Graphical Abstract
INTRODUCTION
Hepatocellular carcinoma (HCC) is the sixth most common cancer and the third most common cause of cancerrelated death worldwide [1]. The major risk factors of HCC are chronic hepatitis B (CHB) or chronic hepatitis C (CHC) infection and the presence of cirrhosis. Major clinical guidelines have recommended surveillance of HCC for patients with these risk factors for HCC [2]. The overall goal is to identify HCC at an early stage, provide effective therapy, and reduce disease-related mortality. The 5-year survival rate exceeds 50% among patients with early-stage HCC, but median survival markedly reduces to 1–2 years in those with advanced stages [3]. Several HCC scoring systems have been proposed for predicting the risk of developing HCC. For example, the GAG-HCC (guide with age, gender, HBV DNA, core promoter mutations and cirrhosis) [4], CU-HCC (Chinese University-HCC) [5], REACH-B (risk estimation for hepatocellular carcinoma in chronic hepatitis B) [6], PAGE-B (platelet age gender–HBV) [7] score were developed to predict CHB-related HCC with or without antiviral therapies.
In recent years, the number of non-B and non-C HCCs has been increasing, mostly attributed to metabolic dysfunction-associated steatotic liver disease (MASLD) (formerly known as non-alcoholic fatty liver disease (NAFLD]) and alcohol-related liver disease (ArLD) [8]. No HCC risk predictors exist for MASLD, ArLD, or other non-viral chronic liver diseases (CLD). A meta-analysis pointed out that 39% of patients with NAFLD-related HCC do not have cirrhosis and do not fit the current HCC surveillance guidelines [9]. There is an unmet clinical need to identify at-risk patients with non-viral CLDs and recommend regular HCC surveillance. A risk predictor using simple clinical parameters is preferred for risk stratification during daily practice.
The Steatosis-Associated Fibrosis Estimator (SAFE) score consists of age, body mass index (BMI), diabetes mellitus (DM), platelets, aspartate aminotransferase (AST), alanine aminotransferase (ALT), and globulins to distinguish significant liver fibrosis (≥stage 2, F2) from F0/1 based on patients with biopsy-proven NAFLD [10]. The model yielded higher areas under receiver operating characteristic curves (AUROC) than those of fibrosis-4 (FIB-4) and NAFLD Fibrosis Scores, and had high negative predictive value (88–92%) in ruling out significant fibrosis. The SAFE score ≥100 increased 1.53-fold risk of mortality in a longterm follow-up cohort, and it may be used in primary care to recognize low-risk NAFLD patients [10].
Liver fibrosis is a significant risk for the development of HCC [11], and steatosis is also prevalent in patients with various CLDs. We hypothesized that high SAFE scores are associated with the development of HCC. This study aimed to explore the performance of the SAFE score to predict the development of HCC in patients with CLD of various etiologies, including CHB, CHC, MASLD, ArLD, and other etiologies, to provide HCC surveillance recommendations.
MATERIALS AND METHODS
Cohort establishment and definition of subgroups
We conducted a retrospective cohort study using the Integrated Medical Database of the National Taiwan University Hospital (NTUH-iMD) [12], a tertiary medical center in Taiwan. The NTUH-iMD contains comprehensive medical records from over ten affiliated hospitals, with 3 million patient visits annually. We queried the NTUH-iMD for patients aged ≥20 years with CLDs diagnosed by the ICD9/10 codes (Supplementary Table 1) from January 2006 to April 2021.
CHB or CHC was diagnosed by at least one inpatient or two outpatient diagnosis codes, positive HBsAg or anti-HCV, or the prescription of antiviral therapy. NAFLD or ArLD was diagnosed by at least one inpatient or two outpatient diagnosis codes and excluded patients with CHB or CHC. Other liver cirrhosis (LC) was diagnosed by at least one inpatient or two outpatient diagnosis of LC and its complications or ultrasonographic diagnosis of cirrhosis more than twice, without a diagnosis of CHB, CHC, ArLD, or NAFLD. Other CLD were patients with liver diseases without any of the preceding diagnoses.
