Explant-based prognostic models for hepatocellular carcinoma recurrence after liver transplantation: A systematic review and meta-analysis

Abstract

BACKGROUND

Liver transplantation (LT) is the preferred curative treatment for early-stage hepatocellular carcinoma (HCC). However, approximately 17% of patients experience post-transplant recurrence. Extrahepatic recurrence and early recurrence (within one year after LT) are associated with poorer post-recurrence survival.

AIM

To assess which explant-based prognostic model best predicts HCC recurrence after LT.

METHODS

A systematic search was performed in PubMed, EMBASE, Web of Science, and the Cochrane Library from inception to January 30, 2025. Nine retrospective studies comprising 5348 patients were included. Three explant-based prognostic models were analyzed: (1) Risk estimation of tumor recurrence after transplant (RETREAT); (2) Decaens; and (3) Predicting Cancer Recurrence Score (PCRS). Primary outcomes included: (1) HCC recurrence rate; and (2) Predictive accuracy of each score over a five-year follow-up.

RESULTS

All studies were retrospective and included validation cohorts from North America, Europe, and Asia. The overall recurrence rate was 7%. For high-risk thresholds, pooled sensitivity and specificity were Risk Estimation of Tumor Recurrence after Transplant (RETREAT) ≥ 5 (0.381/0.953), Decaens ≥ 4 (0.676/0.817), and PCRS ≥ 3 (0.217/0.987). Among high-risk patients, recurrence reached 45% (95%CI: 35.1-57.0). Area under the curve comparisons showed no statistically significant differences among models. Thus, no model demonstrated clear superiority.

CONCLUSION

Although several explant-based models exist, their limited sensitivity suggests that many patients at risk of recurrence remain unidentified. The RETREAT score, developed in a large cohort, remains the most extensively validated. Future approaches should focus on developing improved prognostic tools using larger, preferably prospective datasets, and integrating artificial intelligence to enhance risk stratification and post-transplant surveillance.

Key Words: Risk factors score; Liver transplantation; Hepatocellular carcinoma; Neoplasm recurrence; Local

Core Tip: This systematic review and meta-analysis assessed the prognostic accuracy of three explant-based models-Risk Estimation of Tumor Recurrence after Transplant (RETREAT), Decaens, and Predicting Cancer Recurrence Score-in predicting hepatocellular carcinoma recurrence after liver transplantation. Analyzing nine retrospective studies with 5348 patients, recurrence was 7%, with no significant differences in area under the curve among models. While RETREAT is the most validated, all models demonstrated limited sensitivity, risking underestimation of recurrence risk. The findings highlight the need for improved prognostic tools, potentially through incorporating larger, prospective datasets and artificial intelligence, to enhance risk stratification, guide post-transplant surveillance, and optimize patient outcomes.

INTRODUCTION

Liver transplantation (LT) is the treatment of choice for patients with early-stage hepatocellular carcinoma (HCC) and decompensated cirrhosis and/or clinically significant portal hypertension, offering the possibility of curing both the tumor and the underlying liver disease[1,2]. However, HCC recurs in approximately 17% of patients after LT and remains the leading cause of death in this population[3,4]. Risk factors associated with recurrence include vascular invasion, tumor size > 5 cm, multifocal disease, and poor differentiation. These histopathologic and clinical characteristics are essential components in developing post-LT predictive models. Moreover, extrahepatic recurrence and early recurrence (within one year after LT) are associated with worse post-recurrence survival[5].

Several prognostic scoring systems have been proposed to guide surveillance after LT in patients transplanted for HCC. These models incorporate pre-transplant and post-transplant variables such as tumor size and number, microvascular invasion, and serum alpha-fetoprotein (AFP) levels[6]. Based on these factors, patients are stratified into different risk categories for HCC recurrence.

However, there are currently no evidence-based, standardized criteria for risk stratification that can be formally recommended in clinical practice. Therefore, the objective of this study was to identify which explant-based risk prediction model most accurately estimates the risk of HCC recurrence post-LT, to support the implementation of individualized surveillance strategies.

MATERIALS AND METHODS

Protocol and registration

This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews of Diagnostic Test Accuracy Studies guidelines[7]. The study protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO), No. CRD42024569977[8].

