease; BCLC, Barcelona Clinic Liver Cancer; ECOG PS, Eastern Cooperative Oncology Group performance status; MASH, metabolic dysfunction-associated steatohepatitis; PD, progressive disease.

In an exploratory sensitivity analysis, we further evaluated the potential impact of portal hypertension by incorporating baseline portal hypertension-related factors, including ascites, hepatic encephalopathy, and variceal status. High-risk varices or prior variceal treatment before Atezo/Bev were identified in 42 of 634 evaluable patients.

In the updated cause-specific Cox models incorporating these portal hypertension-related factors, the principal findings remained largely unchanged. AFP remained associated with PD-driven discontinuation, whereas age, ECOG PS, and ALBI remained associated with AE-driven discontinuation. Among the portal hypertension-related factors, high-risk varices or prior variceal treatment was independently associated with hepatic dysfunction-driven discontinuation (HR 3.04, 95% CI 1.33–6.93; p = 0.008), but not with PD-driven or AE-driven discontinuation ( Fig. S2 ).

Clinical stratification by AFP×mALBI (four strata)

Because PD-driven discontinuation was primarily associated with a higher baseline AFP, whereas AE-driven and hepatic dysfunction-driven discontinuation were strongly associated with worse baseline liver function (ALBI), we stratified patients into four clinically interpretable groups using commonly applied cut-offs: mALBI grade 1/2a vs. 2b/3 and AFP <400 vs. ≥400 ng/ml.

Using an AFP cut-off of 400 ng/ml and mALBI (1–2a vs. 2b–3), the group sizes were as follows: G1, low AFP and good mALBI (n = 291); G2, high AFP and good mALBI (n = 118); G3, low AFP and poor mALBI (n = 165); and G4, high AFP and poor mALBI (n = 111). At 1 year, PD-driven CIF was the highest in G2 (0.520) and G4 (0.460), whereas toxicity-related CIF (AE + hepatic dysfunction) was the highest in G3 (0.339) (Gray’s test across strata: PD p = 1.214×10 -4 ; toxicity p = 2.071×10 -8 ) ( Fig. 4 ). Notably, within G3, toxicity-related discontinuation exceeded PD-related discontinuation at 1 year, whereas G2 showed predominantly PD-driven discontinuation. Fig. 4 Discontinuation patterns stratified by baseline AFP and hepatic reserve. Patients were stratified by baseline AFP (<400 vs. ≥400 ng/ml) and mALBI grade (1–2a vs. 2b–3). (A) One-year cumulative incidence of PD-driven discontinuation and toxicity-related discontinuation (AE + hepatic) is shown as grouped bars (Aalen–Johansen). (B) CIFs for PD-driven discontinuation by stratum. (C) CIFs for toxicity-related discontinuation (AE + hepatic) by stratum. Gray’s test across the four AFP×mALBI strata (ρ = 0): PD p = 1.214×10 -4 ; toxicity (AE + hepatic) p = 2.071×10 -8 . AE, adverse event; AFP, alpha-fetoprotein; CIF, cumulative incidence function; mALBI, modified albumin–bilirubin; PD, progressive disease.

To address the potential influence of non-first-line use, we performed a sensitivity analysis restricted to patients who received atezolizumab plus bevacizumab as first-line systemic therapy (n = 482). In multivariable cause-specific Cox models in this subgroup (complete-case n = 448), log 10 (AFP+1) remained associated with PD-driven discontinuation (HR 1.16, 95% CI 1.05–1.29; p = 0.004), whereas ALBI score remained associated with AE-driven discontinuation (HR per 0.1-point increase 1.09, 95% CI 1.04–1.15; p <0.001), hepatic dysfunction-driven discontinuation (HR per 0.1-point increase 1.28, 95% CI 1.18–1.39; p <0.001), and toxicity-related discontinuation (HR per 0.1-point increase 1.13, 95% CI 1.09–1.18; p <0.001). These findings were consistent with those of the overall cohort and were further supported by the AFP×mALBI stratification analysis ( Fig. S3 ).

