Impact of surveillance for hepatocellular carcinoma on survival in patients with compensated cirrhosis

MARKOV SIMULATION METHOD

A multistate Markov microsimulation model was constructed simulating a cohort of 50-year old patients with compensated cirrhosis. The cycle length of the model was 1 month and the Markov cycle was repeated for 180 cycles or 15 years. HCC surveillance would be performed every 6 months. Figure 1 demonstrates the Markov state transition model used in this study. In the model, the benefits of surveillance would result from HCCs diagnosed at an early stage, which may be treated with a potentially curative modality, namely LTx, surgical resection, or radiofrequency ablation (RFA). A cohort of 100,000 subjects was modeled under two scenarios; no surveillance (n = 50,000) and surveillance (n = 50,000)

Subjects in all groups could, over time, experience hepatic decompensation and/or develop HCC. HCC would progress from early to advanced stages. Subjects under surveillance would be more likely to be diagnosed with an early stage HCC, whereas those not under surveillance would be more likely to be diagnosed at an early stage less frequently. The extent to which subjects under surveillance are diagnosed with early stage HCC would depend on the sensitivity of the surveillance test. Subjects with HCC may die because of treatment complications, progressive or recurrent HCC, decompensated liver disease, or unrelated causes.

The primary outcome of interest was mortality from all causes and from HCC. In addition to the number of deaths in each cohort, we calculated the number of subjects that needed to undergo surveillance to save one life (number needed to be in surveillance [NNS]) from all-cause and HCC-specific mortality. This was obtained following the standard method to compute the number needed to treat (inverse of the absolute risk reduction). In addition, relative risk reduction was calculated by the ratio of the difference between mortality without surveillance and mortality with surveillance divided by mortality without surveillance. TreeAge Pro (v20151120; Williamstown, MA) was used to generate the Markov model.

INPUT VARIABLES

The main input data included in this analysis were extracted from the published literature; however, because not all of the variables needed to execute the model could be populated in a robust, evidence-based fashion, an expert opinion was used to supplement the available data. Sensitivity analyses were undertaken (as described in detail below) to evaluate the degree to which the study conclusion would be affected by the uncertainty in the data.

Key transition probabilities in the model included annual incidence of HCC (3%; range, 1%-5%)9, 10 and median annual mortality of incurable HCC (50%; range, 45%-55%).11, 12 The mean interval of tumor progression between early, detectable lesions to advanced stage was set to be 1 year, based on the reported doubling time of HCC of approximately 6 months.13 Median time-to-event for annual HCC mortality and mean tumor progression were converted into transition probabilities using standard declining exponential functions (1 − e^−ht, where h is ln(2)/Median Time or 1/Mean Time). We modeled cirrhosis progression stochastically as well. Transition probabilities for the four stages of cirrhosis were used based on data from a large meta-analysis, and the probability of death from hepatic decompensation was based on cirrhosis stage.14

In patients undergoing LTx, postoperative mortality was set to be 9% in the first 12 months and then 2% annually thereafter.15 Procedure-related mortality was assigned to be 1% and 5% in the first 6 months after RFA16, 17 and liver resection,18-20 respectively. There would be no procedure-related mortality for RFA or liver resection beyond 6 months after the treatment. We modeled that 70% (combining metastatic recurrence at 4.4% annually over 5 years and de novo HCC or intrahepatic “early” HCC recurrence at 17.8% annually) of patients receiving RFA or resection develop HCC recurrence over 5 years, whereas 13% of patients receiving LT develop metastatic recurrent disease over 5 years. 21-23 Progression of cirrhosis stage and the risk of death from hepatic decompensation would persist except in patients undergoing LTx. In all cohorts, age-specific mortality from unrelated causes was taken from the life expectancy data of the United States population.24

BASE CASE SCENARIO AND SENSITIVITY ANALYSES

We estimated that 30% in the no-surveillance cohort would be diagnosed at an early stage of HCC and would be eligible for curative treatment in the same fashion as early HCC patients of the surveillance cohort.25 We have modeled that 75% of individuals with early stage HCC would receive potentially curative treatment (LTx, RFA or resection).26, 27