Eligibility criteria and outcomes

Studies were included based on the following predefined criteria: (1) Published in a peer-reviewed journal; (2) Included patients who underwent LT after the publication of the Milan criteria as listing criteria for HCC[9]; (3) Confirmed pathological diagnosis of HCC post-transplant; (4) Identified risk factors for HCC recurrence and developed or validated individual-level risk prediction models; (5) Reported at least one outcome of interest: Post-LT HCC recurrence rate or the predictive performance of models over a five-year follow-up; and (6) Provided data to calculate true positive (TP), false positive (FP), false negative (FN), and true negative (TN) rates. No language restrictions were applied. Studies were excluded if they assessed only pre-transplant data for recurrence prediction, or were in the form of abstracts, editorials, letters, narrative reviews, systematic reviews, or meta-analyses. The literature search covered all studies published until January 30, 2025.

Search strategy and study selection

A comprehensive literature search was conducted in PubMed, EMBASE, Web of Science, and the Cochrane Library using the following terms: ("liver transplantation" OR "hepatic transplantation") AND (HCC OR "hepatocellular carcinoma" OR hepatoma OR "liver cancer" OR "hepatic cancer" OR "liver neoplasia" OR "hepatic neoplasia") AND (recurrence OR relapse) AND ("predictive model" OR "predictive models" OR "AFP-R3" OR "R3-AFP" OR "UCLA nomogram" OR UCLA OR AFP OR "AFP score" OR RETREAT OR "Combo-MORAL" OR "Post-MORAL" OR MORAL).

The reference lists of eligible studies, prior systematic reviews, and meta-analyses were also manually screened for additional studies. Two reviewers (Christofoli de Barros I, Rodríguez Villafuerte S) independently conducted the search, imported citations into Zotero software (version 6.0.26, Vienna, Austria), and screened titles and abstracts. Full texts of potentially eligible studies were assessed independently. Discrepancies were resolved through discussion with a third reviewer (Vanzin Fernandes M).

Data extraction

Data extracted from each study included: (1) Study characteristics: First author, publication year, country, sample size, and follow-up duration; (2) Patient characteristics: Number of patients, age, liver disease etiology, prior HCC therapies, and transplant listing criteria; and (3) Outcomes: HCC recurrence rates and data on TP, TN, FP, and FN. For studies that assessed multiple models or thresholds, data were extracted separately for each model.

Risk of bias assessment

Two investigators (Christofoli de Barros I, Rodríguez Villafuerte S) independently assessed the methodological quality of the included studies using the revised Quality Assessment of Diagnostic Accuracy Studies tool (QUADAS-2)[10]. Disagreements were resolved with a third investigator (Vanzin Fernandes M). The QUADAS-2 tool evaluates four domains: (1) Patient selection; (2) Index test; (3) Reference standard; and (4) Flow/timing. Each domain was rated as having low, high, or unclear risk of bias. Results are summarized in Supplementary material.

Statistical analysis

Pooled sensitivity and specificity for each prognostic model were calculated using a random-effects model, along with 95%CI. Forest plots and summary receiver operating characteristic (ROC) curves were constructed using the bivariate method, and the area under the ROC curve (AUROC) was determined. When applicable, multiple thresholds were analyzed separately. Comparative analysis of area under the curve (AUC) across models was conducted using both direct and indirect bivariate approaches. Heterogeneity was assessed using the inconsistency index (I²). All statistical analyses were performed using R software version 4.3.2. A two-sided P < 0.05 was considered statistically significant.

RESULTS

Study characteristics

The initial search retrieved 4603 records (Figure 1). After removing duplicates and screening titles and abstracts, 68 full-text articles were reviewed. Nine retrospective studies, encompassing 5348 liver transplant recipients and evaluating three predictive scores, met the eligibility criteria for inclusion[11-19]. One study by Abdelfattah et al[20] was excluded due to overlapping population data with a previously published study from 2021[11].

Figure 1 Preferred Reporting Items for Systematic Reviews and Meta-Analyses flow diagram of study selection process.