Subsequent treatment availability according to discontinuation cause

Subsequent treatment availability significantly differed according to discontinuation cause ( p <0.001; Table S2 ). Patients with hepatic dysfunction-driven discontinuation had the lowest rate of subsequent treatment availability (27.1%) and the highest rate of no subsequent treatment (70.8%). In addition, subsequent TKI/anti-VEGF systemic therapy was introduced in only 10.4% of patients with hepatic dysfunction-driven discontinuation, compared with 49.1% of those with PD-driven discontinuation.

PDS according to discontinuation cause

PDS differed significantly according to discontinuation cause (log-rank p = 9.854×10 -9 ; Fig. S1 ). Patients with hepatic dysfunction-driven discontinuation had the shortest median PDS (168 days), whereas median PDS was 287 days in the PD-driven discontinuation group and 277 days in the AE-driven discontinuation group. In contrast, the response/strategy-change group showed the most favourable prognosis, and the median PDS was not reached.

Discussion

In unresectable HCC treated with Atezo/Bev, durable disease control and longer treatment exposure have been linked to improved overall survival. 20 Therefore, aside from planned strategy-driven discontinuation ( e.g. conversion), treatment discontinuation is a clinically important event. In our multicentre real-world cohort (n = 685), 586 patients discontinued Atezo/Bev; PD was the most frequent reason (n = 334, 48.8%), followed by AE (n = 125, 18.2%) and hepatic dysfunction (n = 48, 7.0%), whereas response/strategy-driven discontinuation occurred in 44 patients (6.4%). The 1-year cumulative incidence estimates were 0.394 for PD, 0.153 for AE, and 0.056 for hepatic dysfunction (Gray’s test p = 1.6×10 -3 ), underscoring that non-PD discontinuation represents a substantial component of Atezo/Bev attrition in routine practice.

Importantly, pretreatment determinants differed by the discontinuation reason. PD-related discontinuation was primarily associated with a higher baseline AFP, whereas AE-related discontinuation was associated with older age, ECOG performance status ≥1, and poorer hepatic reserve. Hepatic dysfunction–related discontinuation was even more strongly linked to age and ALBI. These findings suggest that the routinely available pretreatment markers can help anticipate the most likely reason for discontinuation and inform more individualised monitoring and supportive care at treatment initiation. Importantly, hepatic dysfunction-driven discontinuation was associated with markedly poor PDS compared with PD-driven or AE-driven discontinuation. These findings suggest that preservation of hepatic reserve may critically influence not only treatment continuation but also post-discontinuation prognosis and access to subsequent therapy.

Our discontinuation profile was broadly consistent with those of previous real-world studies of Atezo/Bev. Prior reports demonstrated that impaired liver function was associated with hepatic decompensation and poor prognosis, whereas tumour progression was associated with higher AFP levels. [17] , [18] , [19] In the present study, PD was the most frequent cause of discontinuation, while hepatic dysfunction-driven discontinuation accounted for 7.0% of patients. These findings support previous evidence linking liver reserve to decompensation-related outcomes while extending prior studies by quantifying cause-specific discontinuation probabilities over time.

In systemic anticancer therapy, discontinuation may occur for mutually exclusive reasons such as progression and toxicity. When competing events exist, conventional Kaplan–Meier methods can overestimate reason-specific probabilities; competing-risk analysis has therefore become a standard approach to separate and quantify each reason. 21 , 22 Competing-risk regression has been used to identify predictors of toxicity-related discontinuation and develop practical risk stratification in other malignancies, including renal cell carcinoma, chronic lymphocytic leukaemia, and in older patients with multiple myeloma. [23] , [24] , [25] Our analyses extend this established framework to Atezo/Bev-treated HCC, where both tumour progression and liver-of treatment-related events are clinically consequential.

