Beyond Single Biomarkers: A Multidimensional Risk Score for Improved Post-Transplant Prognosis Prediction in Hepatocellular Carcinoma

Abstract

Introduction

Inflammation- and nutrition-related biomarkers are associated with prognosis in hepatocellular carcinoma (HCC), but most studies have focused on single markers. This study aimed to develop and internally validate a multidimensional inflammatory-nutritional risk score (RS) for predicting post-transplant outcomes in HCC patients.

Materials and methods

This single-center retrospective cohort included 239 patients with pathologically confirmed HCC who underwent liver transplantation between April 2013 and July 2023. Fourteen preoperative inflammation- and nutrition-related markers were calculated from laboratory data obtained within one week before transplantation. A bootstrap LASSO-Cox framework with 2000 resampling iterations was used to identify stable prognostic markers, and those selected in >70% of iterations were retained for the primary RS. Overall survival (OS) and recurrence-free survival (RFS) were assessed. Model performance was evaluated using Cox regression, C-index, time-dependent ROC analysis, calibration, decision curve analysis, and bootstrap optimism-corrected internal validation.

Results

During a median follow-up of 53 months, 83 OS events and 89 RFS events occurred. Four markers, including fibrinogen-to-albumin ratio, gamma-glutamyl transpeptidase-to-prealbumin ratio, gamma-glutamyl transpeptidase-to-albumin ratio, and aspartate aminotransferase-to-neutrophil ratio index, were selected to construct the primary RS. Patients with high RS had significantly worse OS and RFS than those with low RS, and RS remained independently associated with both outcomes in multivariable Cox analyses. Adding RS improved prognostic performance beyond Milan criteria, Hangzhou criteria, UCSF criteria, and clinicopathological models incorporating alpha-fetoprotein, tumor burden, portal vein tumor thrombus, and microvascular invasion. The final RS-plus clinicopathological model achieved apparent C-indices of 0.766 for OS and 0.759 for RFS, with bootstrap optimism-corrected C-indices of 0.756 and 0.749, respectively.

Conclusion

A four-marker inflammatory-nutritional RS based on routinely available preoperative biomarkers was independently associated with OS and RFS after liver transplantation for HCC and provided incremental prognostic information beyond established criteria and clinicopathological factors. Because the RS-plus model incorporates microvascular invasion, it should be interpreted as a postoperative prognostic tool. External validation is required before clinical application.

Hepatocellular carcinoma (HCC) remains one of the most common malignancies and a leading cause of cancer-related mortality worldwide, with a persistently high disease burden despite advances in surveillance and treatment.¹ For selected patients with HCC, liver transplantation (LT) offers the unique advantage of removing both the tumor and the underlying cirrhotic liver, thereby representing a potentially curative treatment strategy.² However, post-transplant tumor recurrence continues to be a major determinant of long-term survival. Although conventional selection systems, such as the Milan, Hangzhou, and UCSF criteria, have improved candidate selection and risk stratification, considerable prognostic heterogeneity remains among patients who undergo LT. Therefore, there is a continuing need for reliable, easily accessible, and clinically interpretable biomarkers that can refine post-transplant prognostic assessment beyond established tumor burden-based criteria.

Systemic inflammation, nutritional status, coagulation activation, cholestasis, and hepatic functional reserve are closely involved in the biological behavior of HCC and may influence recurrence and survival after LT. Inflammation contributes to tumor progression by promoting immune dysregulation, angiogenesis, invasion, and metastatic dissemination.^3,4 Meanwhile, impaired nutritional status and poor hepatic functional reserve may reflect both tumor-host interaction and reduced physiological resilience, which can adversely affect postoperative recovery and long-term outcomes. Accordingly, a variety of inflammation- and nutrition-related biomarkers have been investigated in HCC, including the fibrinogen-to-albumin ratio (FAR),⁵ neutrophil-to-lymphocyte ratio (NLR),⁶ platelet-to-lymphocyte ratio (PLR),⁷ monocyte-to-lymphocyte ratio (MLR),⁸ systemic immune-inflammation index (SII),⁹ prognostic inflammatory index (PII),¹⁰ aspartate aminotransferase-to-neutrophil ratio index (ANRI),¹¹ albumin-to-globulin ratio (AGR),¹² gamma-glutamyl transpeptidase-to-albumin ratio (GAR),¹³ gamma-glutamyl transpeptidase-to-prealbumin ratio (GPAR),¹⁴ albumin-to-alkaline phosphatase ratio (AAPR),¹⁵ prognostic nutritional index (PNI),¹⁶ albumin-bilirubin (ALBI) grade,¹⁷ and platelet-albumin-bilirubin (PALBI) grade.¹⁸

We previously reported that FAR was an independent prognostic factor in HCC patients after LT using the same institutional cohort.⁵ However, that study focused on a single biomarker and did not determine whether a multidimensional combination of inflammatory and nutritional markers could provide more stable and incremental prognostic information. Indeed, most existing studies have evaluated individual biomarkers separately, which may not fully capture the complex interplay among inflammation, coagulation, nutrition, cholestasis, and liver functional reserve. Moreover, the optimal combination of these markers and their added value beyond established transplant criteria and clinicopathological risk factors remain insufficiently defined.