MASLD was defined by the recently proposed criteria [13,14]. Patients should have hepatic steatosis by abdominal ultrasonography [15], and one of the five following clinical criteria: (1) BMI ≥23 kg/m2; (2) fasting serum glucose ≥100 mg/dL or HbA1c ≥5.7% or a diagnosis for pre-DM or type 2 DM or specific treatment for type 2 DM; (3) a diagnosis for hypertension or specific antihypertensive drug treatment; (4) plasma triglycerides ≥150 mg/dL or lipid-lowering treatment; (5) plasma high-density lipoprotein-cholesterol ≤40 mg/dL for men or ≤50 mg/dL for women or lipid-lowering treatment [14]. By the definition of MASLD, concurrent HBV infection is included, while HCV is not.
This study was approved by the Institutional Review Board of National Taiwan University Hospital (202108125RINB) and conformed to the ethical principles for medical research involving human subjects of the Declaration of Helsinki and Istanbul. Written informed consent from participants was waived because it was a retrospective review of medical records with de-linked patient information.
Data collection and SAFE score calculation
From the NTUH-iMD, we recorded patients’ birth date, sex, BMI, AST, ALT, platelet, albumin, total protein, and the presence of DM. The baseline SAFE score was (2.97xage)+(5.99xBMI [BMI >40 set to 40])+(62.85xDM [0 if absent, 1 if present])+(154.85xLn [AST, U/L]–(58.23xLn [ALT, U/L])+(195.48xLn [globulins, g/dL])–(141.61xLn [platelets, 109/L])–75, where the globulin was calculated by total protein minus albumin. The SAFE scores were graded as low (<0), intermediate (0–100), and high (≥100) for fibrosis severity stratification (https://medcalculators.stanford.edu/safe) [10].
HCC surveillance and outcome definition
HCC surveillance, including abdominal ultrasound and alpha-fetoprotein (AFP) examinations every 6–12 months, was performed routinely and reimbursed by the National Health Insurance of Taiwan [16]. We identified HCC using the cancer registry database diagnostic codes, which were diagnosed according to the AASLD guidelines [17,18]. Patients were traced from baseline until the development of HCC, death, or the last medical visit. Patients with occurrence of HCC before or within 3 months after the baseline date were excluded because we evaluated the SAFE score as a predictive marker for HCC.
External validations cohorts
To validate the prediction of SAFE score for HCC, we included two external validation cohorts from a hospital and community. From 2000 to 2022, patients with CLD with the same diagnostic criteria as this study were consecutively included in Taichung Veterans General Hospital (VGH), a tertiary medical center in Taiwan. They received standard medical care and were traced for the development of HCC, diagnosed by the Cancer registry.
Another community MASLD cohort of healthy individuals over 30 years who participated in a health screening program of 336,866 participants in Taiwan between 1997 and 2013 was included. MASLD was identified by abdominal ultrasonography and cardiometabolic profiles. Their outcomes were linked to the National Cancer Registry for the newly diagnosed HCC until December 31, 2019 [19].
Statistical analysis
Continuous variables were presented as median (interquartile range), and categorical data as number (percentage). Differences between groups were evaluated using the Student’s t-test or the chi-square statistic. The performance of the SAFE score for classifying the risk of HCC was examined by the time-dependent AUROC, area under the precision-recall curve (AUPRC) and adjusted for competing mortality. The observed outcomes and predicted probabilities of HCC were examined by the calibration plot, and the goodness of fit for the SAFE score was examined by the Brier score, which ranges from 0 (perfect prediction) to 1 (worst prediction). The most favorable SAFE score cut-off was calculated using Youden’s index with a sensitivity greater than 80%. The HCC risks were estimated using the cumulative incidence function with adjustment for death as the competing risk event, and compared by the Gray‘s test or modified log-rank test [20]. The sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were reported. Clinically relevant and significant (P<0.05) variables associated with HCC in univariate analysis were included in the sub-distribution hazard model and adjusted for competing risk of death developed by Fine and Gray [21]. Sub-distribution hazard ratios (SHRs) and 95% confidence intervals (CIs) were reported. All statistical analyses were performed using R (Version 4.1.0, R Foundation for Statistical Computing, Vienna, Austria). All tests were two-sided, and a P-value of <0.05 was considered statistically significant.
