Phase 3 Validation of PAaM for Hepatocellular Carcinoma Risk Stratification in Cirrhosis

Abstract

Background & Aims

Hepatocellular carcinoma (HCC) risk stratification is an urgent unmet need for cost-effective HCC screening and early detection in patients with cirrhosis to improve poor HCC prognosis.

Methods

Molecular (prognostic liver secretome signature with α-fetoprotein) and clinical (aMAP [age, male sex, albumin-bilirubin, and platelets] score) variable–based scores were integrated into PAaM (prognostic liver secretome signature with α-fetoprotein plus age, male sex, albumin-bilirubin, and platelets), which was subsequently validated in 2 phase 3 biomarker validation studies: the state-wide Texas HCC Consortium and nationwide HCC Early Detection Strategy prospective cohorts, following the prospective specimen collection, retrospective blinded evaluation design. The associations between baseline PAaM and incident HCC were assessed using Fine-Gray regression, with overall death and liver transplantation as competing events.

Results

Of 2156 patients with cirrhosis in the Texas HCC Consortium, PAaM identified 404 (19%) high-risk, 903 (42%) intermediate-risk, and 849 (39%) low-risk patients with annual HCC incidence rates of 5.3%, 2.7%, and 0.6%, respectively. Compared with low-risk patients, high- and intermediate-risk groups had sub-distribution hazard ratios for incident HCC of 7.51 (95% CI, 4.42–12.8) and 4.20 (95% CI, 2.52–7.01), respectively. Of 1328 patients with cirrhosis in the HCC early detection strategy, PAaM identified 201 high-risk (15%), 540 intermediate-risk (41%), and 587 low-risk (44%) patients, with annual HCC incidence rates of 6.2%, 1.8%, and 0.8%, respectively. High- and intermediate-risk groups were associated with sub-distribution hazard ratios for incident HCC of 6.54 (95% CI, 3.85–11.1) and 1.77 (95% CI, 1.02–3.08), respectively. Sub-group analysis showed robust risk stratification across HCC etiologies, including metabolic dysfunction–associated steatotic liver disease and cured hepatitis C infection.

Conclusions

PAaM enables accurate HCC risk stratification in patients with cirrhosis from contemporary etiologies.

Cirrhosis, the terminal stage of chronic liver disease, most commonly from viral and metabolic etiologies, affects 1%–2% of the global population and leads to 1.32 million deaths annually.^1,2 The prevalence of chronic liver disease has increased by 15% over the past decade in parallel with the epidemic of obesity and associated metabolic disorders.^1,2 Hepatocellular carcinoma (HCC) is a major complication of cirrhosis and the fastest-rising cause of cancer-related death in the United States due to frequent late diagnosis at advanced stages when curative therapies are no longer applicable.³ To improve patient survival, professional societies recommend semi-annual ultrasound-based HCC screening in all patients with cirrhosis for early tumor detection and receipt of curative treatment.^4,5 However, this “one-size-fits-all” approach to HCC screening has been ineffective and infeasible due to the increasing number of patients with cirrhosis from metabolic dysfunction–associated steatotic liver disease (MASLD), alcohol-associated liver disease (ALD), and cured hepatitis C virus (HCV) infection.^1,6 The vast size of the target population and lower associated HCC incidence rates have decreased the cost-effectiveness of HCC screening in all patients with cirrhosis.^7,8 Thus, tools to identify a subset of patients with cirrhosis with elevated HCC risk are needed to enable cost-effective HCC screening.⁶ Furthermore, identifying low-risk patients in whom screening is unnecessary can minimize risks of screening-related physical, financial, and psychological harms.^9–11 A prior Markov model–based simulation showed that risk-stratified HCC screening, that is, altering screening intensity according to predicted HCC risk, would be cost-effective when an HCC risk-stratification biomarker distinguishes at least a 2-fold difference (ie, hazard ratio [HR], >2) in HCC incidence between patient groups.¹² An accurate HCC risk-stratification method would enable individualized and cost-effective early HCC detection in patients with cirrhosis.