In HCC, underlying cirrhosis can drive treatment discontinuation through hepatic decompensation independent of radiological progression. 7 , 26 During treatment with Atezo/Bev, VEGF exposure adds a distinct risk profile, including variceal bleeding in patients with portal hypertension, emphasising careful baseline evaluation and prophylaxis. 27 Although exploratory, the sensitivity analysis incorporating portal hypertension-related factors suggested that clinically relevant variceal status may contribute to hepatic dysfunction-driven discontinuation, supporting the importance of baseline endoscopic assessment before Atezo/Bev initiation. Moreover, HCC includes a clinically favourable category of response/strategy-driven discontinuation ( e.g. conversion to locoregional therapy, planned drug-free intervals, or elective switching under disease control). These scenarios differ fundamentally from PD- or toxicity-related treatment discontinuation; therefore, they should be distinguished to preserve interpretability of the discontinuation endpoints. 12 , 26

The discontinuation timing differed by reason. Median TTD was 105 days for AE (IQR 49–275), 144 days for hepatic dysfunction (IQR 83–276), and 155 days for PD (IQR 68–310), whereas response/strategy-driven discontinuation occurred later (median 296 days, IQR 191–474; Kruskal–Wallis p = 4.0×10 -5 ). PDS was particularly poor after hepatic dysfunction (median 168 days) compared with that after PD (287 days) or AE (277 days). Together, these results support an early-phase focus on proactive AE management and close hepatic monitoring, especially in patients with limited hepatic reserve, whereas patients with sustained disease control should be periodically evaluated for optimal strategy changes such as conversion.

Tumour- and liver-driven risks were separable; therefore, we explored a pragmatic pretreatment stratification using routinely available markers, including AFP (<400 vs. ≥400 ng/mL) 28 and hepatic reserve (mALBI 1–2a vs. 2b–3, as applicable), 29 yielding four clinical groups. This simple 2×2 framework distinguishes progression-prone profiles from liver- or AE-prone profiles: high AFP with preserved hepatic reserve group tended to discontinue mainly because of PD, whereas impaired hepatic reserve (particularly in older patients) was associated with discontinuation as a result of toxicity or hepatic dysfunction. Such risk awareness at treatment initiation may help in tailoring the visit frequency, laboratory monitoring, dose-intensity decisions, and supportive care to maintain treatment when appropriate.

In practical terms, AFP reflects tumour aggressiveness and adverse biology, whereas ALBI/mALBI reflects susceptibility to treatment- or liver-related discontinuation. Using both may help in prioritising monitoring and supportive interventions beyond OS/progression-free survival alone. 7 , 12 , 29

In the aetiology-restricted cohort, viral vs. non-viral aetiology showed only modest differences in the unadjusted cumulative incidence curves, and these differences were attenuated after multivariable adjustment, suggesting that baseline clinical profiles and management practices may explain much of the apparent aetiology effect.

This study has some limitations. First, its retrospective multicentre design entails missing data, including the number of intrahepatic lesions, and heterogeneity in assessment and management. ALD/MASH labels came from site records, whereas data on metabolic factors and alcohol intake were incomplete. As standardised re-adjudication and dose–response analyses were not possible, the interpretation of aetiology results requires caution. Longitudinal ALBI trajectory data, the number of intrahepatic lesions, anticoagulation use and cardiovascular comorbidities were not uniformly available across centres. Although bevacizumab withholding was captured as BEV-skip, detailed information on the timing, frequency, and duration of bevacizumab omissions was not uniformly available. Second, discontinuation reasons were determined from the medical records; when PD co-occurred with AE and/or hepatic dysfunction, cases were preferentially classified as PD, which may underestimate non-PD discontinuation. Although we used an operational definition of hepatic decompensation, the classification of the primary reason for treatment discontinuation relied on the treating physician’s assessment at each participating centre without central adjudication. Accordingly, inter-institutional differences in diagnostic thresholds and documentation may have introduced some heterogeneity and potential misclassification in the discontinuation causes. In particular, overlap between immune-mediated liver injury and cirrhotic/hepatic decompensation may have complicated the attribution of discontinuation causes in some cases. Nevertheless, we attempted to mitigate these issues by prespecifying definitions and explicitly separating events accompanied by radiological tumour progression. Third, portal hypertension management, AE mitigation, and post-discontinuation treatments varied across institutions, leading to residual confounding. Fourth, because the reasons for discontinuation were retrospectively determined from medical records and treating-physician judgement, without central adjudication, inter-physician and inter-centre variability may have influenced the attribution of AE-driven or hepatic dysfunction-driven discontinuation. Finally, the cohort predominantly comprised Japanese patients; therefore, generalisability to other populations requires confirmation.