All included studies focused on patients with confirmed HCC on explant analysis following LT. Patient characteristics are summarized in Table 1[11-19,21-32]. Most study participants were male, with a mean age of approximately 60 years. The leading etiologies of cirrhosis were chronic viral hepatitis (particularly hepatitis C virus) and alcohol-related liver disease. Six studies reported locoregional therapy use while patients were on the transplant waiting list. All transplant procedures used deceased donor grafts. The Milan criteria were the most applied for waitlist eligibility, although one center also used the University of California, San Francisco criteria, and another exclusively applied the revised United Kingdom criteria. All included studies were retrospective in design; six were conducted in Europe, two in North America, and one in Asia. Study-specific characteristics are detailed in Supplementary Table 1[11-19,21-32].

Table 1 Baseline characteristics of patients, and allocation criteria for hepatocellular carcinoma transplantation in the studies included in systematic review.

Ref.	Age (years)	Male (%)	Etiology of liver disease (%)	Hepatocellular carcinoma allocation criteria	Within Milan criteria (%)	Loco regional therapy1 (%)
Parfitt et al[24], 2007	54	84	HCV (42.7%), alcohol (21.3%), HBV (17.3%)	N/A	66.72	N/A
Chan et al[19], 2008	56.2	89.7	HCV (58%), HBV (33%), alcohol (17%)	N/A	N/A	N/A
Agopian et al[21], 2015	59.5	73.3	HCV (58%), HBV (16%), alcohol (9%)	Milan/UCSF	84	60
Mehta et al[12], 2017	58.2	75.2	HCV (58.8%), HBV (16.1%), MAFLD (8.0%)	Milan	50.32	91.5
Halazun et al[25], 2017	57.8	79.9	HCV (69.3%), HBV (15.3%), alcohol (11.8%)	Milan	66.6	81.7
Costentin et al[13], 2017	52.7	85.2	HBV or HCV (61.8%), alcohol (28.8%), others (9.4%)	Milan	64.22	67.7
Mehta et al[14], 2018	60	76.7	HCV (62.8%), MAFLD (8.0%), alcohol (7.3%)	Milan	852	91.1
Mirón Fernández et al[15], 2019	56.2	76.2	HCV (39%), mixed (alcohol/HCV): (22.9%), HBV (16.2%)	N/A	59.4	N/A
Feng et al[26], 2019	59	82.2	HBV (88.1%), HCV (5.9%), alcohol (2%)	Milan	79.22	75
Sánchez Segura et al[27], 2020	59	82.8	HCV (38.4%), alcohol (25.3%), mixed (alcohol/HCV): (23.2%)	N/A	62.6	94.9
Hasan et al[28], 2021	65.5	75	HCV (55.8%), MAFLD (25%), alcohol (5.8%)	Milan	100	58
Ma et al[29], 2021	51.98	92.1	HCV (54.9%), alcohol (23.5%), HBV (7.8%)	N/A	37.12	N/A
Abdelfattah et al[11], 2021	57.1	68.5	HCV (60.3%), HBV (26%), others (13.7%)	Milan	100	N/A
Åberg et al[16], 2021	57.4	79.3	HCV (59.2%), alcohol (28.4%), HBV (17.9%)	N/A	61.52	35.7
Aziz et al[30], 2021	62	89	HCV (46%), alcohol (11%), HBV (5%)	Milan/UCSF/total tumor volume < 15 cm3	49.3	68.5
Costentin et al[22], 2022	58	82.7	HBV or HCV (57.8%), alcohol (31.3%), others (10.8%)	Milan/alpha-fetoprotein score	62.32	68.5
Reddy et al[17], 2022	59	73.8	HCV (41.9%), alcohol (18.5%), HBV (11.2%)	Revised United Kingdom criteria	84.3	76.4
Van Hooff et al[18], 2022	61.2	79.8	HCV (20.7%), alcohol (16.3%), HBV (13.8%)	N/A	79.32	75.4
Cuadrado et al[31], 2023	60	87.9	Alcohol (34.8%), HCV (31.8%), mixed (alcohol/HCV): (19.7%)	N/A	98.5	77.3
Tran et al[23], 2023	58.0	77.9	HCV (61.6%), alcohol (10.7%), HBV (10.6%)	Milan	71.9	86.2
Brandão et al[32], 2024	60.5	71.9	HCV (87.5%), alcohol (9.4%), HBV (3.1%)	Brazilian Milan criteria	71.6	65.2

¹Loco regional therapy (includes downstaging in some studies).