We previously developed an 8-protein prognostic liver secretome signature (PLSec), derived from a thoroughly validated hepatic transcriptome signature,^13–16 for long-term HCC risk prediction together with α-fetoprotein (PLSec-AFP).¹⁷ PLSec-AFP was successfully validated in independent cirrhosis cohorts, achieving the threshold of a 2-fold difference in HCC incidence.¹⁷ However, there was room for improvement in identifying low-risk patients with cirrhosis with negligible HCC risk who could be spared from potential harms of semi-annual HCC screening.^9–11 Existing clinical variable–based HCC risk scores identify only a small proportion of low-risk patients in cirrhosis cohorts, ranging from 3% to 10%.^18–22 In addition, our previously reported PLSecAFP addressed these issues, but annual HCC incidence rate in the low-risk patients (1.5%) was still suboptimal. Given independent prognostic capability of PLSec-AFP from clinical variables and its molecular nature covering a variety of hepatic cells driving hepatocarcinogenesis,¹⁷ we hypothesized that integration of PLSec-AFP with a clinical variable^–based score might overcome the remaining challenge. Thus, we aimed to develop an integrative molecular/clinical HCC risk indicator for further refinement of risk stratification, especially identifying low-risk patients with negligible HCC incidence, and validate the score in 2 independent prospective cohorts of patients with cirrhosis from multicenter consortia.

Materials and Methods

Study Design

Three externally validated and etiology-agnostic clinical variable–based HCC risk scores, aMAP (age, male sex, albumin-bilirubin, and platelets) score,²³ Toronto HCC Risk Index (Supplementary Table 1),²⁴ and ADRESS (age, diabetes, race, etiology of cirrhosis, sex, and severity) HCC score,²⁵ were compared for HCC risk-predictive performance, and the best performing score was integrated with PLSec-AFP in a single-center prospective cohort of patients with cirrhosis (integrative score derivation cohort) (Figure 1A). As recommended by the Early Detection Research Network paradigm for biomarker validation, we then conducted a phase 3 biomarker validation study, following the PRoBE (prospective specimen collection, retrospective blinded evaluation) design (Supplementary Table 2).^6,7,26 Specifically, the integrative score was externally validated in 2 independent prospective cohorts: the statewide Texas HCC Consortium (THCCC) and nationwide HCC Early Detection Strategy (HEDS) cohorts. The study was approved by the Institutional Review Board for the analysis of deidentified samples from the participating institutions (STU 062018-058, STU 052018-034, and STU 082017-013). Exclusion of patients due to missing data did not affect clinical demographic characteristics in all study cohorts (Supplementary Table 3).

Figure 1. Derivation of PAaM score.

(A) Study design. (B) The association of PAaM classification with incident HCC in the integrative score derivation cohort.

Patient Cohorts

Integrative score derivation cohort

Patients with cirrhosis from mixed etiologies were prospectively and consecutively enrolled at University of Michigan Hospital between January 2004 and September 2006, and regularly followed up using ultrasound with AFP every 6 months for incident HCC until June 2020.^17,27

Texas Hepatocellular Carcinoma Consortium validation cohort

THCCC is a statewide prospective multicenter cohort supported by Cancer Prevention and Research Institute of Texas.²⁸ Patients with cirrhosis were prospectively enrolled at 7 Texas tertiary centers (Michael E. DeBakey Veterans Affairs Medical Center and Baylor St Luke’s Medical Center in Houston; University of Texas Southwestern, Parkland Health, and Baylor Scott & White Research Institute in Dallas; Doctor’s Hospital at Renaissance in McAllen; and Texas Liver Institute in San Antonio) between December 2016 and September 2022, and followed semi-annually until June 2023.

Hepatocellular Carcinoma Early Detection Strategy validation cohort

The HEDS cohort is a nationwide prospective multicenter cohort supported by the National Cancer Institute Early Detection Research Network.²⁹ Patients with cirrhosis were prospectively enrolled at 7 US tertiary centers (University of Texas Southwestern Medical Center, Stanford University, St Louis University, Mayo Clinic, Mount Sinai Hospital, University of Michigan, and University of Pennsylvania) between April 2013 and June 2021 and followed semi-annually until September 2023.

Specific details of the inclusion and exclusion criteria, clinical follow-up, and HCC diagnosis for each cohort are summarized in the Supplementary Materials. Briefly, cirrhosis was diagnosed on the basis of histologic and/or clinical features. For all cohorts, clinical data and serum samples were collected and aliquoted at enrollment and stored at −80°C until use. All patients were followed semi-annually per usual care using ultrasound-based screening during the observation period. HCC was diagnosed based on histologic or imaging-based examinations according to the American Association for the Study of Liver Disease practice guidance.⁴ There were no overlapping patients between the cohorts.