In the present study, we systematically evaluated a panel of preoperative inflammation- and nutrition-related biomarkers in HCC patients undergoing LT. Using a bootstrap LASSO-Cox framework, we aimed to identify the most stable prognostic markers and develop a multidimensional inflammatory-nutritional risk score (RS). We further assessed whether this RS could improve prediction of overall survival and recurrence-free survival beyond conventional transplant criteria and clinicopathological models. By integrating complementary biological dimensions into a clinically accessible scoring system, this study sought to provide a robust and interpretable tool for individualized post-transplant risk stratification in patients with HCC.

Methods

Patient Selection

This retrospective study analyzed patients with hepatocellular carcinoma (HCC) who underwent liver transplantation at Beijing Chaoyang Hospital, Capital Medical University, from April 2013 to July 2023. All cases were pathologically confirmed as HCC by postoperative histology. The Ethics Committee of Beijing Chaoyang Hospital approved the retrieval and analysis of patient clinical data (No. 2020-D.-303, approved 28 May 2020).

The inclusion criteria were as follows: histopathologically confirmed primary HCC, no concurrent malignancies or extrahepatic metastatic lesions, available follow-up information, complete clinical and pathological records, and comprehensive preoperative blood test results obtained within one week before surgery. Patients were excluded if they had concurrent malignancies or metastatic lesions, died during the perioperative period, or lacked critical clinical, pathological, follow-up, or laboratory data. A total of 260 patients were initially screened, and 239 patients were finally included in the analysis. The detailed patient selection process is summarized in Figure 1.

Figure 1.

Study flowchart of patient selection. A total of 260 patients diagnosed with hepatocellular carcinoma (HCC) who underwent liver transplantation (LT) were initially screened. Patients were excluded if they died within 1 month after LT, lacked sufficient blood records or clinical data, or had other tumors or metastases. Finally, 239 patients were included in the final analysis.

Data Collection

Patient data were obtained from the hospital information system for individuals who underwent liver transplantation for HCC. Laboratory results were cross-verified using the electronic medical record system. An Excel database was used for systematic data entry, checking, and management. The collected variables included sex, age, hepatitis B surface antigen status, alpha-fetoprotein (AFP), preoperative hematological and biochemical parameters, tumor-related pathological variables, and follow-up information. Laboratory variables included albumin, globulin, prealbumin, total bilirubin, fibrinogen, alkaline phosphatase, gamma-glutamyl transpeptidase, aspartate aminotransferase, white blood cell count, neutrophil count, lymphocyte count, monocyte count, and platelet count. Tumor-related variables included T stage, N stage, TNM stage, number of intrahepatic tumors, maximum tumor diameter, total intrahepatic tumor diameter, gross tumor morphology, histological type, tumor differentiation grade, macrovascular invasion, microvascular invasion (MVI), and portal vein tumor thrombus (PVTT). Transplant criteria, including Milan, Hangzhou, and UCSF criteria, were also recorded.

Missing values were checked across all clinical, pathological, laboratory, and follow-up variables. Complete case analysis was performed, and patients with missing data in key variables required for model construction or survival analysis were excluded.

Follow-Up

All transplant recipients were followed regularly after surgery. The median follow-up duration was 53 months, with a range of 4–134 months. Follow-up visits were scheduled monthly during the first six postoperative months. Between six months and two years after transplantation, follow-up was generally performed every three to six months. After two years, follow-up was continued at regular intervals according to clinical condition.

Routine surveillance included serum tumor markers and contrast-enhanced abdominal computed tomography (CT) or magnetic resonance imaging (MRI). Imaging examinations were generally performed every six months. When local recurrence or distant metastasis was suspected, additional targeted imaging examinations, including CT, MRI, bone scanning, or positron emission tomography-computed tomography, were performed according to clinical indications.

Overall survival (OS) was calculated from the date of liver transplantation to death from any cause or the last confirmed follow-up. Recurrence-free survival (RFS) was calculated from the date of liver transplantation to the first documented tumor recurrence, death, or last follow-up, whichever occurred first. Patients without recurrence or death were censored at the date of last follow-up.

Definition of Inflammatory and Nutritional Related Markers

Inflammatory and nutritional related markers were calculated using laboratory test results obtained within one week before liver transplantation. The following markers were evaluated:

FAR = fibrinogen (g/L) / serum albumin (g/L)

NLR = neutrophil count (10⁹/L) / lymphocyte count (10⁹/L)

PLR = platelet count (10⁹/L) / lymphocyte count (10⁹/L)

MLR = monocyte count (10⁹/L) / lymphocyte count (10⁹/L)

SII = neutrophil count (10⁹/L) × platelet count (10⁹/L) / lymphocyte count (10⁹/L)

PII = neutrophil count (10⁹/L) × monocyte count (10⁹/L) / lymphocyte count (10⁹/L)

ANRI = aspartate aminotransferase (U/L) / neutrophil count (10⁹/L)

AGR = serum albumin (g/L) / serum globulin (g/L)

GAR = gamma-glutamyl transpeptidase (U/L) / serum albumin (g/L)

GPAR = gamma-glutamyl transpeptidase (U/L) / serum prealbumin (g/L)

AAPR = serum albumin (g/L) / alkaline phosphatase (U/L)

PNI = serum albumin (g/L) + 5 × lymphocyte count (10⁹/L)

ALBI = log₁₀ total bilirubin (μmol/L) × 0.66 − albumin (g/L) × 0.085

PALBI = 2.02 × log₁₀ total bilirubin (μmol/L) − 0.37 × [log₁₀ total bilirubin (μmol/L)]² − 0.04 × serum albumin (g/L) − 3.48 × log₁₀ platelet count (10⁹/L) + 1.01 × [log₁₀ platelet count (10⁹/L)]²

Development of the Risk Scoring System

To improve the robustness of variable selection and reduce instability caused by single biomarker screening, bootstrap resampling combined with LASSO-Cox regression was used to identify stable inflammatory-nutritional predictors. Fourteen inflammatory and nutritional indicators were included as candidate markers: FAR, NLR, PLR, MLR, SII, PII, ANRI, AGR, GAR, GPAR, AAPR, PNI, ALBI, and PALBI.