RESULTS
From January 2006 to April 2021, we screened the NTUH-iMD yielding 23,742 patients with CLD. We excluded 9,604 patients without BMI or incomplete laboratory data, and 1,175 patients developed HCC before or at baseline. Finally, 12,963 patients with CLD with SAFE scores were included in the study (NTUH cohort) (Supplementary Fig. 1).
The median age was 59, and 51.5% were men (Table 1). The median BMI was 23.7, and 19.7% of subjects had DM. The median SAFE score was 38, and 38.3%, 28.3%, and 33.4% had low (<0), intermediate (0–100), and high SAFE scores (≥100), respectively. The diagnosis breakdown was HBV (n=5,449), HCV (n=1,819), HBV/HCV (n=433), MASLD (n=2,958), NAFLD (n=1,575), ArLD (n=324), other LC (n=433), and other CLD (n=2,930) groups. The median SAFE score was 19, 109, 87, 27, 24, 126, 178, and 19, respectively. The other LC group has the highest SAFE score, supporting that the SAFE score correlates with liver fibrosis.
SAFE score stratifies risks of HCC in subgroups of CLD
After a median follow-up of 4 years, 258 patients (2.0%) developed HCC. Overall, the 1-, 3-, 5-, and 10-year cumulative incidence of HCC was 0.3%, 1.3%, 2.2% and 3.6%, respectively. Regarding the etiologies of CLDs, the 5-year cumulative incidence of HCC was lowest in the other CLD (0.06%) group, followed by NAFLD (0.3%), MASLD (0.6%), HBV (2.3%), HBV/HCV (2.9%), other LC (3.1%), ArLD (4.1%), and HCV (5.4%) (Supplementary Fig. 2).
Using the SAFE score to stratify the risk of HCC, we found that there was a significantly higher yearly risk of HCC in the high (≥100) SAFE score group (1.0%), followed by the intermediate (0–100) (0.2%) and low-risk (<0) groups (0.1%, P<0.001) (Fig. 1A).
The HCC risks correlate with their baseline SAFE scores in individual groups. The SAFE score categories (low, intermediate, high) correlated with the risk of HCC in HBV (P<0.001), HCV (P<0.001), HBV/HCV (P=0.003), MASLD (P=0.015), and NAFLD (P=0.023). There was a trend for ArLD (P=0.093), other LC (P=0.174), and other CLD groups (P=0.227) (Supplementary Fig. 3).
SAFE score independently predicts the development of HCC
The time-dependent AUROC using the SAFE score to classify 1-, 3-, and 5-year HCC risk was 0.84, 0.82, and 0.81, respectively (Supplementary Fig. 4). The 5-year AUROCs were, by and large, similar in the subgroups (0.64–0.93) (Supplementary Table 2). Generally, the AUROCs by the SAFE score are numerically greater than those by the FIB-4 index, which is another fibrosis estimator. Specifically, the AUROC by the SAFE score was significantly greater than those of the FIB-4 index in classifying the MASLD or NAFLD-related HCC (all P<0.05).
Various cut-off values of the SAFE score in classifying HCC risks were calculated (Supplementary Table 3). Numerically, the SAFE score of 129 is associated with the highest Youden index, with a sensitivity of 79% and specificity of 74%. At the same time, the original cut-off of 100 to predict a high probability of liver fibrosis was not too far off, with a sensitivity of 82% and specificity of 68%.
Table 2 summarizes the results of the univariate and multivariable sub-distribution hazard model. Age, sex, albumin, and the SAFE score were associated with the risk of HCC in univariate analysis. After adjustment of these factors, SAFE score ≥100 and 0–100 were associated with an adjusted SHR (aSHR) of 11.4 (95% CI 6.80–19.0; P<0.001) and 2.04 (95% CI 1.11–3.75; P=0.022) respectively, compared with SAFE score <0. There is a 6% increase risk of HCC per 10-scores of SAFE score increase (Supplementary Table 4). After adjusting the CLD etiologies, a high SAFE score of ≥100 and 0–100 had a significantly greater risk of HCC (aSHR 8.44; 95% CI 5.05–14.13; P<0.001 and aSHR 1.89; 95% CI 1.03–3.48; P=0.040) compared with SAFE score <0 (Supplementary Table 5).
