A single-center, retrospective analysis was performed on 707 patients with HCC who underwent adult LDLT at Seoul National University Hospital between 2003.01.01 and 2018.12.31. The last follow-up date was 2024.01.31 (Figure 1). The inclusion criteria were as follows: 1. age ≥19 years; 2. diagnosis of HCC following LDLT based on explant pathology reports; and 3. with available AFP and PIVKA-II measurements prior to transplantation. The exclusion criteria were as follows: 1. non-diagnosis of HCC following LDLT based on explant pathology; 2. re-transplantation; 3. missing AFP and PIVKA-II measurements; and 4. combined intra-hepatic carcinoma-hepatocellular carcinoma or intra-hepatic carcinoma. Diagnosis of HCC was based on postoperative histopathological examination. The study patients were subsequently categorized into the Milan cohort (MC) and Beyond Milan cohort (BMC) based on explant pathology.

Preoperative data, including demographic and clinical information, such as age, sex, body mass index (BMI), arterial hypertension, diabetes mellitus, underlying disease (such as HCV and HBV infections), AFP level, PIVKA-II level, and Model for End-stage Liver Disease scores, were extracted from electronic medical records (EMRs). Records of pre-transplant interventions, such as hepatectomy, transarterial chemoembolization, radiofrequency ablation, and percutaneous ethanol injection therapy, were reviewed from the EMRs. Tumor characteristics, including vascular invasion status (microvascular or macrovascular), tumor size, number of tumors, tumor stage, and differentiation grade, were extracted from post-LT pathological reports. HCC diagnosis, total necrosis, and graft-to-recipient weight ratio were confirmed by explant pathology findings. Overall survival (OS) was defined as the period from the date of liver transplantation until death from any cause. Patients who were alive at the last follow-up were censored. Recurrence-free survival (RFS) was defined as the duration from the date of liver transplantation to the first radiological or pathological evidence of HCC recurrence or death from any cause, whichever occurred first. Patients without recurrence or death were censored at the last follow-up date.

Preoperative AFP and PIVKA-II levels, as individual predictors, and four prognostic scores for HCC recurrence (SNAPP, RETREAT, MoRAL, and R3-AFP) were evaluated in both the MC and BMC [14–17]. AFP, PIVKA-II, SALT, and METROTICKET 2.0 scores were evaluated to predict post-LT mortality for both the MC and BMC [18, 19]. All prognostic scores were calculated retrospectively utilizing explant pathology reports, preoperative measurements, and radiological data. Time-dependent receiver operating characteristic (ROC) analysis was performed using Uno’s integrated area under the curve (iAUC), with inverse probability of censoring weighting (IPCW), to evaluate the dynamic predictive performance of the Cox regression models applied to both composite scores and individual biological tumor markers for HCC recurrence and mortality [32]. Furthermore, to assess the overall predictive ability of each model, Harrell’s C-index was calculated from the Cox regression models for both recurrence and mortality outcomes [33]. Confidence intervals for Uno’s iAUC were derived from 1,000 bootstrap samples, and iAUC values were compared between models using 1,000 bootstrap iterations to ensure the robustness and statistical reliability of the estimate [34, 35].

Continuous variable Cox regression models utilized calculated scores from prognostic models or measurements of biological markers, such as AFP and PIVKA-II levels, to predict HCC recurrence or survival following LDLT. In contrast, the threshold value binary Cox regression models were constructed using previously reported or validated threshold values (Supplementary Table S3). The Cox model specifications were as follows: AFP ≥200 (ng/mL) and PIVKA-II ≥400 (mAU/mL) cutoff values for predicting HCC recurrence and mortality were employed, as previously described [36, 37]. The SNAPP score was calculated using AFP, PIVKA-II, tumor size, number, and PET metabolic status, with a SNAPP ≥5 score used as a cutoff to indicate a high risk of HCC recurrence, as previously described [15]. The RETREAT score was calculated using the explant microvascular invasion status, largest tumor size, and preoperative AFP level, with RETREAT ≥5 indicating the high-risk group and as a threshold, as previously described [16]. The MoRAL score was calculated using preoperative AFP and PIVKA-II levels with a cutoff of 314.8, as previously described [14]. The R3-AFP score was based on the number of nodules, size of the largest tumor nodule, AFP level, microvascular invasion status, and tumor differentiation grade (Edmonson and Steiner grade >2), with a cutoff R3-AFP ≥3 score for cox models, as previously described [17]. SALT was based on the risk score, with a risk score ≥4.07 used as a threshold for higher mortality, as previously described [18]. The METROTICKET transplantability score was based on three categories: 1. If AFP <200 ng/mL, the sum of the number and size ≤7; 2; if 200 ≤ AFP <400 ng/mL, the sum of the number and size ≤5; 3. if 400 ≤ AFP <1,000 ng/mL, the sum of the number and size ≤4, as described previously [19].