Bootstrap resampling was performed for 2000 iterations. In each bootstrap sample, the sample size was the same as that of the original analysis cohort, and patients were sampled with replacement. Candidate variables were standardized using Z-scores before LASSO-Cox regression. LASSO-Cox regression was performed with α set to 1, and the optimal λ.min was determined by 10-fold cross-validation. Variables with non-zero coefficients in each bootstrap iteration were recorded. The selection frequency of each marker was then calculated as the proportion of bootstrap iterations in which the marker was retained.

Markers selected in more than 70% of bootstrap iterations were retained to construct the primary risk score (RS). This threshold was selected to balance variable stability and model parsimony. To further examine the stability of the selection strategy, additional RS models were constructed using alternative selection-frequency thresholds of >50%, >60%, and >80%. These models were compared with the primary >70% RS in sensitivity analyses.

For survival modeling, the RS was analyzed as a continuous variable. For survival curve visualization and subgroup analyses, the optimal cutoff for RS was identified using maximally selected rank statistics. Patients were then stratified into high-RS and low-RS groups according to this cutoff.

Model Construction and Validation

The final RS-plus clinicopathological model was constructed using Cox proportional hazards regression. The model incorporated RS, maximum tumor size, tumor number, PVTT, and MVI. These variables were selected based on clinical relevance, pathological importance, non-overlapping prognostic information, and model parsimony.

Nomograms were constructed to predict 1-, 3-, and 5-year OS and RFS. Model discrimination was evaluated using Harrell’s concordance index (C-index) and time-dependent receiver operating characteristic curves. Time-dependent area under the curve (AUC) values were calculated at 1, 3, and 5 years. Calibration curves were generated to assess the agreement between predicted and observed survival probabilities. Decision curve analysis was used to evaluate the clinical net benefit of the RS-plus model compared with the clinicopathological model without RS, the treat-all strategy, and the treat-none strategy.

The incremental prognostic value of RS was evaluated by adding RS to established transplant criteria and clinicopathological models, including Milan criteria, Hangzhou criteria, UCSF criteria, AFP plus tumor burden, AFP plus tumor burden plus PVTT, and AFP plus tumor burden plus PVTT plus MVI. Tumor burden was defined by maximum tumor size and tumor number. Incremental value was assessed using changes in C-index, likelihood-ratio tests, AIC, and bootstrap 95% confidence intervals for ΔC-index.

Sensitivity Analyses

Several sensitivity analyses were performed to evaluate the robustness of the primary RS. First, alternative RS models were constructed using different bootstrap selection-frequency thresholds, including >50%, >60%, >70%, and >80%. Second, marker-redundancy analyses were performed by removing GAR, GPAR, or both GAR and GPAR from the primary RS. Third, endpoint-specific derivation analyses were conducted by deriving RS models using OS and RFS separately and then cross-applying them to both endpoints.

Internal validation of the final RS-plus clinicopathological model was performed using 1000 bootstrap resamples. The optimism-corrected C-index was calculated by subtracting the mean optimism from the apparent C-index.

Exploratory subgroup analyses were also performed to evaluate whether the prognostic association of high RS was consistent across clinically relevant subgroups, including age, sex, AFP level, maximum tumor size, tumor number, total tumor size, TNM stage, PVTT, MVI, pathological differentiation, and Milan, Hangzhou, and UCSF criteria. Hazard ratios and 95% confidence intervals were estimated for high RS versus low RS within each subgroup. P for interaction was used to assess heterogeneity across subgroup strata.

Statistical Analysis

Statistical analyses were performed using SPSS 26.0 and R 4.4.2. Normally distributed continuous variables are reported as mean ± standard deviation and were compared between groups using independent samples t-tests. Non-normally distributed continuous variables are expressed as median with interquartile range and were compared using Kruskal–Wallis tests. Categorical variables are presented as counts and percentages and were compared using χ²-tests or Fisher’s exact tests, as appropriate. When multiple pairwise comparisons were performed, P values were adjusted using the Bonferroni method.

Survival analyses were performed using the Kaplan–Meier method, and differences between groups were compared using the Log rank test. Prognostic factors were evaluated using univariable and multivariable Cox proportional hazards regression models. The proportional hazards assumption was assessed using Schoenfeld residuals. Events per variable were calculated for the primary multivariable Cox models. Model discrimination was assessed using C-index and time-dependent AUC. Calibration curves and decision curve analysis were used to evaluate model calibration and clinical utility. A two-sided P value <0.05 was considered statistically significant. Potential nonlinear associations between RS and survival outcomes were assessed using restricted cubic spline Cox regression adjusted for maximum tumor size, tumor number, PVTT, and MVI. The nonlinearity test was based on the nonlinear component of the spline model.