Prognostic Liver Secretome Signature With α-Fetoprotein Assessment

The PLSec assay, including vascular cell adhesion molecule 1, insulin-like growth factor-binding protein 7, gp130, matrilysin, interleukin-6, and C-C motif chemokine ligand 21 as high-risk proteins, and angiogenin and protein S as low-risk proteins, was performed at the University of Texas South-western Medical Center BioCenter, as reported previously¹⁷ (Supplementary Table 4; see Supplementary Materials for details). PLSec-AFP was calculated as 0.175 × PLSec + 0.325 × log₂ (1 + AFP [ng/mL]), and a high-risk prediction was made based on the predefined cutoff of ≥1.66.¹⁷ We previously reported high within- and inter-batch reproducibility of this assay (r² = 0.9997 and 0.971, respectively).¹⁷ All assays were performed in a blinded fashion.

Statistical Analysis

Time to HCC development was defined as the interval between the dates of serum collection and HCC diagnosis, liver transplantation, overall death, or last follow-up. Cumulative HCC incidence rates at 3 and 5 years and the sub-distribution hazard ratio (sHR) for incident HCC were assessed by accounting for liver transplantation and overall death before incident HCC as competing events using Aalen-Johansen’s³⁰ and Fine’s methods,³¹ respectively. The annual HCC incidence rates were calculated using the declining exponential approximation of life expectancy method³² based on the cumulative 3-year incidence rate. HCC incidence was compared using Gray’s method.³³ Performance of the integrative score was regarded as clinically meaningful when HR was >2 for association of the integrative score with time to incident HCC (for cost-effective risk-stratified HCC screening) and annual HCC incidence rate in patients with low-risk prediction was <1% to enable cost-effective risk-stratified HCC screening.^4,12

HCC risk association of 3 clinical variable-based scores, aMAP score, Toronto HCC Risk Index, and ADRESS-HCC, was compared using Royston’s D-index,^34,35 time-dependent area under receiver operating characteristics curve, Harrell’s c-index, and calibration slope and intercept at 3 years in the integrative score derivation cohort (see Supplementary Materials for details of the scores). The best-performing clinical score was integrated with PLSec-AFP by using regression coefficients from their multi-variable Fine-Gray regression modeling for incident HCC, accounting for liver transplantation and overall death as competing risks. In the integrative score derivation cohort, the cutoff values for the risk stratification of the patients were defined to ensure an annual HCC incidence rate of <1% in the low-risk group and to maximize the annual HCC incidence and achieve an HR of >2 (compared with the low-risk group) in the high-risk group (see Supplementary Materials for details).

For validation studies, the estimated sample size to detect an HR of 2 as statistically significant was 1245 under the assumption that ≥25% of the patients show a high-risk score and ≥7% of the patients develop HCC (expected incidence rate over 5 years based on an annual HCC incidence rate of 1.5%, the traditionally used threshold to trigger semi-annual HCC screening) at a statistical power of 0.8 and an α error of .05. Predefined subgroup analyses for sex, age, race and ethnicity, and liver disease etiology were performed to explore the clinical characteristics that influence the HCC risk association for the integrative score. Multiple hypothesis testing was adjusted using Bonferroni correction. Statistical analyses were performed using R software, version 4.2.1 (www.r-project.org).

Results

Derivation of PAaM Score

We first compared 3 externally validated clinical variable–based HCC risk scores for their association with incident HCC in the integrative score derivation cohort of 327 patients with cirrhosis, among which 44 patients developed HCC during a median follow-up time of 3.5 (interquartile range [IQR], 1.7–8.4) years (Table 1). The aMAP score showed a superior HCC risk association compared with the other 2 scores; aMAP also demonstrated a complementary prognostic association with PLSec-AFP (Supplementary Figure 1, Table 2, and Supplementary Table 5). Based on these results, we integrated the molecular and clinical scores into the PLSec-AFP and aMAP (PAaM) score by using regression coefficients from multi-variable Fine-Gray regression modeling (see Supplementary Table 5 and Supplementary Materials):

$$\text{PAaM}\text{score}=0.941\times \text{PLSec}-\text{AFP}+0.052\times \text{aMAP}\text{score}$$

Table 1. Patient Characteristics of Derivation and Validation Cohorts.

