Effects of Cirrhosis and Diagnosis Scenario in Metabolic-Associated Fatty Liver Disease-Related Hepatocellular Carcinoma

Study Design and Patient Population

We performed an international, multicenter, retrospective, cohort study of MAFLD-related HCC at 11 medical centers. Details of the centers are shown in Supporting Table S1. The study period was from 2005 to 2017. Inclusion criteria were MAFLD and a first diagnosis of HCC, as verified by chart review. As recommended by a recent international consensus forum,^{(3, 4)} MAFLD was diagnosed based on any of the following: diabetes, race-adjusted overweight (body mass index ≥23 for Asians and ≥25 for non-Asians), or both dyslipidemia and hypertension. We also excluded participants with other etiologies of liver disease, such as viral hepatitis, excessive alcohol use, and autoimmune diseases as we have already reported findings for patients with liver diseases of other etiologies. HCC was diagnosed based on 2010 American Association for the Study of Liver Diseases criteria.⁽³¹⁾ Presence of steatosis was not a requirement as it is often not present by the time patients develop cirrhosis and HCC. Patients with prior HCC or liver transplant were excluded.

This study was approved by the Institutional Review Board at Stanford University (Stanford, CA) and at each of the other participating centers.

Definition of Cirrhosis and Screening

Laboratory values, imaging findings, and HCC treatments were obtained from individual medical records. Patients were designated as having cirrhosis based on histologic or noninvasive elastographic methods demonstrating fibrosis stage 4, clinical portal hypertension (i.e., otherwise unexplained splenomegaly or platelet count <120,000/μL, ascites, or varices on imaging/endoscopy), prior hepatic decompensation (hepatic encephalopathy, ascites, variceal gastrointestinal bleeding), or decreased synthetic function (total bilirubin >2.0 mg/dL or international normalized ratio >1.2 without alternative explanation).

Screening was estimated using two methods. First, we assessed whether patients had had liver imaging within the year before HCC diagnosis; it could not be reliably determined whether these imaging studies were obtained for screening purposes or other indications. Second, we assessed the scenario in which patients were diagnosed: without symptoms (implying that diagnosis was made based on screening) versus with symptoms of HCC, such as new-onset or worsening hepatic decompensation, abdominal pain, or weight loss at the time of HCC diagnosis.

Tumor Staging and Survival Outcomes

Tumor characteristics were determined by multiphase contrast-enhanced computed tomography or magnetic resonance imaging. Patients were followed from the date of diagnosis with HCC and either date of death or last follow-up. For U.S. sites, the medical record was supplemented by a National Death Index registry search up to March 19, 2019. The National Death Index registry is a centralized database of death record information on file in state vital statistics offices; it has over 90% completion for most states and 99% for the state of California where the Stanford cohort is located.⁽³²⁾

Statistical Analysis

Descriptive statistics were reported as percentage for categorical variables and mean ± SD or median (interquartile range) for continuous variables. Normally distributed continuous variables were compared by the Student t test, and Wilcoxon rank-sum statistics were applied when continuous variables were not normally distributed. The chi square test was used to compare categorical variables.

For time-to-event analyses, the primary outcome was overall survival. Follow-up was calculated for each patient as the time from dates of HCC diagnosis to the date of death or to the date when patients were last known to be alive. Mortality rates by various disease states were calculated and expressed per 100 person-years. The Kaplan-Meier method was used to estimate overall survival of patients; those lost to follow-up were censored. Differences in overall survival among subgroups were compared using the log-rank test. A Cox proportional hazards model was developed to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) relating cirrhosis, imaging history, and other risk factors with overall survival. For the multivariable model, we used the albumin-bilirubin (ALBI) score as a measurement for liver function and Milan criteria for tumor staging instead of the Child-Pugh Model for End-Stage Liver Disease (MELD) score or Barcelona Clinic Liver Cancer (BCLC) stage because we had the most complete data for the ALBI score and Milan criteria. We conducted lead-time analyses, as reported, using an estimated sojourn of 6 months.^{(30, 33)}

Statistical significance was defined as two-tailed P < 0.05. All statistical analyses were performed using R version 3.5.1 (R Foundation for Statistical Computing, Vienna, Austria; www.r-project.org) with the tidyverse package (www.tidyverse.org).

In addition, we used machine learning to determine factors associated with prognosis. The dependent variable was overall survival based on a rank-based method to account for censored data.⁽³⁴⁾ Values that were missing in over half of patients were removed; following this step, missing values were imputed using the “mice” package with five imputations and a maximum of 50 iterations.⁽³⁵⁾ The machine learning algorithm used was a random forest model with tune length 5 with the “party” package version 1.3.3.⁽³⁶⁾ We used the Stanford University Medical Center, Ogaki Municipal Hospital, and National University Health System cohorts as validation cohorts because together they provided a balance of U.S. and Asian patients and constituted approximately 27% of the overall cohort size. All other centers’ data were used as the training cohort. The model was trained using 10-fold cross-validation on the training cohort and then applied to the validation cohort. We chose the top three nonredundant contributors to survival and added them into a risk score based on their relative importance.

