The training cohort was a territory-wide retrospective cohort identified using data from the Clinical Data Analysis and Reporting System (CDARS) under the management of Hospital Authority, Hong Kong [19]. Different territory-wide studies have been conducted using CDARS [20-23]. The International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) coding system was used in CDARS. The use of ICD-9-CM codes in CDARS to identify hepatic events has been found 99% accurate when referenced to clinical, laboratory, imaging and endoscopy results from the electronic medical records [24]. The derived HCC risk score was externally validated in an independent cohort of CHB patients from Severance Hospital, Seoul, Korea.

Consecutive adult CHB patients with positive hepatitis B surface antigen for 6 months and/or defined by the ICD-9-CM diagnosis codes between 2000 and 2017 in Hong Kong were identified. Those patients with serum HBV DNA and at least one of the serum fibrosis scores (APRI, FIB-4, and Forns index) measured between January 2000 and June 2018 were included. Follow-up duration were censored at the start date of first antiviral treatment for patients ever received antiviral treatment. HCC surveillance was done every 6 to 12 months by ultrasound and alpha-fetoprotein (AFP) testing under the Hospital Authority. APRI, FIB-4, and Forns index were calculated as described (Supplementary Table 1) [15,25,26]. We excluded the following patients: those co-infected with hepatitis C virus and/or hepatitis D virus based on ICD-9-CM diagnosis codes, viral and serological markers and/or use of antiviral therapy for hepatitis C; co-infected with human immunodeficiency virus based on ICD-9-CM diagnosis codes and/or use of antiviral therapy for human immunodeficiency virus; had other coexisting autoimmune and metabolic liver diseases based on ICD-9-CM diagnosis codes; had cancer including HCC and/or liver transplantation before or within the first 6 months after baseline date; follow-up duration less than 6 months; received rituximab and any chemotherapeutic agents (Supplementary Tables 1, 2). We also excluded patients with ICD-9-CM diagnosis codes of acute hepatitis B and/or positive immunoglobulin M to hepatitis B core antigen unless it was more than 6 months apart from a positive hepatitis B surface antigen result, patients with Child-Pugh class B or C or had any hepatic events before baseline. For validation cohort, CHB patients identified in 2000–2017 with available HBV DNA results in Severance Hospital, Seoul, Korea were included. We excluded patients age younger than 18 years; develop HCC before or within first 6 months after baseline; follow-up less than 6 months; without available serum fibrosis scores at baseline; with Child Pugh B or C and/or had any hepatic events before baseline.

Patients were followed until diagnosis of HCC, death, liver transplantation, starting antiviral treatment, last follow-up date (December 31, 2018), or up to 5 years of follow-up, whichever came first. The study protocol was approved by the Joint Chinese University of Hong Kong - New Territories East Cluster Clinical Research Ethics Committee.

Data of training cohort were retrieved from CDARS in January 2019. Baseline date was set on the first serum HBV DNA assay. Demographic data including date of birth and gender were captured. At baseline, liver and renal biochemistries, hematological and virologic parameters were collected. Thereafter, serial liver and renal biochemistries as well as HBV viral markers were collected until the last follow-up date. We also retrieved data on other relevant diagnoses, procedures, concomitant drugs, laboratory parameters. Drug data of any antiviral treatment for CHB including any approved nucleos(t)ide analogues, namely lamivudine, adefovir dipivoxil, entecavir, telbivudine, tenofovir disoproxil fumarate and tenofovir alafenamide, as well as (pegylated)-interferon were also retrieved during the study period; patients’ followup was censored at the time of first antiviral treatment.

The primary endpoint was HCC. HCC was identified based on diagnosis codes (155.0: HCC and 155.2: carcinoma of liver) or procedure codes for HCC treatment according to ICD-9-CM codes from CDARS. Liver cirrhosis was identified by ICD-9-CM diagnosis codes for cirrhosis and its related complications (Supplementary Table 2). Diabetes mellitus (DM) was defined by exposure to any anti-diabetic agents, and/or hemoglobin A1c ≥6.5% and/or fasting plasma glucose ≥7.0 mmol/L in one measurement, and/or the ICD-9-CM diagnosis codes for DM.