Variable	Integrative score derivation cohort (n = 327)	THCCC cohort (n = 2156)	HEDS cohort (n = 1328)
Age, y, median (IQR)	52 (47–57)	61 (54–67)	60 (54–66)
Male sex, n (%)	193 (59)	1345 (62)	703 (53)
Race and ethnicity, n (%)
Black or African American	8 (2)	344 (16)	86 (7)
Hispanic or Latino	8 (2)	511 (24)	126 (10)
Non-Hispanic White	308 (94)	1125 (52)	1065 (81)
Other	3 (1)	176 (8)	45 (3)
BMI, kg/m2, median (IQR)	28.4 (25.0–33.9)	30.1 (26.2–34.6)	30.2 (26.4–35.2)
Obesity,a n (%)	139 (43)	1079 (51)	681 (51)
Etiology, n (%)
HCV	159 (49)	810 (38)	544 (41)
Active infection	152 (46)	317 (15)	297 (22)
Cured infection	7 (2)	489 (23)	197 (15)
Hepatitis B virus	12 (4)	53 (2)	27 (2)
MASLD	20 (6)	559 (26)	292 (22)
Alcohol-associated	60 (18)	445 (21)	207 (16)
Other	76 (23)	289 (17)	258 (19)
Aspartate aminotransferase, IU/L, median (IQR)	60 (42–92)	38 (27–56)	41 (30–60)
Alanine aminotransferase, IU/L, median (IQR)	48 (34–78)	30 (22–44)	33 (23–49)
Albumin, g/dL, median (IQR)	3.4 (2.9–3.8)	3.8 (3.4–4.2)	3.8 (3.4–4.2)
Total bilirubin, mg/dL, median (IQR)	1.2 (0.8–1.9)	0.9 (0.6–1.6)	0.9 (0.6–1.5)
Platelet count, ×103/mL, median (IQR)	95 (68–138)	123 (83–174)	110 (76–160)
Child-Pugh class, A/B/C, n (%)	121/178/25 (37/55/8)	1340/671/97 (63/32/5)	974/353/0 (73/27/0)
AFP, ng/mL, median (IQR)	3.9 (2.3–7.9)	4.0 (2.7–6.0)	3.8 (2.5–6.1)

BMI, body mass index.

^aObesity was defined as BMI ≥30 kg/m², according to World Health Organization criteria.³⁶

Table 2. Model performance of PAaM, PLSec-AFP, and Clinical Hepatocellular Carcinoma Risk Scores in the Integrative Score Derivation Cohort.

Scores	Discrimination					Calibrationa
	Royston’s D-indexb (95% CI)	Time-dependent AUROC at 3 y (95% CI)	Time-dependent AUROC at 5 y (95% CI)	Harrell’s c-indexc (95% CI)		Slope	Intercept, %
PAaM	1.01 (0.55 to 1.46)	0.74 (0.64 to 0.84)	0.73 (0.62 to 0.84)	0.74 (0.61 to 0.80)		1.02	–0.2
PAaM (internal validation)d	1.00 (0.61 to 1.39)	0.74 (0.65 to 0.83)	0.74 (0.63 to 0.82)	0.72 (0.61 to 0.80)		0.94 (0.42 to 1.56)	0.6 (–4.8 to 5.8
PLSec-AFP	0.90 (0.41 to 1.39)	0.69 (0.58 to 0.81)	0.66 (0.54 to 0.78)	0.74 (0.59 to 0.80)		1.19	–1.6
aMAP score	0.68 (0.18 to 1.18)	0.67 (0.54 to 0.79)	0.68 (0.56 to 0.79)	0.61 (0.45 to 0.73)		1.34	–3.8
THCC Risk Index	0.29 (–0.12 to 0.70)	0.58 (0.46 to 0.69)	0.63 (0.52 to 0.73)	0.56 (0.42 to 0.69)		1.33	–2.4
ADRESS-HCC	0.58 (–0.03 to 1.19)	0.61 (0.48 to 0.75)	0.62 (0.49 to 0.75)	0.56 (0.39 to 0.70)		1.79	–8.3

AUROC, area under receiver operating characteristic curve.

^aThe ideal calibration slope and intercept are 1 and 0%, respectively.

^bHigher values for the D-index indicate greater discrimination, where an increase of 0.1 over other risk scores is a good indicator of improved prognostic separation.³⁷

^c95% CIs were calculated using 1000-time bootstrapping.

^dInternal validation was performed using 1000-time bootstrapping.

The performance of the score was internally validated using a bootstrapping procedure (Table 2).