Overall Characteristics and Treatment

In total, 1,382 patients with MAFLD-related HCC met study inclusion criteria and were included in this study. Overall characteristics for the cohort are shown in Table 1. The cohort was 68% men, 46% Asian, 34% Caucasian, and 7% Hispanic, with a mean age of 67 years. Cirrhosis was present in 62% of patients. There was no difference in age or sex between patients with or without cirrhosis, but patients with cirrhosis were more frequently Caucasian or Hispanic and less frequently Asian, more frequently had diabetes, and were less frequently obese. Having undergone abdominal imaging within the year before HCC diagnosis was more common among patients with cirrhosis (30% vs. 16%). Compared to patients without cirrhosis, patients with cirrhosis had smaller maximal tumor size (median 3.3 vs. 5.7 cm; P < 0.001); they also more frequently had early stage (BCLC 0 or A) disease (43% vs. 29%) and less frequently had intermediate-stage (BCLC B) disease (27% vs. 46%; P < 0.001 for both) (Table 2). Despite this, patients with cirrhosis less often received potentially curative therapy (i.e., resection, ablation, or liver transplant) and more often received liver-directed therapy with radiation or transarterial therapy (Table 2).

We estimated the effect of screening in two ways. First, we assessed the effect of abdominal imaging (ultrasound or contrast-enhanced computed tomography or magnetic resonance imaging) within the year before HCC diagnosis; data were available in 23% of patients (n = 313; Supporting Table S2). Second, we investigated the scenario in which HCC was diagnosed, this being asymptomatic (implying that HCC was diagnosed by screening) versus symptomatic, including new/worsening hepatic decompensation or abdominal pain at time of diagnosis (Supporting Table S3). Data on diagnosis scenario were available in 51% of patients (n = 706). With both the comparisons of imaging versus no imaging and symptoms versus no symptoms, there were no differences in age, sex, or presence of diabetes or obesity. Patients with recent imaging or with no symptoms had smaller and/or fewer tumors, more frequently had early stage disease based on BCLC stage, and were more likely to receive curative therapy.

Survival

We next turned our attention to factors associated with overall survival. Unadjusted mortality in multiple subgroups are shown in Table 3. Cirrhosis did not affect overall survival (HR, 1.14; 95% CI, 0.94-1.37) (Table 4; Fig. 1A). Other factors associated with poorer survival included greater age, higher MELD score, Child-Pugh score, and BCLC stage; being within Milan criteria, receiving curative therapy, Asian race, and more recent period of diagnosis (2008 and later) were associated with better prognosis (Table 4). History of abdominal imaging within a year of HCC diagnosis was associated with improved prognosis (HR, 0.55; 95% CI, 0.37-0.82), whereas symptoms at HCC diagnosis were associated with poorer outcomes (HR, 2.16; 95% CI, 1.68-2.78) (Table 4; Fig. 1B,C). After lead-time adjustment with an estimated sojourn of 6 months, the association between symptoms and mortality remained significant (HR, 1.66; 95% CI, 1.28-2.13; P < 0.001) while that between history of imaging and mortality did not (HR, 0.70; 95% CI, 0.47-1.03; P = 0.07).

We stratified by cirrhosis status and either receipt of imaging in the year before HCC diagnosis or symptoms. Among patients both with and without cirrhosis, there was an insignificant trend toward improved survival based on screening history (Supporting Fig. S1A,B), whereas presence of symptoms was associated with poorer survival in both groups (Supporting Fig. S2A,B). Patients with symptoms and without cirrhosis had poorer survival than those with symptoms and with cirrhosis (Supporting Figs. S1D, S2D); there was no difference in survival based on cirrhosis status based on having undergone screening/surveillance or among asymptomatic patients.

We constructed a multivariate Cox proportional hazards model for overall survival that included cirrhosis. The model was adjusted for factors associated with overall survival on univariate analysis at P < 0.10. To avoid collinearity with ALBI score and Milan criteria, Child-Pugh and MELD scores, BCLC stage, and treatment type were not included in the multivariable model. On this analysis, cirrhosis was not associated with survival.

When presence of symptoms at HCC diagnosis was added to the model, symptoms were associated with poorer overall survival (HR, 1.82; 95% CI, 1.30-2.53; P < 0.001; Supporting Table S4). There was a trend toward an association between prior imaging and improved prognosis (HR, 0.61; 95% CI, 0.38-1.00; P = 0.051; Supporting Table S5). However, many patients had missing data on both symptoms and prior imaging.

We conducted a sensitivity analysis where we used the nonalcoholic fatty liver disease fibrosis score (NFS) to determine fibrosis based on cutoff >0.675 indicating advanced fibrosis/cirrhosis (fibrosis stage 3-4), <−1.455 indicating no significant fibrosis (stage 0-2), and a score from −1.455 to 0.675 as indeterminate. NFS >0.675 was not associated with increased mortality relative to NFS <−1.455 (HR, 0.84; 95% CI, 0.61-1.15) or relative to NFS ≤0.675 (HR, 1.21; 95% CI, 0.97-1.52).

Machine Learning

Finally, we constructed a random forest machine learning model to identify prognostic factors in an attempt to create a risk score for survival in MAFLD-related HCC. The factors most strongly associated with rank-based survival were treatment type, Child-Pugh score, and maximal tumor size (Supporting Fig. S3). Based on these results, we created a score that allocated up to 3 points based on therapy type, 1 point for maximal tumor size <4.3 cm (median tumor size in the derivation cohort), and 1 point for Child-Pugh score <6 (median score in the derivation cohort) (Fig. 2A). This score predicted three distinct risk groups in both the training and validation cohorts (Fig. 2B,C). Notably, in a Cox proportional hazards model incorporating treatment type, tumor size, and Child-Pugh score, variance inflation factors were low (<2 for all), indicating no meaningful multicollinearity.