The Kaplan–Meier method was used to estimate overall survival and recurrence-free survival. Chi-square or Fisher’s exact tests were used to compare categorical variables, and Student’s t-test was used for continuous variables. Statistical analyses were performed using SPSS version 29 (IBM SPSS Inc., Armonk, NY, United States) and R version 4.4.1 ¹ . Uno’s iAUC was calculated using “survAUC” package, and Harrell’s C-index was computed using “survcomp” package in R. Cox models were fitted using “survival.” All statistical tests were two-sided, with a significance threshold of 0·05, and were performed within an exploratory framework.

This study adhered to the Declaration of Helsinki and Istanbul guidelines and was approved by the Institutional Review Board of Seoul National University Hospital (IRB-H-2502-060-1612). The need to obtain informed patient consent was waived owing to the retrospective nature of the study. This study adhered to the STROBE guideline for retrospective study and check lists are provided in the Supplementary Material.

From 707 screened patients, 117 were excluded due to missing data, combined cholangiocarcinoma-hepatocellular carcinoma, or re-transplantation history. Finally, 590 patients who underwent LDLT were categorized into the MC (n = 437) and BMC (n = 153) groups (Figure 1). The baseline characteristics were similar between the groups (Table 1). Hepatitis B was more common in the MC (81.0% vs. 75.8%), whereas hepatitis C was more common in the BMC (13.1% vs. 8.2%, p < 0.01). BMC showed elevated tumor markers with higher AFP (15,125.0 vs. 48.4 ng/mL, p = 0.03) and PIVKA-II levels (3,556.6 vs. 69.4 mAU/mL, p < 0.01). More BMC patients had AFP ≥200 ng/mL and PIVKA-II ≥400 mAU/mL. Prognostic scores, tumor characteristics, and PET-CT hypermetabolic activity were worse in the BMC (44.4% vs. 11.7%, p < 0.01, Table 1; Supplementary Table S1). BMC patients had higher recurrence rates (50.3% vs. 10.3%), mortality (46.4% vs. 15.6%), and HCC-specific deaths (87.3% vs. 45.6%, Table 2). The median follow-up was longer in the MC group (113.05 vs. 70.54 months, Table 2), whereas recurrence-free survival was shorter in the BMC group (51.8 vs. 89.8 months, p < 0.01). The median follow-up duration for the entire population was 104.6 months (IQR: 68.0–145.9 months).