Over the range of the PAaM score in the derivation cohort, the annual HCC incidence rate was <1% at the 25^th percentile cutoff (<4.318) to define the low-risk group and plateaued at approximately 6% at the 75^th percentile cutoff (≥5.072) to define the high-risk group (Supplementary Figure 2). In the meanwhile, only 8%, 9%, and 0.3% of the patients were classified into low-risk group by aMAP, Toronto HCC Risk Index, and ADRESS-HCC scores, respectively, consistent with previous reports.^18–22 Defining the remaining patients as the intermediate-risk group, the annual HCC incidence rates were 5.9%, 3.1%, and 0.5% in the high-, intermediate-, and low-risk groups, respectively (Figure 1B). The high- and intermediate-risk groups showed higher HCC incidence rates with sHRs of 4.79 (95% CI, 1.81–12.7) and 2.77 (95% CI, 1.08–7.13), respectively, compared with the low-risk group. The PAaM score showed consistently superior D-index, time-dependent area under receiver operating characteristics curve, c-index, and calibration compared with the PLSec-AFP and aMAP score alone (Table 2). The scores and cut-offs were subsequently tested in independent validation cohorts with no modifications.

Validation of PAaM Score in Texas Hepatocellular Carcinoma Consortium Cohort

The THCCC cohort included 2156 evaluable patients with cirrhosis, among which 148 patients developed HCC and 316 died or underwent liver transplantation before HCC development during a median follow-up time of 2.3 years (IQR, 1.1–4.0 years). The annual incidence rate of HCC in the entire cohort was 2.3%. Median age was 61 years (IQR, 54–67 years) and 1345 patients (62%) were male. The most prevalent etiologies were MASLD (26%), cured HCV (23%), and ALD (21%). Most patients (95%) had Child-Pugh class A or B cirrhosis (Table 1).

The PAaM score classified 404 (19%), 903 (42%), and 849 (39%) patients as having high, intermediate, and low HCC risks, respectively (Figure 2A). The annual HCC incidence rates in high-, intermediate-, and low-risk patients were 5.3%, 2.7%, and 0.6%, respectively (P < .001) (Figure 2B). The cumulative HCC incidence rates in the high-, intermediate-, and low-risk groups were 14.7%, 7.9%, and 1.7% at 3 years and 20.6%, 14.2%, and 3.7% at 5 years, respectively. Compared with low-risk patients, high- and intermediate-risk PAaM were significantly associated with incident HCC with sHRs of 7.51 (95% CI, 4.42–12.8) and 4.20 (95% CI, 2.52–7.01), respectively. PAaM showed consistently superior model performance metrics compared with PLSec-AFP and aMAP score alone (Table 3).

Figure 2. Associations of PAaM with incident HCC development in THCCC cohort.

(A) Pattern of PLSec protein abundance and distribution of clinical variables in THCCC cohort. (B) Association of PAaM with time to HCC development. CCL-21, C-C motif chemokine 21; HBV, hepatitis B virus; IGFBP-7, insulin-like growth factor-binding protein 7; IL-6, interleukin 6; MMP-7, matrilysin; VCAM1, vascular cell adhesion molecule 1.

Table 3. Comparisons of PAaM, PLSec-AFP, and aMAP in Texas Hepatocellular Carcinoma Consortium and Hepatocellular Carcinoma Early Detection Strategy Cohorts.

Cohort	HCC risk score	Risk prediction	n (%)	Annual HCC incidence rate, %	Prognostic association sHR (95% CI)	Discrimination					Calibrationa
						Royston’s D-index (95% CI)b	Time-dependent AUROC at 3 y (95% CI)	Time-dependent AUROC at 5 y (95% CI)	Harrell’s c-index (95% CI)c		Slope	Intercept, %
THCCC (n = 2156)	PAaM	High risk	404 (19)	5.3	7.51 (4.42–12.8)	1.04 (0.82–1.25)	0.74 (0.70–0.78)	0.74 (0.68–0.79)	0.67 (0.63–0.71)		0.89	0.3
		Intermediate risk	903 (42)	2.7	4.20 (2.52–7.01)	—	—	—	—		—	—
		Low risk	849 (39)	0.6	Reference	—	—	—	—		—	—
	PLSec-AFP	High risk	577 (27)	3.6	2.33 (1.69–3.22)	0.67 (0.4–0.90)	0.63 (0.58–0.67)	0.62 (0.57–0.67)	0.58 (0.54–0.62)		0.78	–0.1
		Low risk	1579 (73)	1.3	Reference	—	—	—	—		—	—
	aMAP	High risk	1316 (61)	3.3	6.77 (1.68–27.4)	0.87 (0.66–1.08)	0.65 (0.62–0.69)	0.65 (0.61–0.70)	0.63 (0.59–0.65)		1.26	–4.3
		Intermediate risk	682 (32)	0.8	1.79 (0.42–7.69)	—	—	—	—		—	—
		Low risk	158 (7)	0.2	Reference	—	—	—	—		—	—
HEDS (n = 1328)	PAaM	High risk	201 (15)	6.2	6.54 (3.85–11.1)	1.30 (1.05–1.56)	0.71 (0.64–0.78)	0.66 (0.59–0.73)	0.66 (0.60–0.73)		1.00	–0.7
		Intermediate risk	540 (41)	1.8	1.77 (1.02–3.08)	—	—	—	—		—	—
		Low risk	587 (44)	0.8	Reference	—	—	—	—		—	—
	PLSec-AFP	High risk	283 (21)	3.9	2.64 (1.74–3.99)	0.89 (0.65–1.13)	0.61 (0.55–0.68)	0.60 (0.54–0.65)	0.59 (0.54–0.65)		0.84	0.0
		Low risk	1045 (79)	1.4	Reference	—	—	—	—		—	—
	aMAP	High risk	802 (60)	2.9	2.02 (0.74–5.49)	1.19 (0.80–1.58)	0.63 (0.58–0.69)	0.61 (0.55–0.66)	0.61 (0.56–0.66)		0.64	0.6
		Intermediate risk	438 (33)	0.4	0.45 (0.14–1.42)	—	—	—	—		—	—
		Low risk	88 (7)	1.8	Reference	—	—	—	—		—	—