²Based on explant pathology.

HBV: Hepatitis B virus; HCV: Hepatitis C virus; MAFLD: Metabolic dysfunction-associated fatty liver disease; N/A: Not available; UCSF: University of California, San Francisco.

Pooled analysis

Three risk models for post-transplant HCC recurrence were assessed across multiple studies and included in the meta-analysis: (1) Risk Estimation of Tumor Recurrence after Transplant (RETREAT), which incorporates AFP at LT, microvascular invasion, the diameter of the largest viable tumor, and the number of viable tumors[12]; (2) Decaens score, based on largest tumor size, number of nodules, and histologic differentiation grade[33]; and (3) Predicting Cancer Recurrence Score (PCRS), which includes explant pathology findings such as largest tumor size, macrovascular invasion, and bilobar involvement[19].

The overall pooled recurrence rate of HCC post-LT was approximately 7% (347/4976), excluding the study by Costentin et al[13], for which recurrence data were unavailable. Among patients classified as high-risk according to predictive models, the pooled recurrence rate was 45.87% (95%CI: 35.15-57.03, I² = 67%), as illustrated in Figure 2[13-17,28,31]. Score-specific recurrence rates for high-risk categories are also presented in Figure 2[13-17,28,31]. The bivariate analyses of sensitivity, specificity, and AUC for each model are presented in Table 2. Positive and negative predictive values are shown in Supplementary Table 3[11-32].

Figure 2 Recurrence rate of hepatocellular carcinoma. GLMM: Generalized Linear Mixed Model; N/A: Not applicable; PCRS: Predicting cancer recurrence score; RETREAT: Risk estimation of tumor recurrence after transplant.

Table 2 Sensitivity, specificity and area under the curve by bivariate analyses for different thresholds of the meta-analyzed scores.

Bivariate analysis
Score	Sensibility	Sensibility 95%CI	Specificity	Specificity 95%CI	AUC	AUC 95%CI
RETREAT ≥ 5	0.386	0.252-0.540	0.950	0.897-0.976	0.793	0.406-0.945
RETREAT ≥ 4	0.601	0.443-0.740	0.861	0.759-0.924	0.807	0.611-0.880
RETREAT ≥ 3	0.686	0.560-0.790	0.736	0.647-0.810	0.771	0.704-0.809
PCRS ≥ 1	0.582	0.306-0.815	0.851	0.797-0.893	0.852	0.521-0.905
PCRS ≥ 3	0.223	0.154-0.311	0.950	0.915-0.971	0.431	0.246-0.952
Decaens ≥ 4	0.591	0.329-0.810	0.821	0.597-0.934	0.772	0.518-0.889

AUC: Area under the curve; PCRS: Predicting cancer recurrence score; RETREAT: Risk estimation of tumor recurrence after transplant.

RETREAT score accuracy

For a RETREAT score ≥ 5, the pooled sensitivity and specificity were 0.381 (95%CI: 0.259-0.520, I² = 87%) and 0.953 (95%CI: 0.907-0.976, I² = 93%), respectively (Figure 3)[11,12,14,16-18]. For a score ≥ 4, the sensitivity and specificity were 0.603 (95%CI: 0.453-0.736, I² = 84%) and 0.864 (95%CI: 0.775-0.922, I² = 96%) (Figure 3)[11,12,14,16-18]. The corresponding AUC based on summary ROC curve analysis was 0.807 (95%CI: 0.611-0.880). For a score ≥ 3, pooled sensitivity and specificity were 0.766 (95%CI: 0.599-0.878, I² = 72%) and 0.743 (95%CI: 0.660-0.811, I² = 93%) (Figure 3)[11,12,14,16-18]. Diagnostic performance (AUC and 95%CI) for all RETREAT thresholds is summarized in Tables 2 and 3.