In this large real-world cohort of unresectable HCC treated with Atezo/Bev, competing-risk analyses showed that PD was the dominant reason for discontinuation, whereas a clinically meaningful proportion stopped as a result of AE or hepatic dysfunction. These discontinuation causes were associated with distinct pretreatment factors: higher AFP for PD-driven discontinuation and poorer hepatic reserve with older age for AE/hepatic dysfunction-driven discontinuation. The AFP×mALBI-based four-strata framework may provide a simple, hypothesis-generating approach to risk stratification for guiding monitoring intensity and supportive care, although external validation is warranted.

Abbreviations

AE, adverse event; AFP, alpha-fetoprotein; ALBI, albumin–bilirubin; ALD, alcoholic liver disease; ALT, alanine aminotransferase; AST, aspartate aminotransferase; Atezo/Bev, atezolizumab plus bevacizumab; BCLC, Barcelona Clinic Liver Cancer; BEV-skip, bevacizumab withholding; CIF, cumulative incidence function; CR, complete response; CTLA-4, cytotoxic T-lymphocyte-associated protein 4; DCR, disease control rate; ECOG PS, Eastern Cooperative Oncology Group performance status; EIS, endoscopic injection sclerotherapy; EM, extrahepatic metastasis; EVL, endoscopic variceal ligation; FIB-4, fibrosis-4 index; HCC, hepatocellular carcinoma; HR, hazard ratio; IQR, interquartile range; mALBI, modified albumin–bilirubin; MASH, metabolic dysfunction-associated steatohepatitis; MVI, macrovascular invasion; NA, not available; NBNC, non-HBV non-HCV; ORR, objective response rate; OS, overall survival; PD, progressive disease; PD-L1, programmed death-ligand 1; PDS, post-discontinuation survival; PFS, progression-free survival; Plt, platelet count; PR, partial response; PT, prothrombin time; SD, stable disease; T.Bil, total bilirubin; TNM, tumour–node–metastasis; TTD, time to treatment discontinuation; VEGF, vascular endothelial growth factor; Vp, portal venous invasion.

Authors’ contributions

Designed the study: GS. Collected the data: TSho, TA, KE, YF, HN, GI, MT, CI, KY, SS, MNakai, MO, TKitagataya, TKatsumi, KM, WS, SM, KA, MF, YY, MB, TM, TO, NT. Performed the statistical analyses: GS. Provided hepatological expertise: TSho, MNinomiya, YU, NSakamoto, HK. Wrote and revised the manuscript: GS. Edited the manuscript: TSho, MNinomiya, YU, NSakamoto, HK. Revised the manuscript for important intellectual content: TSho, MNinomiya, YU, NSakamoto, HK. Reviewed the manuscript: TSho, TA, KE, YF, HN, GI, MT, CI, KY, SS, MNakai, MO, TKitagataya, TKatsumi, KM, WS, SM, KA, MF, YY, MB, TM, TO, NT. Read and approved the final manuscript: all authors.

Data availability

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

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

During the preparation of this work, the authors used generative AI tools for language editing and to improve readability. The authors reviewed and edited all AI-assisted output and take full responsibility for the content of the manuscript.

Financial support

This study was supported in part by grants from the 10.13039/100009619 Japan Agency for Medical Research and Development (AMED Grant Nos. JP25fk0210126, JP25fk0310545, JP24fk0210121, JP25fk0210142, JP25fk0310551, JP25fk0210123, JP25fk0310535, JP25fk0210157, JP25fk0310543, JP25fk0210172, JP25fk0210174, and JP25fk0210143). The sponsor had no role in study design; in the collection, analysis, and interpretation of data; in the writing of the report; or in the decision to submit the article for publication.

Conflicts of interest

NS received lecture fees from Chugai Pharmaceutical Co. Ltd. and research grants from Gilead Sciences Inc. and AbbVie Inc. GS received research grants from Gilead Sciences. YU has received lecture fees from Chugai Pharmaceutical Co., Ltd. and Eisai Co., Ltd. HK has received lecture fees from Eisai Co., Ltd., Chugai Pharmaceutical Co., Ltd., and AstraZeneca. All other authors declare no conflicts of interest.

Please refer to the accompanying ICMJE disclosure forms for further details.