Results

Patient Characteristics

A total of 239 patients with HCC who underwent LT were included in the final analysis. According to the optimal RS cutoff, 148 patients were assigned to the low-RS group and 91 patients to the high-RS group. Compared with the low-RS group, patients in the high-RS group had more aggressive tumor characteristics. The high-RS group had higher proportions of advanced TNM stage, MVI, multiple tumors, maximum tumor size >5 cm, total tumor size >8 cm, poor or undifferentiated pathological differentiation, AFP >400 ng/mL, PVTT, and patients beyond Milan, Hangzhou, and UCSF criteria. In contrast, age, sex, and HBsAg status did not significantly differ between the two RS groups (Table 1).

Table 1.

Baseline Clinicopathological Characteristics of Patients with Hepatocellular Carcinoma Undergoing Liver Transplantation Stratified by Risk Score Group

Variables	Overall (n = 239)	Low RS Group (n = 148)	High RS Group (n = 91)	p-value
Gender (Male/Female)	220/19	134/14	86/5	0.331
TNM stages (III/I–II)	51/188	17/128	34/56	<0.001
Age, years (>55/≤55)	105/134	72/76	33/58	0.081
Positive HBsAg (+/−)	216/23	132/16	84/7	0.503
MVI (Yes/No)	131/108	68/80	63/28	<0.001
Tumor number (Multiple/Single)	129/110	72/76	57/34	0.045
Maximum tumor size, cm (>5/≤5)	62/177	20/128	42/49	<0.001
Total tumor size, cm (>8/≤8)	48/191	13/135	35/56	<0.001
Differentiation (Poor, undifferentiated/ Well, moderate)	184/55	106/42	78/13	0.012
AFP, ng/mL (>400/≤400)	56/183	21/127	35/56	<0.001
PVTT (+/−)	33/206	12/136	21/70	0.002
Milan criteria (Yes/No)	111/128	96/52	15/76	<0.001
Hangzhou criteria (Yes/No)	193/46	134/14	59/32	<0.001
UCSF criteria (Yes/No)	141/98	115/33	26/65	<0.001

No missing data were observed for OS follow-up time, OS event status, RFS follow-up time, RFS event status, RS, or the key clinicopathological variables included in the analysis (Table S1). There were 83 OS events and 89 RFS events. The EPV values were 16.6 for OS and 17.8 for RFS, indicating adequate event numbers for the primary multivariable Cox models (Table S2).

Construction of the Primary RS

Bootstrap LASSO-Cox selection identified four inflammatory-nutritional markers with selection frequencies greater than the prespecified 70% threshold (Figure 2A and B): FAR, GPAR, GAR, and ANRI. These four markers were used to construct the primary RS:

Figure 2.

Construction and evaluation of the inflammatory-nutritional risk score (RS). (A) Bootstrap selection frequencies of candidate inflammatory-nutritional markers using OS as the derivation endpoint. The red dashed line indicates the prespecified 70% selection-frequency threshold. (B) Bootstrap selection frequencies of candidate inflammatory-nutritional markers using RFS as the derivation endpoint. (C) Correlation matrix of the candidate markers retained under the broader selection pool. Color intensity represents the magnitude of correlation between markers. (D) Variance inflation factor (VIF) analysis of the four markers included in the primary RS. The red dashed line indicates a VIF threshold of 2. All included markers showed acceptable collinearity.

Abbreviations: OS, overall survival; RFS, recurrence-free survival; RS, risk score; VIF, variance inflation factor; FAR, fibrinogen-to-albumin ratio; GPAR, gamma-glutamyl transpeptidase-to-prealbumin ratio; GAR, gamma-glutamyl transpeptidase-to-albumin ratio; ANRI, aspartate aminotransferase-to-neutrophil ratio; PLR, platelet-to-lymphocyte ratio; AGR, albumin-to-globulin ratio; PALBI, platelet-albumin-bilirubin score; AAPR, albumin-to-alkaline phosphatase ratio; ALBI, albumin-bilirubin score.

RS = 11.524911 × FAR + 0.000058 × GPAR + 0.025626 × GAR − 0.001611 × ANRI

The correlation matrix demonstrated varying degrees of correlation among the candidate markers retained in the broader selection pool, supporting further assessment of potential multicollinearity (Figure 2C). Correlation analysis of candidate markers was performed to evaluate potential marker redundancy. VIF analysis showed acceptable collinearity among the four markers included in the primary RS (Figure 2D).

Association Between RS and Survival Outcomes

The optimal RS cutoff was 0.8, which was used to stratify patients into high- and low-RS groups (Figure 3A). Kaplan–Meier analysis showed that patients in the high-RS group had significantly worse OS and RFS than patients in the low-RS group (Figure 3B and C).

Figure 3.

Prognostic performance of the primary inflammatory-nutritional RS.(A) Distribution of the RS and determination of the optimal cutoff using maximally selected rank statistics. Patients were classified into high- and low-RS groups according to the cutoff value of 0.8. (B) Kaplan–Meier curves for OS according to RS group. Patients in the high-RS group showed significantly worse OS than those in the low-RS group. (C) Kaplan–Meier curves for RFS according to RS group. Patients in the high-RS group showed significantly worse RFS than those in the low-RS group. Shaded areas indicate 95% confidence intervals, and P values were calculated using the Log rank test. (D) Restricted cubic spline curve for the association between RS and overall survival. (E) Restricted cubic spline curve for the association between RS and recurrence-free survival.

Abbreviations: RS, risk score; OS, overall survival; RFS, recurrence-free survival.