AUROC, area under receiver operating characteristic curve.

^aCalibrations were calculated using linear regression between predicted HCC incidence rates of each risk group in the integrative score derivation cohort and observed incidence rates of the corresponding risk group in the validation cohorts at 3 years. Ideal calibration slope and intercept are 1 and 0%, respectively.

^bHigher values for the D-index indicate greater discrimination, where an increase of 0.1 over other risk scores is a good indicator of improved prognostic separation.³⁷

^c95% CIs were calculated using 1000-time bootstrapping.

Validation of PAaM Score in Hepatocellular Carcinoma Early Detection Strategy Cohort

HEDS cohort included 1328 patients with cirrhosis, among which 93 patients developed HCC, and 247 patients died or underwent liver transplantation before HCC development during a median follow-up time of 2.8 years (IQR, 1.1–5.2 years). The annual HCC incidence rate in the entire cohort was 2.3%. The median age of the cohort was 60 years (IQR, 54–66 years) and 703 (53%) patients were male. The most prevalent etiologies were active HCV infection (22%) and MASLD (22%). All patients had Child-Pugh class A or B cirrhosis (Table 1).

PAaM score classified 201 (15%), 540 (41%), and 587 (44%) patients as having high, intermediate, and low HCC risk, respectively (Figure 3A). The annual HCC incidence rates in the high-, intermediate-, and low-risk patients were 6.2%, 1.8%, and 0.8%, respectively (P < .001) (Figure 3B). The cumulative HCC incidence rates in the high-, intermediate-, and low-risk groups were 16.9%, 5.3%, and 2.3% at 3 years and 22.9%, 7.1%, and 4.7% at 5 years, respectively. Compared with low-risk patients, high- and intermediate-risk PAaM scores were significantly associated with incident HCC with sHRs of 6.54 (95% CI, 3.85–11.1) and 1.77 (95% CI, 1.02–3.08), respectively. PAaM showed consistently superior model performance metrics compared with PLSec-AFP and aMAP score alone (Table 3).

Figure 3. Associations of PAaM with incident HCC development in HEDS cohort.

(A) Pattern of PLSec protein abundance and distribution of clinical variables in HEDS cohort. (B) Association of PAaM with time to HCC development. CCL-21, C-C motif chemokine 21; IGFBP-7, insulin-like growth factor-binding protein 7; IL-6, interleukin-6; MMP-7, matrilysin; VCAM1, vascular cell adhesion protein 1.

Subgroup Analyses in Texas Hepatocellular Carcinoma Consortium and Hepatocellular Carcinoma Early Detection Strategy Cohorts

In predefined subgroup analyses of sex, age, race and ethnicity, and liver disease etiology, the HCC risk association of PAaM score achieved sHRs of >2 in all subgroups, with sHRs ranging from 5 to 10 in both validation cohorts (Figure 4). In the subgroup analyses, higher sHRs of high-risk PAaM (vs low-risk) were observed in the THCCC and HEDS cohorts for subgroups of patients defined by younger age (sHRs, 10.1 and 9.54, respectively), Hispanic ethnicity (sHRs, 14.9 and 12.2, respectively), and ALD (sHRs, 10.4 and 10.5, respectively), whereas lower sHRs were observed in Black patients (sHRs, 5.27 and 3.78, respectively).