A high SAFE score identifies an increased risk of HCC
Given the dramatic increase in the risk of HCC in the group with the highest tier of SAFE, we combined the two lower risk tiers together, namely SAFE ≥100 vs. SAFE <100. The cumulative incidence of HCC in all patients with a high SAFE score was significantly higher than in low-risk patients (1.0% vs. 0.1% per year; P<0.001) (Fig. 1B).
SAFE ≥100 (vs. SAFE <100) significantly increased risks of HCC in HBV (1.2% vs. 0.2% per year; P<0.001), HCV (1.8% vs. 0.2% per year; P<0.001), HBV/HCV (1.2% vs. 0% per year; P<0.001), MASLD (0.3% vs. 0.05% per year; P<0.001), NAFLD (0.2% vs. 0% per year; P=0.006), and ArLD (1.3% vs. 0.2% per year; P=0.034) (Fig. 2). The difference was replicated in the other LC and other CLD group (Supplementary Fig. 5). However, the actual incidence of HCC still varied by diagnosis. For example, only 8 patients (1.1%, 8/750) developed HCC in MASLD patients with high SAFE scores, while 8.6% (81/946) developed HCC in HCV patients with high SAFE scores.
Notably, the SAFE score tiers stratified patients with viral hepatitis regardless of their treatment status. For example, among patients with HBV, the SAFE score tiers correctly classified the annual incidence of HCC among those undergoing antiviral therapy (1.6% vs. 0.2%; P<0.001) and those who remained untreated (1.0% vs. 0.15%; P<0.001). Similarly, the incidence of HCC was 2.4% vs. 0% (P=0.013) among treated HCV patients and 1.7% vs. 0.3% (P<0.001) among untreated patients for high and low SAFE score tiers, respectively (Supplementary Fig. 6). Among 519 HBV patients who received antiviral therapy with undetectable HBV DNA level, a SAFE score ≥100 significantly increased the risk of HCC compared with a SAFE score <100 (P<0.001). In 61 HCV patients who achieved SVR after antiviral therapy, SAFE scores ≥100 increased risks of HCC compared with low SAFE score patients (P=0.179) (Supplementary Fig. 7). In our HBV cohorts, we compared the performance of SAFE scores to some HBV-related HCC risk scores. The SAFE score has comparable AUROC compared with the existing risk calculators in HBV-related HCC, such as the REACH-B, CU-HCC, and PAGEB scores (Supplementary Table 6).
In the multivariable sub-distribution hazard model after adjustment of confounding factors and competing mortality, SAFE score ≥100 was associated with an aSHR of 7.54 (95% CI 5.38–10.60; P<0.001), compared to SAFE score <100 (Table 3). Patients with a high SAFE score (≥100) increased HCC risks in subgroups of HBV (aSHR 4.94; 95% CI 3.24–7.51), HCV (aSHR 8.22; 95% CI 3.61–18.70), HBV/HCV (aHR 10.10; 95% CI 1.37–75), and MASLD (aHR 4.23; 95% CI 1.43–12.50).
In patients with non-viral hepatitis, the overall aSHR was 11.10 (95% CI 3.97–31.30). Specifically, all HCC cases developed in the SAFE score ≥100 patients of the NAFLD subgroup (thus, the SHR cannot be calculated) with an annual incidence of 0.6% in the first year. There is a trend of increased HCC risks in ArLD (aSHR 8.07; 95% CI 0.93–69.90), other LC (aSHR 3.17; 95% CI 0.83–12.10), and other CLD (aSHR 9.67; 95% CI 0.48–195) group but statistically insignificant because of a small number of HCC cases.