In Cox regression analyses, RS was significantly associated with both OS and RFS. For OS, RS was associated with worse survival in univariable analysis and remained independently associated with OS in multivariable analysis after adjustment for clinicopathological factors, with an adjusted HR of 2.057 (Table 2). For RFS, RS was also significantly associated with adverse outcome and remained independently associated with RFS in multivariable analysis, with an adjusted HR of 1.903 (Table 3).

Table 2.

Univariate and Multivariate Cox Analyses of Baseline Characteristics and Risk Score on Overall Survival in Patients with Hepatocellular Carcinoma Undergoing Liver Transplantation

Variables	Univariate Analysis			Multivariate Analysis
	HR	95% CI	p-value	HR	95% CI	p-value
Age, years (>55vs≤55)	0.801	0.515–1.246	0.325
TNM stage (III vs I+II)	3.507	2.261–5.439	<0.001	1.473	0.806–2.691	0.208
MVI (Yes vs No)	4.757	2.717–8.329	<0.001	2.814	1.516–5.223	0.001
AFP, ng/mL (>400 vs ≤400)	0.805	0.503–1.286	0.364
Maximum tumor size, cm (>5 vs ≤5)	3.001	1.944–4.634	<0.001	1.609	0.799–3.239	0.183
Tumor number (Multiple vs Single)	1.753	1.116–2.753	0.015	1.167	0.651–2.094	0.605
Gender (Male vs Female)	1.601	0.648–3.955	0.308
Differentiation (Poor, undifferentiated vs Well, moderate)	2.623	1.661–4.143	<0.001	1.052	0.639–1.732	0.843
Positive HBsAg (+ vs −)	3.236	1.022–10.249	0.046	1.931	0.594–6.27	0.274
PVTT (+ vs −)	2.935	1.786–4.822	<0.001
Total tumor size, cm (>8 vs ≤8)	3.156	2.011–4.954	<0.001	1.118	0.547–2.284	0.761
RS	2.709	2.108–3.481	<0.001	2.057	1.505–2.811	<0.001

Table 3.

Univariate and Multivariate Cox Analyses of Baseline Characteristics and Risk Score on Disease-Free Survival in Patients with Hepatocellular Carcinoma Undergoing Liver Transplantation

Variables	Univariate Analysis			Multivariate Analysis
	HR	95% CI	p-value	HR	95% CI	p-value
Age, years (>55vs≤55)	0.823	0.538–1.259	0.369
TNM stage (III vs I+II)	3.457	2.259–5.289	<0.001	1.486	0.838–2.635	0.176
MVI (Yes vs No)	4.279	2.548–7.185	<0.001	2.547	1.429–4.537	0.001
AFP, ng/mL (>400 vs ≤400)	0.866	0.554–1.358	0.531
Maximum tumor size, cm (>5 vs ≤5)	2.973	1.952–4.528	<0.001	1.346	0.664–2.731	0.41
Tumor number (Multiple vs Single)	1.657	1.074–2.556	0.022	1.096	0.633–1.898	0.742
Gender (Male vs Female)	1.696	0.688–4.181	0.251
Differentiation (Poor, undifferentiated vs Well, moderate)	2.777	1.786–4.319	<0.001	1.102	0.675–1.797	0.698
Positive HBsAg (+ vs −)	4.711	1.124–11.24	0.031	2.237	0.693–7.221	0.178
PVTT (+ vs −)	3.121	1.937–5.029	<0.001	1.612	0.956–2.717	0.073
Total tumor size, cm (>8 vs ≤8)	3.381	2.191–5.218	<0.001	1.133	0.555–2.314	0.731
RS	2.533	2.002–3.203	<0.001	1.903	1.412–2.564	<0.001

Among clinicopathological variables, MVI was consistently associated with worse outcomes in multivariable analyses. Other tumor-related factors, including TNM stage, maximum tumor size, tumor number, PVTT, total tumor size, and differentiation, showed significant associations in univariable analyses but were attenuated after multivariable adjustment.

Restricted cubic spline Cox regression further showed significant overall associations between RS and survival outcomes. No strong evidence of nonlinearity was observed after adjustment for maximum tumor size, tumor number, PVTT, and MVI (P for nonlinearity > 0.05; Figure 3D and E). Therefore, RS was modeled as a continuous linear variable in the primary Cox regression analyses.

Performance of the RS-Plus Clinicopathological Model

The final RS-plus clinicopathological model incorporated RS, maximum tumor size, tumor number, PVTT, and MVI. RS-plus nomograms were constructed for individualized prediction of 1-, 3-, and 5-year OS and RFS (Figure 4A and B, respectively).

Figure 4.

RS-plus clinicopathological nomograms and incremental prognostic value of the RS. (A) Nomogram for predicting 1-, 3-, and 5-year OS based on the RS-plus clinicopathological model. The model incorporated RS, maximum tumor size, tumor number, PVTT, and MVI. (B) Nomogram for predicting 1-, 3-, and 5-year RFS based on the RS-plus clinicopathological model. (C) Time-dependent AUC curves comparing the clinicopathological model with the RS-plus clinicopathological model for OS and RFS. The RS-plus model showed consistently higher or comparable predictive performance across 1-, 3-, and 5-year time points. (D) Incremental prognostic value of adding RS to established transplant criteria and clinicopathological models. Points represent changes in C-index after adding RS, and horizontal bars represent bootstrap 95% confidence intervals. The dashed vertical line indicates no improvement. Asterisks indicate statistical significance by likelihood-ratio test.