Figure 4. Association of high-risk PAaM with incident HCC in various subgroups.

Discussion

Available resources to effectively implement semi-annual HCC screening have become increasingly challenged with the growing at-risk population, particularly patients with MASLD or cured HCV, who have relatively lower but still clinically significant risk of HCC.¹ To address this problem, HCC risk stratification in the emerging at-risk liver disease population is urgently needed. Multiple clinical variable–based scores and biomarkers have been reported as methods for HCC risk stratification, although their performance is generally limited in validation studies.⁶ In addition, certain types of biomarkers may not have clinical utility as risk stratification biomarkers due to limited performance by their nature. For example, single nucleotide polymorphisms have been extensively explored as potential HCC risk indicators, but a recent nationwide biobank study suggested that single nucleotide polymorphisms, alone or in combination, showed saturated HCC risk-predictive performance, adding limited information to readily available clinical variable–based scores.³⁸ Thus, there is no HCC risk-stratification biomarker that has advanced to the late phases of cancer biomarker development^6,7 (Supplementary Table 2). To our knowledge, this study represents the largest and first phase 3 biomarker validation study that establishes an integrative molecular/clinical score for HCC risk stratification.

PAaM score achieved a benchmark of HR >2 to enable cost-effective risk-stratified HCC screening in both multi-center validation cohorts, with an anticipated reimbursed cost of PLSec test at $796 based on existing multi-analyte molecular tests in our previously published Markov model–based simulation analysis.¹² There were 15% to 20% of patients in each cohort who were identified as high risk, with annual HCC incidence rates exceeding 5%, a threshold at which magnetic resonance imaging–based screening was previously found to be cost-effective.^39,40 Use of magnetic resonance imaging in high-risk patients would improve early detection compared with ultrasound-based screening, particularly among obese individuals and those with nonviral liver disease in whom ultrasound has particularly poor visualization and sensitivity for early-stage HCC.^41–43 Conversely, PAaM also identified a sizable proportion of low-risk patients, accounting for approximately 40% of the cohorts, and aMAP score identified only 7% of low-risk patients in both of the validation cohorts (Table 3). Thus, PAaM resolved the critical issues for the clinical variable–based scores (ie, infrequent low-risk patients) and PLSec-AFP (ie, non-negligible HCC incidence in low-risk patients) via independent validation in 2 prospective statewide and nationwide multicenter cohorts. Patients with low-risk PAaM score may be spared from regular HCC screening, given the low annual HCC incidence rates below the threshold to justify the semi-annual ultrasound as a cost-effective intervention. Alternatively, less intensive screening may be considered for the low-risk patients by using more accessible modalities (eg, blood test–only screening without ultrasound) less frequently (eg, annual, instead of semi-annual, testing).⁶ Intermediate-risk patients may undergo recommended or less intensive screening as suggested in our prior Markov model–based cost-effectiveness analysis.¹² Such risk-stratification methods will help target HCC screening to those in most need and mitigate screening-related harms in those who are less likely to benefit. Clinical significance of such method cannot be overemphasized, given that <25% of patients with cirrhosis currently receive the guideline-recommended semi-annual HCC screening.^44,45 Despite the variations in clinical demographic characteristics between the cohorts, we observed robust association of PAaM with HCC risk, as shown by the sHRs that stably exceed the benchmark of >2 in the subgroup analysis (Figure 4), supporting its general applicability to heterogeneous global cirrhosis patient populations. This is important, given the global heterogeneity in implementation of HCV screening and treatment, prevalence of HBV infection, alcohol use and policy, access to emerging costly therapies for metabolic disorders, among several others.

Subgroup analysis showed that PAaM met the benchmark of HR >2 across all clinical subgroups evaluated, while the magnitude of HCC risk associations varied by HCC etiology and patient race and ethnicity. This finding suggests that HCC risk assessment may be further refined according to etiology and/or race and ethnicity. Our previous proof-of-concept study in patients with MASLD showed that the addition of etiology-specific “plug-in” biomarker may serve as an approach to address the issue.⁴⁶ Indeed, this “plug-in” strategy enabled identification of low-risk patients who were free from incident HCC over 15 years of longitudinal follow-up, and in whom HCC screening appears to be of low value.