In the base case scenario, LTx would be available for 17% of patients who would have died of hepatic decompensation.28 We estimated that 33% of early stage HCC patients would be eligible for LTx listing; the remaining two thirds would receive either RFA or resection, depending on their cirrhosis stage. Patients with stage 1 cirrhosis would receive resection, whereas RFA would be available for patients in all other stages. The waiting time before receiving LTx for HCC would be 12 months, at which time 80% would undergo LTx, whereas the remaining 20% of HCC patients listed for LTx would drop out due to tumor progression, death from cirrhosis, death from natural causes, or other reasons.29, 30

We assumed that 18% of patients with early stage HCC had very early stage HCC (single tumor <2 cm).31 Because patients with Barcelona Clinic Liver Cancer (BCLC) stage 0 HCC do not qualify for Model for End-Stage Liver Disease exception score in the current organ allocation system in the United States, we have modeled that they receive resection or RFA. Alternatively, they can wait for 6 months without any HCC treatment to simulate a situation in which patients wait until the tumor grows into United Network for Organ Sharing T2 stage to be listed for LTx. LTx would be considered as a treatment option in 33% of patients for BCLC stage 0 patients only after waiting for 6 months. We have assumed that there is no metastatic recurrence after curative treatment for very early stage HCC, but patients continue to have risk of developing de novo HCC after curative treatment.

One-way sensitivity analysis was undertaken to identify variables that have an important impact on the results of the analyses. These include (1) no progression of cirrhosis beyond stage 1, (2) annual incidence of HCC (range, 1%-5%), (3) sensitivity of HCC surveillance test (range, 50%-90%), (4) availability of LTx for hepatic decompensation (range, 0%-100%), (5) time to develop a late stage cancer from an early stage (range, 4-36 months), and (6) annual mortality due to late stage HCC (range, 40%-80%). Finally, we incorporated discounting for future years of life, recognizing the time preference in evaluation of life over time. Without making a value judgment about the effect of hepatic decompensation and HCC on the utility of the patient, we used the standard discount rate of 3% uniformly in the model and assessed its effect on the results.

Probabilistic sensitivity analysis (PSA) was performed to address uncertainty beyond limited one-way sensitivity analyses. Uncertainty regarding the base case value of each variable (i.e., second-order uncertainty) and the impact of such uncertainty on base case results was examined in PSA by sampling all variables over their ranges (Table 1), using triangular or beta distributions. The Markov analysis was performed for 100 iterations (100,000 subjects in each iteration, 50,000 per strategy), where each iteration represented a different sampling of variables from the assumed distributions.

BASE CASE SCENARIO

Figure 2 summarizes the overall result indicating that HCC surveillance improves survival. The median survival in the surveillance cohort was 11.2 years compared with 10.4 years in the no-surveillance cohort. In Table 2, mortality from all causes at 15 years improved from 68.2% without surveillance to 64.7% with surveillance. Without surveillance, HCC was the cause of death in 25.1% of patients and hepatic decompensation in 43.6% of patients. In the surveillance cohort, the proportion of patients dying of HCC decreased to 17.8%, whereas that of hepatic decompensation increased to 47.1%.

Table 2 also provides the NNS and relative risk reduction calculations. The NNS to reduce one death decreased progressively from 3 years to 15 years. At 15 years, the NNS to avoid one HCC death was 18 and one death from all cases was 28. Relative risk reduction was fairly constant regardless of the time frame of assessment. Surveillance was associated with an approximate 30%-35% relative reduction in HCC mortality throughout. Relative reduction in overall mortality was 5.2% with surveillance at 15 years.