References

1 Singal A.G. Kanwal F. Llovet J.M. Global trends in hepatocellular carcinoma epidemiology: implications for screening, prevention and therapy Nat Rev Clin Oncol 20 2023 864 884 10.1038/s41571-023-00825-3 37884736

2 Singal A.G. Llovet J.M. Yarchoan M. AASLD Practice Guidance on prevention, diagnosis, and treatment of hepatocellular carcinoma Hepatology 78 2023 1922 1965 10.1097/HEP.0000000000000466 PMC10663390 37199193

3 European Association for the Study of the Liver Clinical Practice Guidelines on the management of hepatocellular carcinoma J Hepatol 82 2025 315 374 10.1016/j.jhep.2024.08.028 39690085

4 Hasegawa K. Takemura N. Yamashita T. Clinical practice guidelines for hepatocellular carcinoma: the Japan society of hepatology 2021 version (5th JSH-HCC guidelines) Hepatol Res 53 2023 383 390 10.1111/hepr.13892 36826411

5 Lau G. Obi S. Zhou J. APASL clinical practice guidelines on systemic therapy for hepatocellular carcinoma-2024 Hepatol Int 18 2024 1661 1683 10.1007/s12072-024-10732-z 39570557

6 Yau T. Galle P.R. Decaens T. Nivolumab plus ipilimumab versus lenvatinib or sorafenib as first-line treatment for unresectable hepatocellular carcinoma (CheckMate 9DW): an open-label, randomised, phase 3 trial Lancet 405 2025 1851 1864 10.1016/S0140-6736(25)00403-9 40349714

7 Finn R.S. Qin S. Ikeda M. Atezolizumab plus bevacizumab in unresectable hepatocellular carcinoma N Engl J Med 382 2020 1894 1905 10.1056/NEJMoa1915745 32402160

8 Cheng A.L. Qin S. Ikeda M. Updated efficacy and safety data from IMbrave150: atezolizumab plus bevacizumab vs. sorafenib for unresectable hepatocellular carcinoma J Hepatol 76 2022 862 873 10.1016/j.jhep.2021.11.030 34902530

9 Tada T. Kumada T. Hiraoka A. Outcomes of patients with hepatocellular carcinoma treated with atezolizumab plus bevacizumab in real-world clinical practice who met or did not meet the inclusion criteria for the phase 3 IMbrave150 trial Aliment Pharmacol Ther 60 2024 233 245 10.1111/apt.18037 38716823

10 Hosoda S. Suda G. Sho T. Low Baseline CXCL9 predicts early progressive disease in unresectable HCC with atezolizumab plus bevacizumab treatment Liver Cancer 12 2023 156 170 10.1159/000527759 PMC10267515 37325489

11 Kohya R. Suda G. Ohara M. Serum FGF21 as a predictor of response to atezolizumab and bevacizumab in HCC JHEP Rep 7 2025 101364 10.1016/j.jhepr.2025.101364 PMC11999275 40242310

12 Ohara M. Suda G. Kohya R. Three-year overall survival in unresectable hepatocellular carcinoma treated with atezolizumab plus bevacizumab Hepatol Int 19 2025 1371 1381 10.1007/s12072-025-10875-7 40875102

13 Yang Z. Suda G. Sho T. Association of proteinuria with improved prognosis in unresectable hepatocellular carcinoma treated with atezolizumab and bevacizumab, and the predictive role of serum vascular endothelial growth factor D levels: a multicenter retrospective study Hepatol Res 55 2025 433 443 10.1111/hepr.14139 39584803

14 Suzuki T. Narita K. Matsuura K. The effectiveness of long-term continuation of atezolizumab plus bevacizumab in patients receiving systemic chemotherapy for advanced hepatocellular carcinoma: a multicenter study Oncology 104 2025 1 9 10.1159/000544051 40159345

15 Kudo M. Yamashita T. Finn R.S. Depth and duration of response are associated with survival in patients with unresectable hepatocellular carcinoma: exploratory analyses of IMbrave150 Liver Cancer 14 2025 539 554 10.1159/000544981 PMC12052357 40331064

16 Putter H. Fiocco M. Geskus R.B. Tutorial in biostatistics: competing risks and multi-state models Stat Med 26 2007 2389 2430 10.1002/sim.2712 17031868