Data were analyzed using SPSS version 25.0 (IBM Corp., Armonk, NY, USA), and R software (4.0.0; R Foundation for Statistical Computing, Vienna, Austria). Continuous variables were expressed in mean±standard deviation or median (interquartile range [IQR]), as appropriate, while categorical variables were presented as frequency (percentage). Time to HCC diagnosis was defined as the time from the baseline date to the date of first HCC diagnosis code, or the date of last follow-up if patients were still alive without HCC at the time of analysis. Kaplan-Meier analysis with log-rank test was used to estimate and compare the cumulative risks of development of HCC in different groups of patients. Cox proportional hazards model was performed to determine the relationship of clinical characteristics of HBV DNA level, gender, age, serum bilirubin level, FIB-4 index and serum hepatitis e antigen (HBeAg) status with the development of HCC. Missing values were excluded from the analysis. A simple HCC risk score was derived by significant variables obtained from backward stepwise multivariable analysis with P<0.05. The score was the weighted sum of those variables of which the weights were defined as the quotient (rounded to nearest integer) of corresponding estimated coefficient from a Cox regression analysis divided by the smallest chi-square coefficient [11].

The performance of serum fibrosis scores and the HCC risk score was assessed by area under the time-dependent receiver operating characteristic curve (AUROC) [27]. The 95% confidence interval (CI) for the comparison of two ROC curves was computed by bootstrap sampling of 1,000 samples. Diagnostic accuracy of the suggested cutoff values was assessed by sensitivity, specificity, positive predictive value, and negative predictive value (NPV) with 95% CI. Youden’s index was used to select the best cutoff value. High sensitivity and NPV were used to further select the cutoff values with comparable Youden’s index to exclude patients with low HCC risk. All statistical tests were two-sided. P<0.05 was taken as statistically significant.

We identified in total 135,395 CHB patients in the public healthcare system in 2000–2017 [28]. There were 60,539 CHB patients with available HBV DNA results at baseline. After excluding patients fulfilling the exclusion criteria, the final number of treatment-naïve CHB patients included in the training cohort was 15,187. We identified 8,559 untreated CHB patients with available HBV DNA measurement in the validation cohort, 4,286 patients were finally analyzed after exclusion (Fig. 1).

During a mean follow-up of 52 months, 180 patients developed HCC in the training cohort. We compared the characteristics of patients developed HCC and without HCC. Patients who developed HCC were older (61.4±11.4 vs. 51.8±13.6 years), and more likely to be male (73.9% vs. 54.8%) than patients without HCC. More patients of non-HCC group were younger than 50 years (41.9%). For liver biochemistries, patients who developed HCC had lower albumin level (38.1±5.6 vs. 41.6±4.9 g/L), platelet count (161.3±68.2 vs. 208.5±68.2×109/L), cholesterol level (4.6±1.2 vs. 4.8±1.1 mmol/L) and higher total bilirubin (14.1 [IQR, 9.2–19.1] vs. 11.0 [IQR, 8.0–15.8] µmol/L), ALT (46.0 [IQR, 31.0– 72.3] vs. 37.0 [IQR, 22.0–73.0] IU/L), gamma-glutaryl transpeptidase (GGT; 70.2 [IQR, 37.0–164.5] vs. 29.0 [IQR, 19.0–58.0] U/L) than patients without HCC. Patients with HCC had cirrhosis (31.1% vs. 5.1%) and higher HBV DNA level (5.0±1.8 vs. 4.4±2.1 log10 IU/mL). Their serum fibrosis scores of APRI (0.7 [IQR, 0.5– 1.2] vs. 0.4 [IQR, 0.3–0.7]), FIB-4 index (2.5 [IQR, 1.6–4.4] vs. 1.3 [IQR, 0.9–2.1]), and Forns index (7.4±2.0 vs. 5.6±2.0) were all higher than those of non-HCC patients (Table 1). The AFP levels in patients developed HCC was significantly higher than patients did not develop HCC (6.1 [IQR, 3.3–11.6] vs. 3.1 [IQR, 2.2–4.9] ng/mL, P<0.001). Patients developed HCC had more proportion of DM than patients did not develop HCC (37.2% vs. 18.6%, P<0.001).

In the validation cohort, 47 patients developed HCC during a mean follow-up of 50 months. Compared with the training cohort, patients in the validation cohort were younger (48.5±12.6 vs. 52.0±13.6) and had more patients with cirrhosis (22.2% vs. 5.4%). Patients had similar albumin level, platelet count, APRI and FIB-4 index in both of the training and validation cohorts (Table 1).