Figure 3 Sensitivity and specificity. A: Sensitivity risk estimation of tumor recurrence after transplant (RETREAT); B: Specificity RETREAT. FN: False negatives; FP: False positive; RETREAT: Risk estimation of tumor recurrence after transplant; TP: True positives.

Table 3 Comparison matrix of area under the curves scores.

Score	RETREAT ≥ 5	RETREAT ≥ 4	RETREAT ≥ 3	PCRS ≥ 1	PCRS ≥ 3	Decaens ≥ 4
RETREAT ≥ 5	-	0.963	0.860	0.809	0.941	0.913
RETREAT ≥ 4	0.963	-	0.632	0.697	0.956	0.779
RETREAT ≥ 3	0.860	0.632	-	0.382	0.931	0.948
PCRS ≥ 1	0.809	0.697	0.382	-	0.903	0.545
PCRS ≥ 3	0.941	0.956	0.931	0.903	-	0.940
Decaens ≥ 4	0.913	0.779	0.948	0.545	0.940	-

PCRS: Predicting cancer recurrence score; RETREAT: Risk estimation of tumor recurrence after transplant.

Decaens score accuracy

For a Decaens score ≥ 4, the pooled sensitivity and specificity were 0.676 (95%CI: 0.581-0.759, I² = 61%) and 0.817 (95%CI: 0.666-0.909, I² = 90%), respectively (Figure 4)[13,15,19]. AUROC values are provided in Tables 2 and 3.

Figure 4 Sensitivity and specificity of predicting cancer recurrence score and Decaens. A: Sensitivity; B: Specificity. FN: False negatives; FP: False positive; PCRS: Predicting cancer recurrence score; TP: True positives.

PCRS score accuracy

For a PCRS score ≥ 1, the pooled sensitivity was 0.543 (95%CI: 0.288-0.777, I² = 64%) and specificity was 0.840 (95%CI: 0.799-0.875, I² = 0%) (Figure 4)[13,15,19]. For a score ≥ 3, sensitivity dropped to 0.217 (95%CI: 0.149-0.305, I² = 0%) while specificity increased to 0.987 (95%CI: 0.740-0.999, I² = 0%) (Figure 4B)[13,15,19]. Corresponding AUROC values are summarized in Tables 2 and 3.

Risk of bias and quality assessment

The risk of bias across individual studies, as evaluated using the QUADAS-2 tool, is presented in Supplementary Table 2.

DISCUSSION

To the best of our knowledge, this is the first systematic review and meta-analysis to evaluate the prognostic performance of predictive models based on explant pathology data for estimating HCC recurrence following LT. A total of nine studies and three predictive scores-RETREAT, Decaens, and PCRS-were included[12,19,33]. These models were generally constructed using multivariable analyses, incorporating variables with significant associations (P < 0.05) in univariable analysis. The most frequently used predictors included AFP, tumor size and number, histological differentiation, and vascular invasion.

Across the included studies (excluding Costentin et al[13], which did not report recurrence rates), the pooled HCC recurrence rate post-LT was approximately 7% (347/4976). Recurrence rates varied significantly, from 4.4% in a large United States cohort of 3276 patients[14] to 20.1% in a smaller Swedish cohort of 169 patients[15].

The PCRS (or Chan score), developed in 2008, was derived from a cohort of 116 United States patients who underwent LT with HCC found on explant-18.9% of which were incidental tumors. This relatively small and heterogeneous cohort may limit generalizability. The score includes four explant variables: (1) Diameter of the largest nodule; (2) Bilobar involvement; (3) Macrovascular invasion; and (4) Tumor differentiation. The score ranges from -3 to 6, stratifying patients into low (≤ 0), intermediate (1-2), and high (≥ 3) risk categories. In the original cohort, no recurrences occurred in the low-risk group, while recurrence rates were 19.3% and 66.7% in the intermediate-risk and high-risk groups, respectively[19]. In our meta-analysis, patients with PCRS ≥ 3 had a recurrence probability of 66% within five years. The sensitivity and specificity for PCRS > 3 were 22.3% (95%CI: 15.4%-31.1%) and 95.0% (95%CI: 91.5%-97.1%), respectively, indicating high specificity but poor sensitivity. The AUROC for this cutoff was 0.431 (95%CI: 0.246-0.952), suggesting limited overall performance and considerable uncertainty due to a wide confidence interval. The limited sample size and heterogeneity of the included cohort, combined with the scarcity of external validation studies, may undermine the reliability of the pooled estimates. The broad confidence intervals and suboptimal sensitivity suggest that PCRS should not be used in isolation for clinical decision-making. Rigorous prospective validation is necessary prior to its routine clinical implementation.