Abbreviations: RS, risk score; OS, overall survival; RFS, recurrence-free survival; AUC, area under the curve; PVTT, portal vein tumor thrombus; MVI, microvascular invasion; CI, confidence interval; AFP, alpha-fetoprotein.

The RS-plus model showed favorable discrimination. For OS, the final model achieved an apparent C-index of 0.766. For RFS, the apparent C-index was 0.759 (Table S3). Time-dependent AUC analysis showed that the RS-plus clinicopathological model provided consistently higher or comparable predictive performance compared with the clinicopathological model without RS across 1-, 3-, and 5-year time points (Figure 4C).

Calibration curves showed good agreement between predicted and observed 1-, 3-, and 5-year OS and RFS probabilities. DCA demonstrated that the RS-plus clinicopathological model provided favorable net benefit compared with the clinicopathological model, treat-all strategy, and treat-none strategy across clinically relevant threshold probability ranges (Figure 5).

Figure 5.

Calibration and decision curve analysis of the RS-plus clinicopathological model. (A–C) Calibration curves for 1-, 3-, and 5-year OS prediction. (D–F) Calibration curves for 1-, 3-, and 5-year RFS prediction. The dashed diagonal line represents perfect agreement between predicted and observed probabilities, and the blue or Orange calibration curves represent the model-estimated calibration performance. (G–I) Decision curve analysis for 1-, 3-, and 5-year OS. (J–L) Decision curve analysis for 1-, 3-, and 5-year RFS. The RS-plus model was compared with the clinicopathological model, the treat-all strategy, and the treat-none strategy. The RS-plus model provided favorable net benefit across clinically relevant threshold probability ranges.

Abbreviations: RS, risk score; OS, overall survival; RFS, recurrence-free survival.

The proportional hazards assumption was assessed using Schoenfeld residuals. No evidence of violation of the proportional hazards assumption was observed. The global test was non-significant for both OS and RFS models (Table S4).

Incremental Prognostic Value of RS

Adding RS improved the prognostic performance of established transplant criteria and clinicopathological models.

For OS, adding RS increased the C-index of the Milan criteria model from 0.654 to 0.730, with a ΔC-index of 0.076. The C-index of the Hangzhou criteria model increased from 0.621 to 0.732, with a ΔC-index of 0.110. The C-index of the UCSF criteria model increased from 0.675 to 0.741, with a ΔC-index of 0.066. Adding RS also improved clinicopathological models. The C-index increased from 0.682 to 0.737 for AFP plus tumor burden, from 0.705 to 0.736 for AFP plus tumor burden plus PVTT, and from 0.743 to 0.767 for AFP plus tumor burden plus PVTT plus MVI. All likelihood-ratio tests were significant, and AIC decreased after adding RS (Figure 4D and Table S5).

Similar improvements were observed for RFS. Adding RS increased the C-index of the Milan criteria model from 0.657 to 0.724, with a ΔC-index of 0.067. The C-index increased from 0.614 to 0.719 for the Hangzhou criteria model and from 0.678 to 0.736 for the UCSF criteria model. For clinicopathological models, adding RS increased the C-index from 0.675 to 0.730 for AFP plus tumor burden, from 0.704 to 0.735 for AFP plus tumor burden plus PVTT, and from 0.743 to 0.759 for AFP plus tumor burden plus PVTT plus MVI. These findings indicate that RS provided additional prognostic information beyond established transplant criteria and conventional clinicopathological predictors (Figure 4D and Table S5).

Sensitivity Analyses

Sensitivity analyses supported the robustness of the primary RS. In the selection-threshold analysis, the >70% model retained four markers, including FAR, GPAR, GAR, and ANRI, and achieved strong discrimination with fewer variables. For OS, the >70% primary RS achieved a C-index of 0.718 and 1-, 3-, and 5-year AUCs of 0.804, 0.720, and 0.761, respectively. For RFS, the corresponding C-index was 0.705, with 1-, 3-, and 5-year AUCs of 0.747, 0.733, and 0.748 (Table S6).

The >60% model showed similar performance but required six variables, whereas the >50% model incorporated eight variables without improving discrimination. The >80% model retained only FAR and GPAR and showed reduced performance. These findings supported the >70% selection-frequency model as the primary RS because it achieved a favorable balance between stability, parsimony, and discrimination (Table S5).

Marker-redundancy analyses showed that removing either GAR or GPAR resulted in only modest performance attenuation. For OS, the C-index decreased from 0.718 to 0.711 after removing GAR and to 0.708 after removing GPAR. For RFS, the corresponding C-index decreased from 0.705 to 0.700 and 0.698. However, simultaneous removal of both GPAR and GAR led to a more evident decrease in discrimination, with C-indices of 0.678 for OS and 0.673 for RFS. These results suggest that GPAR and GAR contributed complementary prognostic information despite partial overlap (Table S6).

Endpoint-specific derivation analyses showed highly consistent performance. The OS-derived RS achieved C-indices of 0.718 for OS and 0.705 for RFS. The RFS-derived RS achieved C-indices of 0.704 for RFS and 0.716 for OS. These findings indicate that the RS was not dependent on a single endpoint definition (Table S6).