17 Celsa C. Cabibbo G. Fulgenzi C.A.M. Hepatic decompensation is the major driver of mortality in patients with HCC treated with atezolizumab plus bevacizumab: the impact of successful antiviral treatment Hepatology 81 2025 837 852 10.1097/HEP.0000000000001026 39028886

18 de Castro T. Jochheim L.S. Bathon M. Atezolizumab and bevacizumab in patients with advanced hepatocellular carcinoma with impaired liver function and prior systemic therapy: a real-world experience Ther Adv Med Oncol 14 2022 17588359221080298 10.1177/17588359221080298 PMC8891886 35251317

19 Piseddu I. Jochheim L.S. Boettcher K. Early mortality in atezolizumab/bevacizumab for HCC is associated with impaired liver function and alterations of systemic immunity JHEP Rep 7 2025 101513 10.1016/j.jhepr.2025.101513 PMC12861967 41635626

20 Shen Y.C. Liu T.H. Nicholas A. Clinical outcomes and histologic findings of patients with hepatocellular carcinoma with durable partial response or durable stable disease after receiving atezolizumab plus bevacizumab J Clin Oncol 42 2024 4060 4070 10.1200/JCO.24.00645 PMC11608592 39197119

21 Dignam J.J. Kocherginsky M.N. Choice and interpretation of statistical tests used when competing risks are present J Clin Oncol 26 2008 4027 4034 10.1200/JCO.2007.12.9866 PMC2654314 18711194

22 Austin P.C. Lee D.S. Fine J.P. Introduction to the analysis of survival data in the presence of competing risks Circulation 133 2016 601 609 10.1161/CIRCULATIONAHA.115.017719 PMC4741409 26858290

23 Kaymakcalan M.D. Xie W. Albiges L. Risk factors and model for predicting toxicity-related treatment discontinuation in patients with metastatic renal cell carcinoma treated with vascular endothelial growth factor-targeted therapy: results from the International Metastatic Renal Cell Carcinoma Database Consortium Cancer 122 2016 411 419 10.1002/cncr.29773 26540173

24 Hampel P.J. Ding W. Call T.G. Rapid disease progression following discontinuation of ibrutinib in patients with chronic lymphocytic leukemia treated in routine clinical practice Leuk Lymphoma 60 2019 2712 2719 10.1080/10428194.2019.1602268 PMC6813846 31014142

25 Okayama Y. Takakuwa T. Shimura Y. Insights into very elderly multiple myeloma treatment from Kansai Myeloma Forum Hematology 30 2025 2496545 10.1080/16078454.2025.2496545 40296514

26 Cabibbo G. Celsa C. Battaglia S. Early hepatic decompensation identifies patients with hepatocellular carcinoma treated with atezolizumab plus bevacizumab or sorafenib at highest risk of death Clin Cancer Res 31 2025 543 550 10.1158/1078-0432.CCR-24-2582 39576248

27 Ha Y. Kim J.H. Cheon J. Risk of variceal bleeding in patients with advanced hepatocellular carcinoma receiving atezolizumab/bevacizumab Clin Gastroenterol Hepatol 21 2023 2421 3.e2422 10.1016/j.cgh.2022.07.035 35944830

28 Hsu C.Y. Liu P.H. Lee Y.H. Using serum alpha-fetoprotein for prognostic prediction in patients with hepatocellular carcinoma: what is the most optimal cutoff? PLoS One 10 2015 e0118825 10.1371/journal.pone.0118825 PMC4349891 25738614

29 Nakagawa M. Inoue M. Ogasawara S. Clinical effects and emerging issues of atezolizumab plus bevacizumab in patients with advanced hepatocellular carcinoma from Japanese real-world practice Cancer 129 2023 590 599 10.1002/cncr.34559 36426410

Supplementary data

The following are the Supplementary data to this article: Multimedia component 1 Multimedia component 2 Multimedia component 3 Multimedia component 4

Acknowledgements

We thank all patients who participated in this study and their families for their cooperation and support.

Author names in bold designate shared co-first authorship

Supplementary data to this article can be found online at https://doi.org/10.1016/j.jhepr.2026.101988 .