It is noteworthy that the HCC risk level measured by PLSec, a main component of PAaM, is therapeutically modifiable, in contrast to other nonmodifiable variables (eg, age, sex, and liver disease etiology) often included in clinical variable–based scores. Clinically, pharmacologic HCV clearance approximately halves future HCC incidence.^47,48 PLSec captures dynamic changes in HCC risk level after the anti-HCV therapies, which are associated with subsequent HCC development.¹⁷ The signature can also be modulated in various in vitro and in vivo experimental systems.⁴⁹ These data suggest that PLSec can be used as a companion biomarker for medical interventions to reduce HCC risk, that is, HCC chemoprevention, including in patients with nonviral etiologies, such as ALD and MASLD. For instance, lipophilic statins, such as atorvastatin and simvastatin, could be potential candidates, given their association with lower PLSec-based HCC risk levels and lower HCC incidence in patients with cirrhosis from viral and metabolic etiologies.^50,51 Phase II HCC chemoprevention trials of atorvastatin (ClinicalTrials.gov Number, NCT05028829) and epigallocatechin gallate (ClinicalTrials.gov Number, NCT06015022) are currently ongoing with PLSec modulation as the primary end point. The use of PLSec-based models may be more feasible for capturing longitudinal HCC risk than for capturing granular data on variations in clinical risk factors, such as weight, glucose control, and medication exposure, particularly considering over-the-counter agents that are poorly recorded in electronic health records.

We acknowledge several limitations of the study. First, the number of HBV-infected patients was limited (n = 80 total); therefore, future evaluation is needed for this etiology group on the current antiviral therapies. Second, associations between longitudinal changes in PAaM/PLSec-AFP levels over time and incident HCC need to be determined in future studies. Such analysis may refine HCC risk prediction by adjusting the estimate for potential influence of medical and/or lifestyle interventions (eg, antiviral therapies, weight loss, and healthy diet). An ancillary HEDS study is ongoing to evaluate association of longitudinal PLSec changes and HCC incidence (protocol ID 536). Nevertheless, our analysis in large prospective cohorts validated that the baseline risk assessment can provide clinically meaningful HCC risk estimate. Third, our previous publications have suggested that our biomarkers can predict prognosis in patients with noncirrhotic liver disease,^46,52 which should be further validated in prospective cohorts.

In conclusion, PAaM was successfully validated for HCC risk stratification in patients with cirrhosis from contemporary HCC etiologies in 2 large multicenter phase 3 biomarker validation cohorts in the United States. This represents a significant step toward the clinical translation of an individual risk-based HCC screening strategy to improve early HCC detection and reduce HCC mortality.

What You Need to Know.

Background and Context

Hepatocellular carcinoma (HCC) risk stratification is an unmet need for cost-effective HCC screening and early detection in patients with cirrhosis to improve poor HCC prognosis.

New Findings

In 2 independent phase 3 biomarker validation studies including 3484 patients with cirrhosis, high-risk prediction of an integrative HCC risk score, prognostic liver secretome signature with α-fetoprotein plus age, male sex, albumin-bilirubin, and platelets (PAaM), was associated with HCC development with subdistribution hazard ratios of 7.51 and 6.54.

Limitations

The association between PAaM and HCC risk should be validated in hepatitis B virus–infected and Asian patients who were underrepresented in this study. Association of longitudinal changes in PAaM with HCC risk is of interest for future studies.

Clinical Research Relevance

This study represents the largest and first phase 3 biomarker validation study that establishes an integrative molecular/clinical score, PAaM, for HCC risk stratification.

Basic Research Relevance

Serum protein panel, which is computationally derived from tissue transcriptome, improves clinical variable-based score for noninvasive prognostic prediction, incorporating biological information from diseased organ.

Abbreviations used in this paper

ADRESS

age, diabetes, race, etiology of cirrhosis, sex, and severity

AFP

α-fetoprotein

ALD

alcohol-associated liver disease

aMAP

age, male sex, albumin-bilirubin, and platelets

HCC

hepatocellular carcinoma

HCV

hepatitis C virus

HEDS

Hepatocellular Carcinoma Early Detection Strategy

HR

hazard ratio

IQR

interquartile range

MASLD

metabolic dysfunction–associated steatotic liver disease

PAaM

prognostic liver secretome signature with α-fetoprotein plus age, male sex, albumin-bilirubin, and platelets

PLSec

prognostic liver secretome signature

sHR

sub-distribution hazard ratio

THCCC

Texas Hepatocellular Carcinoma Consortium

Data Availability

The research team will provide an e-mail address for communication once the information sharing is approved. The proposal should include detailed aims, statistical plan, and other information/materials to guarantee the rationality of requirement and the security of the data. The related patient data will be shared after review and approval of the submitted proposal and any related requested materials. Of note, data with patient names and other identifiers cannot be shared.