Individual tumor markers showed modest predictive performance (Table 3; Figure 2). In the Milan cohort, AFP (continuous, Table 3; Figure 2A) yielded an iAUC of 0.58 (95% CI: 0.47–0.68) and PIVKA-II 0.68 (95% CI: 0.60–0.75). In the Beyond Milan cohort, AFP and PIVKA-II achieved iAUCs of 0.69 (95% CI: 0.61–0.79) and 0.68 (95% CI: 0.58–0.78, Table 3; Figure 2B), respectively. Threshold models followed similar trends. AFP ≥200 had C-indices of 0.71 (95% CI: 0.53–0.90) in Milan and 0.76 (95% CI: 0.66–0.86) in Beyond Milan; PIVKA-II ≥400 reached 0.73 (95% CI: 0.52–0.93) and 0.71 (95% CI: 0.60–0.83). Combination marker models (AFP + PIVKA-II) showed improved performance. The continuous model yielded iAUCs of 0.64 (95% CI: 0.55–0.72) in Milan and 0.72 (95% CI: 0.62–0.80) in Beyond Milan. The threshold version yielded C-indices of 0.68 (95% CI: 0.50–0.86) and 0.73 (95% CI: 0.66–0.81). The MoRAL model, designed for Beyond Milan populations, achieved an iAUC of 0.65 (95% CI: 0.58–0.73, Figure 2D) and a C-index of 0.75 (95% CI: 0.65–0.86). Models incorporating morphology and vascular invasion showed stronger performance. R3-AFP (continuous, Table 3; Figure 2A) yielded iAUCs of 0.64 (95% CI: 0.54–0.74, Figure 2A) in Milan and 0.79 (95% CI: 0.71–0.86, Figure 2B) in Beyond Milan. R3-AFP ≥3 showed C-indices of 0.76 (95% CI: 0.58–0.93) and 0.77 (95% CI: 0.67–0.87). RETREAT, evaluated specifically in the Milan cohort, showed an iAUC of 0.60 (95% CI: 0.48–0.72), and RETREAT ≥5 achieved a C-index of 0.75 (95% CI: 0.30–1.00). The SNAPP model, which incorporates tumor biology, morphology, vascular invasion, and PET metabolism, showed an iAUC of 0.54 (95% CI: 0.42–0.65) in Milan and 0.72 (95% CI: 0.63–0.82) in Beyond Milan. SNAPP ≥5 demonstrated (Table 3; Figure 2) the strongest threshold performance: 0.88 (95% CI: 0.72–1.00) in Milan and 0.78 (95% CI: 0.69–0.87) in Beyond Milan.

Individual tumor markers, including AFP and PIVKA-II, showed moderate predictive accuracy, with slightly better performance in the Beyond Milan cohort (Table 3). While combining these markers modestly improved discrimination, complex models such as R3-AFP and SNAPP ≥5 consistently showed higher iAUCs and C-indices across both cohorts. Recurrence-free survival for each threshold model and marker was evaluated, and results are shown in Figure 3.

In mortality prediction, PIVKA-II (iAUC 0.71, Table 4; Figure 4) outperformed AFP (0.53) in Milan, with improvement when combined (0.66). SALT ≥4.07 had the highest C-index in Milan (0.86; 95% CI: 0.72–1.00). In Beyond Milan (Table 4; Figure 4), SALT and METROTICKET 2.0 showed C-indices of 0.80 (95% CI: 0.71–0.89) and 0.74 (95% CI: 0.62–0.86), respectively. Overall survival rates for each model are presented in Figure 5.

We further analyzed location-specific HCC recurrence by intrahepatic and extrahepatic patterns (Supplementary Figures S1,S2; Supplementary Table S2). In the Milan cohort, complex prognostic models incorporating tumor morphology, vascular invasion, and PET metabolic activity showed improved performance for intrahepatic recurrence. R3-AFP achieved an iAUC of 0.82 (95% CI 0.63–0.94), RETREAT ≥5 a C-index of 0.93 (95% CI 0.79–1.00), and SNAPP ≥5 a C-index of 0.96 (95% CI 0.85–1.00). In the Beyond Milan cohort, R3-AFP yielded an iAUC of 0.68 (95% CI 0.52–0.82), SNAPP ≥5 a C-index of 0.71 (95% CI 0.52–0.90), and MoRAL ≥314.8 a C-index of 0.78 (95% CI 0.61–0.94).

For extrahepatic recurrence in the Milan cohort (Supplementary Figure S2; Supplementary Table S2), AFP and PIVKA-II showed iAUCs of 0.56 (95% CI 0.45–0.67) and 0.66 (95% CI 0.57–0.75), respectively. The combined AFP + PIVKA-II model reached an iAUC of 0.61 (95% CI 0.51–0.70). Composite models outperformed individual markers: R3-AFP reached 0.60 (95% CI 0.49–0.72), and SNAPP ≥5 achieved the highest C-index at 0.81 (95% CI 0.56–1.00). In the Beyond Milan cohort, R3-AFP showed the highest iAUC at 0.83 (95% CI 0.75–0.89), followed by SNAPP at 0.77 (95% CI 0.67–0.96) and AFP + PIVKA-II at 0.73 (95% CI 0.62–0.92). Threshold models R3-AFP ≥3 and SNAPP ≥5 demonstrated C-indices of 0.86 (95% CI 0.76–0.95) and 0.80 (95% CI 0.71–0.90), respectively.