Bootstrap internal validation demonstrated limited optimism. The final RS-plus clinicopathological model had an apparent C-index of 0.766 for OS and 0.759 for RFS. After 1000 bootstrap resamples, the optimism-corrected C-indices were 0.756 for OS and 0.749 for RFS, indicating stable internal model performance (Table S7).

Exploratory Subgroup Analyses

Exploratory subgroup analyses showed that the adverse prognostic association of high RS was generally consistent across clinically relevant subgroups.

For OS, high RS was associated with increased mortality risk in the overall cohort, with an HR of 4.09. The direction of association was broadly consistent across age, sex, AFP level, maximum tumor size, tumor number, total tumor size, TNM stage, PVTT, MVI, pathological differentiation, and transplant criteria. No significant interaction was observed across these subgroup strata (Figure S1).

For RFS, high RS was also associated with increased risk of recurrence or death in the overall cohort, with an HR of 4.01. The adverse association of high RS remained directionally consistent across major clinical subgroups, and no significant interaction was observed. These findings suggest that the prognostic effect of RS was not strongly modified by major clinicopathological characteristics. However, given the limited sample size in some subgroups, these analyses should be interpreted cautiously (Figure S2).

Discussion

In this retrospective cohort study of patients with HCC undergoing LT, we developed and internally validated a multidimensional inflammatory-nutritional risk score based on routinely available preoperative laboratory markers. The primary RS consisted of four markers, including FAR, GPAR, GAR, and ANRI, which were selected using a bootstrap LASSO-Cox framework with a prespecified 70% selection-frequency threshold. The RS was significantly associated with both OS and RFS and remained independently prognostic after adjustment for established clinicopathological factors. More importantly, adding RS improved the discriminative performance of conventional transplant criteria and clinicopathological models, including Milan, Hangzhou, UCSF, AFP plus tumor burden, and models incorporating PVTT and MVI.^19–21 These findings suggest that the preoperative inflammatory-nutritional state contains clinically meaningful prognostic information that is not fully captured by tumor morphology or pathological features alone.

The selection of FAR, GPAR, GAR, and ANRI is biologically plausible. FAR integrates fibrinogen and albumin, thereby reflecting both coagulation activation and nutritional reserve. Elevated fibrinogen may promote tumor progression through inflammation, angiogenesis, extracellular matrix remodeling, and metastatic potential,^22,23 whereas reduced albumin reflects impaired nutritional status, systemic inflammation, and compromised hepatic functional reserve. We previously reported the prognostic value of FAR in HCC patients after LT; however, a single biomarker is unlikely to capture the complex host-tumor interaction that determines post-transplant outcomes. The present study extends our previous observation by embedding FAR into a multidimensional score that also includes markers related to cholestasis, hepatic injury, nutritional reserve, and systemic inflammatory status.

GPAR and GAR both incorporate gamma-glutamyl transpeptidase, a marker closely related to cholestasis, oxidative stress, hepatic inflammation, and aggressive tumor biology.²⁴ Their denominators, prealbumin and albumin, represent different but complementary aspects of nutritional and hepatic synthetic function.²⁵ Prealbumin has a shorter half-life than albumin and may be more sensitive to acute changes in nutritional and inflammatory status, whereas albumin reflects more stable hepatic reserve and systemic protein metabolism. Although GPAR and GAR partly share the same numerator, our marker-redundancy sensitivity analyses showed that removal of either marker resulted in only modest performance attenuation, whereas simultaneous removal of both led to a more evident decrease in discrimination. This finding suggests that GPAR and GAR are not merely redundant variables but may capture complementary pathophysiological information. ANRI, which combines aspartate aminotransferase and neutrophil count, may reflect liver injury, inflammatory activation, and tumor-associated host response. Together, the four components of the RS represent a compact but multidimensional signature of coagulation, nutrition, cholestasis, hepatic injury, and inflammation.

A major strength of this study is the use of bootstrap LASSO-Cox selection rather than a single data-driven screening procedure.²⁶ In biomarker studies, variable selection may be unstable, especially when multiple correlated inflammatory and nutritional indices are evaluated simultaneously. To reduce this instability, we used repeated bootstrap resampling to estimate marker selection frequency and selected markers that consistently appeared across iterations. The sensitivity analysis across selection thresholds further supported the choice of the >70% model. The >70% RS retained only four markers but achieved discrimination comparable to the >60% model and better than the >50% and >80% models. Therefore, the primary RS achieved a favorable balance between marker stability, parsimony, and predictive performance. The endpoint-specific sensitivity analysis also demonstrated that OS-derived and RFS-derived RS models showed highly consistent performance, supporting the robustness of the selected marker combination across survival endpoints.

The RS also provided incremental prognostic value beyond established transplant criteria. The Milan, Hangzhou, and UCSF criteria primarily rely on tumor burden and related morphological features, and they have played an important role in candidate selection for LT. However, these criteria do not directly incorporate systemic inflammation, nutritional reserve, coagulation activation, or cholestatic injury. In our study, adding RS improved the C-index of all three transplant criteria for both OS and RFS. Similar improvements were observed when RS was added to clinicopathological models incorporating AFP, tumor burden, PVTT, and MVI. These results indicate that RS provides prognostic information complementary to conventional tumor-centered variables. Clinically, this may be particularly relevant because patients with similar tumor burden can still have heterogeneous post-transplant outcomes due to differences in host inflammatory-nutritional status and hepatic functional reserve.