We performed univariate and multivariable analysis for APRI, FIB-4 index separately. Forns index is not generally used in clinical assessment of Hong Kong and there were more than half of missing values in our cohort. So, we did not select Forns index to develop the new score. APRI >2 was not associated with HCC development in both univariate and multivariable analysis (Supplementary Table 3). FIB-4 >3.25, male gender, age older than 50 years old, high total bilirubin (bilirubin >18 µmol/L), low platelet (platelet <150×10⁹/L), high HBV DNA level (HBV DNA >2,000 IU/mL) and DM were correlated with HCC development by univariate analysis [29]. There were 2,155 patients (14.2%) without FIB-4 index and these patients were excluded from the Cox proportional hazards model. Male gender, age older than 50, high HBV DNA level and high FIB-4 could predict HCC development significantly in multivariable analysis (Table 2).

FIB-4 score combined with gender, age and HBV DNA level which were significant in the multivariable analysis comprised the new HCC prediction score named Liang score and ranged from 0 to 22 (Table 3).

We listed out several potential cutoff values and calculated the diagnostic accuracy of these values (Table 4). Among these values, cutoff value of 10 had the highest Youden’s index of 0.406, followed by the cutoff value of 9 with Youden’s index of 0.391. As the corresponding sensitivity and NPV of 9 were 83.7% (95% CI, 73.9–90.5%) and 99.8% (95% CI, 99.7–99.9%), which were higher than those of the cutoff value of 10 (sensitivity, 74.4%; specificity, 99.7%), we finally selected 9 as the cutoff value of the Liang score. By using the cutoff value of 9, patients were divided into two groups: low-risk group (7,191 patients) with Liang score ≤9 and high-risk group (5,841 patients) with Liang score >9. The corresponding number of patients developed HCC were 14 (0.19%) and 72 (1.12%) in the low- and high-risk groups at 5 years’ follow-up duration. The cumulative incidence of HCC was significantly different between low- and high-risk patients by Kaplan-Meier analysis (log-rank test, P<0.001). The annual HCC incidences of low-risk group were less than 0.2%, which is the threshold of offering cost-effective HCC surveillance (Fig. 2A).

We compared the performance of Liang score in patients never receive antiviral treatment and patients received antiviral treatment and censored at the start date of the treatment (Supplementary Table 4). The sensitivity (81.82% vs. 84.38%) and NPV (99.91% vs. 99.61%) were similar between these two groups.

In the validation cohort, there were 2,686 patients (62.8%) classified as low-risk groups, among which three patients (0.11%) developed HCC. The corresponding sensitivity and NPV of cutoff of 9 were 90.0% (95% CI, 72.3–97.4%) and 99.9% (95% CI, 99.6–99.97%), respectively (Table 5). The cumulative incidence of HCC in low-risk patients was significantly lower than that in highrisk patients (log-rank test, P<0.001), and was less than 0.2% per year (Fig. 2B). Liang score performed well in patients with or without cirrhosis and subgroups of patients with different AFP levels (Supplementary Tables 5, 6).

We compared the performance of Liang score with other HCC prediction scores which were developed in treatment-naïve CHB patients: CU-HCC score, GAG-HCC score and REACH-B score in the validation cohort. We selected conventionally used cutoff values of 5, 100 and 8 for CU-HCC score, GAG-HCC score and REACH-B score, respectively [11,30,31]. The sensitivity and NPV results were 83.3% and 99.8% for CU-HCC score; 70.0% and 99.8% for GAG-HCC score and 65.2% and 99.6% for REACH-B score, which were all lower than those of Liang score (90.0% and 99.9%). The number of low-risk patients defined by CU-HCC score, GAG-HCC score and REACH-B score were 2,403, 3,639, and 1,933 with HCC incidence of 0.21%, 0.25%, and 0.41%, respectively (Table 5). Time-dependent AUROC was used to compare the predictive ability of Liang score with other HCC prediction scores. The time-dependent AUROC of Liang score was 0.79 (95% CI, 0.70–0.89) which was not significantly different to those of CU-HCC score (0.82; 95% CI, 0.75–0.89; P=0.434) and GAG-HCC score (0.83; 95% CI, 0.74–0.91; P=0.342). The time-dependent AUROC of Liang score was significantly higher than that of REACH-B score (0.61; 95% CI, 0.53–0.70; P<0.001; Supplementary Fig. 1).

Patients with Liang score of 9 or lower may be exempted from HCC surveillance which include the following groups: female patients younger than 50; female patients age between 50 and 60 with low FIB-4 score or male patients younger than 50 with low FIB-4 score (Table 3). The corresponding annual HCC incidences of these groups were all less than 0.2% in the training cohort.