We further evaluate the AUPRC, calibration plots, and Brier score on the discrimination, calibration, and goodness of fit of the SAFE score in the prediction of HCC. The 1-, 3-, and 5-year AUPRC are 0.0144, 0.0541, and 0.0704. The predicted 1-, 3-, and 5-year HCC risks align well with the estimated actual HCC risks in the calibration plot (Fig. 3). The 1-, 3-, and 5-year Brier scores are 0.0033, 0.0128, and 0.0210, confirming the model‘s goodness of fit. We also calculate the diagnostic performance of the SAFE score with the cut-off value of 100. The sensitivity, specificity, PPV, NPV, and accuracy for diagnosing HCC were 0.84, 0.67, 0.04, 0.996, and 0.68 at 5 years (Table 4). The 5-year AUROC, specificity, and accuracy of the MASLD, NAFLD and other CLD subgroups were higher than other etiologies. The NPV of a SAFE score of 100 is very high and can identify patients with low risk of HCC.
External validation of SAFE score for HCC prediction
To validate the prediction of SAFE score for HCC, we included two external cohorts from another hospital and communities. The data set from Taichung Veterans General Hospital included 8,103 patients with a median follow-up of 5 years (Supplementary Table 7). Their median SAFE score was 38, and 133 patients developed HCC afterward. A SAFE score ≥100 significantly increases 7.13-fold risk of HCC compared with the low-risk groups (aSHR 7.13; 95% CI 4.45–11.40) (Fig. 4A, Supplementary Table 8). In the subgroup of 2,643 MASLD patients, a SAFE score ≥100 significantly increased 3.18-fold risk of HCC compared with the low-risk groups (aSHR 3.18; 95% CI 1.22–8.27) (Fig. 4B, Supplementary Table 9).
We included healthy individuals who participated in a health screening program to reduce the referral or selection bias of patients with more advanced diseases or lacking globulin levels in hospitals. These participants were not on HCC surveillance and can support the SAFE score in a real-world setting. From the community cohort, 120,166 participants with MASLD were evaluated (Supplementary Table 7). After a median follow-up of 15 years, 787 developed HCC with an annual incidence of 0.15%. Patients with a high SAFE score (≥100) increased HCC risks significantly compared with those with low SAFE scores (<100) (P<0.001) (Fig. 4C).
DISCUSSION
Our study demonstrated that the SAFE score stratifies the risk of HCC in CLD patients regardless of etiologies, including MASLD, HBV, and HCV. A high SAFE score (≥100) is associated with a 7.54-fold increase of HCC risks compared with low-risk groups. A high SAFE score stratifies high HCC risk patients in both viral and non-viral CLD including MASLD. These findings were validated by independent hospital and community cohorts. Our results suggested the SAFE score using simple clinical parameters could serve as a useful universal HCC predictor in patients with CLD, regardless of their etiologies.
The current guidelines recommend HCC surveillance by abdominal ultrasound in high-risk groups (e.g., CHB, cirrhosis) [2,22-24]. Several risk predictors for HCC have been proposed, such as etiology-specific markers, tumor markers, and fibrosis markers. Most risk predictors for CHB-related HCC are etiology-specific, such as viral profiles. For example, HBV DNA had been included in the GAG-HCC [4], CU-HCC [5], and REACH-B scores [6] for treatment-naïve CHB patients. The second risk score type is HCC tumor markers, such as AFP, PIVKA-II, and AFP-L3 or their combination [25]. Serial tumor marker increase (e.g., AFP) further predicted HCC [26]. The third type of predictors is related to liver fibrosis or liver reserves, such as the FIB-4 index [27] and M2BPGi level [28,29]. Theoretically, disease-specific predictors may perform better than predictors using common laboratory values. However, disease-specific markers (e.g., HBV DNA) are not regularly monitored in real-world settings. Some tumor markers, such as AFP or PIVKA-II, have not been endorsed by all clinical guidelines and would not be checked if these patients do not undergo HCC surveillance. An HCC predictor using routine laboratory markers will be easier to calculate in our daily practice and help identify high-risk patients not on the list of HCC surveillance.