The final RS-plus clinicopathological model incorporated RS, maximum tumor size, tumor number, PVTT, and MVI. This model achieved favorable discrimination, with apparent C-indices of 0.766 for OS and 0.759 for RFS, and bootstrap-corrected C-indices of 0.756 and 0.749, respectively. The low optimism observed in bootstrap validation suggests that the model was not excessively overfitted despite being derived from a single-center cohort. Calibration analysis showed acceptable agreement between predicted and observed survival probabilities, and decision curve analysis demonstrated favorable net benefit across clinically relevant threshold probability ranges. These results suggest that the RS-plus model may provide additional prognostic information for postoperative risk stratification. However, the present study did not test RS-guided surveillance, immunosuppression adjustment, or treatment strategies. Therefore, its clinical use should be considered hypothesis-generating and requires external validation and prospective evaluation.

The exploratory subgroup analyses suggested a generally similar direction of association across clinically relevant strata within the study cohort. The adverse prognostic association of higher RS was directionally consistent across clinically relevant subgroups, including age, sex, AFP level, tumor burden, TNM stage, PVTT, MVI, pathological differentiation, and transplant criteria. Most interaction tests were non-significant, suggesting no strong evidence that the prognostic effect of RS was limited to a specific subgroup. However, these subgroup analyses should be interpreted cautiously because several strata, such as the female subgroup and PVTT-positive subgroup, had limited sample sizes and wide confidence intervals. Therefore, subgroup findings should be considered exploratory rather than confirmatory.

This study has several limitations. First, it was a retrospective single-center study, and selection bias cannot be fully excluded. Although bootstrap internal validation showed limited optimism, external validation in independent cohorts is still required before broader clinical application. Second, all inflammatory and nutritional markers were derived from preoperative laboratory tests, and dynamic postoperative changes were not incorporated. Longitudinal changes in inflammation, nutrition, liver function, and immunosuppression may provide additional prognostic information. Third, although we evaluated multiple routinely available inflammatory-nutritional indices, the study did not include molecular, radiological, or immune microenvironmental features. Integration of RS with imaging, pathological, and molecular data may further improve risk stratification. Fourth, the RS was developed using laboratory data from a specific transplant population, and its applicability to patients undergoing hepatectomy, locoregional therapy, or systemic therapy requires separate evaluation. Fifth, Several transplant-specific factors were not fully captured in the current retrospective dataset and may have influenced OS and RFS. These include bridging or downstaging therapy before LT, waiting time, immunosuppressive regimens, antiviral therapy, donor characteristics, early graft function, postoperative complications, treatment after recurrence, and changes in transplant selection criteria or perioperative management during the 10-year study period. Finally, although the RS is biologically interpretable, the mechanistic relationship between the selected markers and post-transplant recurrence remains to be clarified in future translational studies.

Conclusion

In conclusion, this study developed a multidimensional inflammatory-nutritional RS based on routinely available preoperative laboratory biomarkers, including FAR, GPAR, GAR, and ANRI, for patients with HCC undergoing LT. The RS was independently associated with OS and RFS and provided incremental prognostic information beyond established transplant criteria and conventional clinicopathological factors. The preoperative RS may help characterize host inflammatory-nutritional status, whereas the RS-plus clinicopathological model incorporating clinicopathological predictors should be interpreted as a postoperative prognostic model because clinicopathological predictors are generally determined from explant pathology. Although the RS-plus model showed favorable discrimination, calibration, decision-curve performance, and bootstrap internal validation, external validation in independent multicenter cohorts is required before clinical use.

Funding Statement

This research received no external funding.

Institutional Review Board Statement

The study was approved by the Institutional Review Board of Beijing chaoyang hospital in accordance with the 1964 Helsinki Declaration and its later amendments (No. 2020-D-303). All donor livers used in this study were from voluntary deceased donors with written informed consent signed by their legal representatives, and the organ donation and transplantation procedures were strictly conducted in accordance with the Declaration of Istanbul. All patient medical record data involved in this study were strictly confidential in accordance with the relevant medical privacy regulations, and no personal identifying information was disclosed in the research process and paper writing.

Data Sharing Statement

The datasets used and analyzed during the current study are available from the corresponding authors (Hezhi Wang, wanghezhi2016@mail.ccmu.edu.cn; Zhili Ji, anzhenjzl@mail.ccmu.edu.cn) on reasonable request.

Informed Consent Statement

This study is a retrospective study. It has been approved by the Ethics Committee of Beijing Chaoyang Hospital, and the requirement for patients to sign an informed consent form has been waived.(No. 2020-D-303).

Author Contributions

Xiaoyong Ye: Conceptualization, Methodology, Formal analysis, Software, Data curation, Visualization, Writing – original draft, Writing – review & editing. Simeng Lei: Investigation, Data curation, Validation, Visualization, Writing – original draft, Writing – review & editing. Guang Jia: Investigation, Data curation, Validation, Methodology, Writing – original draft, Writing – review & editing. Qiang He: Investigation, Resources, Data curation, Validation, Writing – review & editing. Zhili Ji: Conceptualization, Methodology, Supervision, Project administration, Resources, Writing – review & editing. Jun Ma: Conceptualization, Supervision, Project administration, Resources, Writing – review & editing. Xiaoyong Ye, Simeng Lei, and Guang Jia contributed equally to this work. All authors took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

A version of this paper was made available as a pre-print: https://doi.org/10.21203/rs.3.rs-7620800/v1. The authors declare no other conflict of interest.