Overall, complex prognostic models provided higher predictive accuracy for both intrahepatic and extrahepatic recurrence than biomarker-only approaches (Supplementary Table S2). Performance was generally higher in the Beyond Milan cohort.

Single tumor markers, including PIVKA-II and AFP, showed comparable predictive accuracy to selected multivariable models in post-LT recurrence prediction. In the Milan cohort, PIVKA-II achieved higher iAUC than AFP (0.68 vs. 0.58, Table 3; Figures 2A,C), while both performed similarly in the Beyond Milan cohort (0.68 vs. 0.69). AFP showed better early discrimination but declined after 2 years, whereas SNAPP, MoRAL, and R3-AFP remained stable [3, 28, 29, 38–41]. AFP was limited in long-term prediction. Combining AFP and PIVKA-II improved prediction in Beyond Milan (iAUC: 0.72), comparable to SNAPP and MoRAL (0.72–0.73). In threshold models, PIVKA-II ≥400 had the highest C-index in Milan (0.73), whereas AFP ≥200 performed better in Beyond Milan (0.76). SNAPP ≥5 showed the highest C-index in Milan (0.88). While AFP remains widely used, dynamic assessment offers better early prediction but weaker long-term value [3, 28, 29, 38–41]. PIVKA-II shows independent accuracy, particularly for early recurrence [36, 40, 42, 43]. Continuous models offer time-sensitive monitoring advantages over fixed thresholds, and AFP and PIVKA-II integration improves risk stratification post-LT [1, 40].

The R3-AFP model, developed using data from 47 Euro-American centers, showed C-index values of 0.76–0.78 in external validation [17]. In this study, it showed the strongest long-term recurrence prediction among evaluated models (iAUC: 0.64 in Milan and 0.79 in Beyond Milan, Table 3; Figures 2A,B), and strong threshold performance (C-index: 0.76 in Milan and 0.77 in Beyond Milan). Although initially developed in DDLT populations, R3-AFP generalized well to LDLT settings. The incorporation of tumor burden and pathology appears to enhance its predictive power compared to single markers [17]. A recent study validated R3-AFP’s prognostic value in LT recipients with mammalian target of rapamycin inhibitor (mTORi)-based immunosuppression, which could potentially decrease HCC recurrence and improve survival [44].

The SNAPP score showed limited performance as a continuous Cox model in the MC (iAUC: 0.58; C-index: 0.57, Table 3; Figure 2) and BMC. However, SNAPP ≥5 performed better, with C-indices of 0.88 in MC and 0.78 in BMC. For intra-hepatic recurrence, continuous SNAPP had the second highest iAUC (0.73) after R3-AFP (0.82), whereas its threshold model achieved the highest C-index (0.96) in the MC. For extrahepatic recurrence, SNAPP ≥5 maintained C-indices above 0.80 in both cohorts (Supplementary Table S2; Supplementary Figure S2). These results align with the model’s design for LDLT populations in HBV-endemic Asian regions [15, 45]. Prior validation showed a C-index of 0.84 [15]; however, as SNAPP was developed for 5-year recurrence prediction, its fifth-year C-indices were below 80% in both cohorts (Supplementary Tables S1, S2).

The MoRAL score, developed in our center to assess recurrence risk beyond Milan criteria previously, was based on data collected between 2001 and 2013 [14]. Although a C-index above 80% was expected, the BMC showed a C-index of 0.75 (Table 3; Figure 2). The 10-year iAUC of MoRAL was 0.65 for BMC, similar to that of AFP and PIVKA-II. For intra-hepatic recurrence, MoRAL showed lower performance (iAUC 0.64), whereas MoRAL ≥314.8 achieved better discrimination (C-index 0.78). Lower early AUCs may reflect the model’s focus on tumor markers. Studies have validated MoRAL in hepatectomy [46], and deep learning integration has improved accuracy [47]. However, partial data overlap with the original cohort requires further validation [14].