MODEL VALIDATION

Using the model, we calculated the expected survival of HCC patients according to the treatment received. Patients who received LTx had a post-LTx survival rate of 71% at 5 years, whereas those who underwent resection or RFA had an overall survival rate of 61% at 5 years, results that are consistent with a general consensus of survival results reported in the literature.21-23 Patients with BCLC stage 0 who were treated with RFA had a 5-year survival rate of 78%, which calibrates with the result seen in a recent meta-analysis.32 Next, we evaluated the overall survival of HCC patients from the onset and diagnosis of HCC (Supporting Figs. S1 and S2, respectively). The median survival of HCC patients was longer in the surveillance cohort than in the no-surveillance cohort (3.1 years versus 2.2 years from the time of HCC onset and 2.8 years versus 1.4 years from the time of HCC diagnosis). The 5-year overall survival from HCC diagnosis was 34% with surveillance and 15% without surveillance, both of which fall well within the range of survival rates reported in the literature.33

COMPETING RISKS ON THE EFFECTIVENESS OF HCC SURVEILLANCE

First, we tested a scenario in which cirrhosis did not progress and thus deaths from hepatic decompensation occurred at a lower frequency than the base case scenario. In Table 3, cirrhosis remained at stage 1 throughout the follow-up, which resulted in a reduction in deaths from hepatic decompensation and an increase in HCC mortality, with a net effect of overall reduction in all-cause mortality. The NNS to prevent one death from all causes at 15 years decreased by 37% (from 28 to 18) and that from HCC decreased by 24% (from 18 to 13).

In addition, we considered a scenario in which liver disease remained stable with concomitant reduction in the incidence of HCC, as would be expected in patients who have cirrhosis with viral hepatitis who have received effective antiviral therapy (Supporting Table S1). As expected, all-cause and HCC-specific mortality decreased compared with the base case scenario. The NNS for all-cause and HCC mortality increased; the NNS over 15 years for overall mortality increased by 57% (from 28 to 44) and that for HCC mortality increased by 90% (from 18 to 34).

OTHER SENSITIVITY ANALYSES

As expected, the annual incidence of HCC affected the effectiveness of HCC surveillance: higher annual incidence rates of HCC were associated with lower NNS to prevent one all-cause or HCC-specific mortality and larger relative risk reduction for all-cause mortality (Figs. 3A and 4A, Supporting Table S2, and Supporting Figs. S3A and S4A). Similarly, a higher sensitivity of HCC surveillance test was associated with a larger effect of surveillance, although its magnitude was smaller than that from variability in HCC incidence (Figs. 3B and 4B, Supporting Table S3, and Supporting Figs. S3B and S4B). It is noteworthy that the effect of surveillance test sensitivity was very small when assessed for the long term (e.g., 15 years) (Figs. 3B and 4B).

Other variables tested had a much smaller effect on the survival benefits of HCC surveillance. For example, availability of LTx had little impact on the effectiveness of surveillance, as gauged both by NNS and relative risk reduction (Supporting Table S4). Similarly, varying access to LTx for hepatic decompensation had little impact on the benefits of HCC surveillance (Figs. 3C and 4C, Supporting Table S5, and Supporting Figs. S3C and S4C). Finally, other variables subjected to sensitivity analyses—namely, time to develop late stage cancer and mortality of late stage HCC—had negligible impact on the overall results (data not shown).

PSA was performed to address the second order uncertainty (uncertainty around the parameters used in the base case analysis) along with first order uncertainty (patient level) (Fig. 5 and Supporting Fig. S5). All-cause and HCC-specific mortality reduction were above 0%, indicating that there are no instances in which there would be no risk reduction; for example, the median 5-year risk reduction in all-cause mortality was 7.8%. The 2.5th and 97.5th percentiles of the PSA were 3.3% and 11.4%, respectively. The median 5-year risk reduction in HCC-specific mortality 34.8%. The 2.5th and 97.5th percentiles of the PSA were 30.8% and 40.7%, respectively.

Finally, a discounted life year analysis was performed with reducing the value of life by 3% per year. HCC surveillance offered, in the base case scenario, an increase of 0.24 life years for a subject relative to no surveillance (discounted life years: 8.52 versus 8.27). In general, it reduced the apparent benefits of surveillance in other scenarios without fundamentally altering the direction of effects (data not shown).