The SAFE score uses widely available variables to estimate liver fibrosis in patients diagnosed with NAFLD. It may be embedded into the electrical medical records systems for primary care physicians to recognize low-risk NAFLD. The SAFE score was derived from Caucasians and validated in an Asian population, showing a better performance in diagnosing significant fibrosis and predicting liver-related events [30]. Furthermore, SAFE ≥100 could predict patients with nearly 10% risk of developing severe liver complications in 15 years [30]. Among the components of the SAFE score, age, BMI, DM [31], AST, ALT, and platelet counts had been associated with the development of HCC [32]. Because fibrosis is a significant predictor for HCC, it is not unexpected that the SAFE score identifies people at risk of HCC as potential candidates for HCC surveillance. We found a dose-dependent risk increase in parallel with the SAFE scores. Patients who receive antiviral therapy for viral hepatitis need to have different risk stratification if viral loads are taken into consideration. We adopted the SAFE score and found a similar AUROC in HBV or HCV patients with and without antiviral therapy.
The incidence of HCC in MASLD without cirrhosis is low, and there is a challenge in diagnosing advanced fibrosis [1]. The challenges of HCC surveillance in NAFLD or MASLD include a higher proportion of NAFLD-related HCC without cirrhosis, poor recognition of at-risk patients, lack of consensus for surveillance in non-cirrhotic NAFLD, subpar effectiveness of surveillance tools related to NAFLD phenotype, and preponderant surveillance underuse among NAFLD patients [33]. There is no surveillance consensus in ArLD patients [1]. The current suggestion is to refer patients with high FIB-4 levels for fibrosis evaluation. The SAFE score has recently outperformed other laboratory-based, non-invasive liver fibrosis tests, such as FIB-4 [30].
We demonstrated that the SAFE score could predict HCC independent of the etiology of CLD. The SAFE score may serve as an HCC predictor for MASLD and ArLD, which currently do not have their HCC risk predictors [1]. We also found that the risk of HCC of MASLD is higher than NAFLD, supporting the fact that the new definition of MASLD can identify more patients with adverse outcomes. Another advantage of the SAFE score is that it consists of modifiable parameters that may be adapted during disease management. Patients with a high SAFE score should receive HCC surveillance for risk stratification, whereas those with intermediate and low SAFE scores may opt out of HCC surveillance or extend the interval. Even with the same high SAFE score category, the risk of HCC still varied by different CLD etiologies. Their underlying etiology still impacts the risk of HCC and the SAFE score. Our analysis demonstrated that a SAFE score ≥100 had a sensitivity of 82% to predict HCC in overall CLD patients and was specifically better in the viral hepatitis group.
According to a recent cost-effective analysis, an HCC incidence >0.4% per year and surveillance adherence >19.5% biannually were necessary for using ultrasound with AFP to be cost-effective compared with no surveillance [33]. According to our data, a SAFE score ≥100 could suggest that patients in HBV, HCV, HBV/HCV, MASLD, ArLD, and other LC groups should receive HCC surveillance.
There are several limitations in this study. First, this is a retrospective cohort study; patients might have missed data for calculating the SAFE score, especially the globulin level. The indication bias for specific patients with complete data may be unavoidable. The diagnosis of steatosis relied on abdominal ultrasound, and we did not use a controlled attenuation parameter or magnetic resonance imaging proton density fat fraction data. The higher HCC incidence may be due to a hospital-based cohort; therefore, we excluded HCC developed within 3 months after calculating the SAFE score to explore the predictability of HCC development. Not all patients with CLD undergo regular HCC surveillance, so the development of HCC may be underestimated. We use two external validation cohorts from hospital and community to support our findings. Because the SAFE score is composed of modifiable parameters, we expected that the dynamic of SAFE scores may further stratify the risk of HCC. Further prospective studies with more longitudinal data should be included to validate our preliminary findings.
In summary, the SAFE score can be used to stratify the risk of HCC, and patients with a high SAFE score (≥100) with a substantial risk of HCC development are recommended to receive surveillance of HCC.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Supplement1
Print