RETREAT, which was primarily validated in DDLT populations, was evaluated in the MC, aligning with its original purpose [16]. The continuous RETREAT model showed an iAUC of 0.60, which was lower than that of PIVKA-II but similar to those of R3-AFP and AFP + PIVKA-II combination. Its threshold version (RETREAT ≥5) performed better, with a C-index of 0.75, ranking third after SNAPP ≥5 and R3-AFP ≥3 (Table 3; Figure 2). For intra-hepatic recurrence, the continuous model showed moderate accuracy (iAUC 0.70, Supplementary Table S2; Supplementary Figure S1), whereas the threshold model demonstrated excellent discrimination (C-index 0.93). RETREAT has been validated in North American cohorts with strong discrimination [48, 49]. UK data confirmed its utility (C-index 0.77) [50], and European data showed a 10-year prediction capability for low-risk HCC recurrence groups [51]. The recent addition of AFP-L3 and PIVKA-II has improved its prognostic performance [52].

PIVKA-II, as an individual marker and in combination with AFP, showed strong and consistent predictive abilities for post-LT mortality in both MC and BMCs (Table 4; Figure 4). In the MC, PIVKA-II clearly outperformed AFP, highlighting the growing relevance of tumor biology markers in long-term risk assessment [36, 40, 42, 43]. Together, AFP and PIVKA-II, particularly when combined, offer a practical and accessible option for risk stratification, although multivariable models, such as SALT, remained superior for long-term individualized prognostication.

This retrospective study has several limitations. Selection bias and variations in clinical management between the MC and BMC cohorts may have affected model performance. Differences in tumor biology and the predominance of viral hepatitis in this cohort may limit generalizability to Western populations, where non-viral etiologies such as metabolic associated liver diseases are more common. The higher recurrence and mortality rates observed in BMC may have led to an overestimation of risk. Furthermore, our study did not adjust for differences in recurrence treatments such as TACE or chemotherapy, which may have introduced bias. Additionally, immunosuppressive agents such as mammalian target of rapamycin inhibitors (mTORi) and steroids were not standardized and may have varied during model development, potentially impacting predictive performance. From a methodological standpoint, although Uno’s iAUC and Harrell’s C-index offer robust time-dependent and overall performance assessments, these statistical measures do not directly translate to clinical decision-making. The clinical relevance of modest differences in performance metrics remains uncertain. The prognostic models evaluated were developed under varying conditions. SNAPP and MoRAL were designed for LDLT populations, whereas RETREAT and R3-AFP were validated in DDLT settings. MoRAL and SALT were derived from single-center data, which may introduce institutional bias. Furthermore, the lack of an external validation cohort remains a major limitation. Future multicenter studies are necessary to confirm the generalizability and clinical utility of these findings.

However, this study provides long-term validation of prognostic models and tumor biomarkers for post-transplant outcomes in MC and BMCs, evaluating four recurrence and two mortality models over 5 years. In settings where LDLT recipients return to local care with limited diagnostic access, preoperative AFP and PIVKA-II levels could serve as accessible risk assessment markers. When applied as continuous variables, they showed moderate to strong predictive performance, with PIVKA-II outperforming AFP, particularly for long-term mortality. Their combination improved accuracy, matching complex models in high-risk populations. Their threshold-based models demonstrated performance comparable to that of complex models, particularly in BMC. R3-AFP showed the highest consistent predictive performance, whereas SNAPP, MoRAL, and SALT also performed well for BMC. Prognostic models and tumor biological scores generally performed better, particularly in the BMC cohort, where tumors were morphologically larger.

In conclusion, preoperative PIVKA-II, alone or in combination with AFP, may serve as an accessible long-term risk assessment marker for HCC recurrence and mortality following LDLT. However, AFP and PIVKA-II do not fully replace validated multivariable models, which remain the preferred approach in centers with advanced diagnostic capabilities.