The Decaens score, published in 2011, was developed to improve LT candidate selection and was intended for pre-transplant use. It was derived from 373 patients transplanted in France between 1988 and 1998. The score incorporates the number of nodules, the maximum tumor diameter (based on imaging at listing), and tumor differentiation-requiring biopsy, which was performed in only 11% of the cohort. Thus, tumor grading was ultimately based on explant pathology. Patients with scores > 4 were at higher risk of recurrence. In the original cohort, 5-year recurrence was 46% for scores ≥ 4 and 20.8% for scores < 4[33]. In our meta-analysis, high-risk patients (score ≥ 4) had a 50% recurrence rate within five years. Sensitivity and specificity were 0.591 (95%CI: 0.329-0.810) and 0.821 (95%CI: 0.597-0.934), respectively, with an AUC of 0.772-indicating acceptable discriminative ability. However, the wide confidence interval for sensitivity warrants caution. Moreover, this score is not applicable to patients with nonviable tumors, limiting its clinical utility.

The RETREAT score, developed from a cohort of 721 United States patients who underwent LT within the Milan criteria between 2002 and 2012, incorporates AFP at transplant, microvascular invasion, and the sum of the largest viable tumor diameter and number of viable tumors. It stratifies recurrence risk across six score levels. In the development cohort, 5-year recurrence increased from 2.9% (score 0) to 75.2% (score ≥ 5), outperforming the Milan criteria[12]. Validation using the UNOS database (3276 patients) confirmed its superior classification performance, as measured by the Net Reclassification Index (NRI) (NRI: 0.28, P < 0.001)[14]. However, the NRI has recently been criticized[34]. In this validation cohort, recurrence at three years ranged from 1.6% (score 0) to 29.0% (score > 5). It is worth noting that the median follow-up in the validation cohort was only 1.9 months, potentially underestimating recurrence compared to the development cohort (11.6% vs 4.4%).

Our analysis revealed that the RETREAT score's sensitivity and specificity varied by cutoff level (Supplementary Table 1)[11-19,21-32]. For RETREAT ≥ 5, specificity was 95.3% (95%CI: 90.7%-98.5%) but sensitivity was low. Lowering the cutoff to ≥ 4 increased sensitivity to 57.3% and reduced specificity to 84.3%; at ≥ 3, sensitivity rose to 77.4%, while specificity declined further to 66.4%. This trade-off highlights the model's high negative predictive value, supporting its utility in identifying patients at low risk who may benefit from less intensive surveillance.

All three scores demonstrated high specificity (≥ 74%), particularly RETREAT, supporting their value in excluding low-risk individuals. Positive predictive value (PPV) increased with higher score thresholds, reinforcing their utility in risk stratification. For instance, RETREAT’s PPV rose from 27.3% (≥ 3) to 51.4% (≥ 5), and PCRS ≥ 3 achieved the highest PPV at 65.7% (95%CI: 48.8%-79.4%).

Although the models show limited sensitivity at higher thresholds, their clinical utility remains evident, especially in identifying high-risk individuals. The Decaens score (≥ 4) provides a balance of sensitivity and specificity, making it a robust tool among those evaluated. These models-despite relying primarily on morphometric and surrogate biological markers (e.g., AFP)-can guide targeted post-transplant strategies.

There remains debate over whether surveillance should be universal or limited to high-risk patients. Current guidelines-including the American Association for the Study of Liver Diseases 2023 Practice Guidance and the International Liver Cancer Association-International Liver Transplantation Society (ILTS) 2024 consensus-recommend surveillance for patients with intermediate to high recurrence risk[2,35]. The RETREAT score is the most widely validated tool in this context[12,14,16-18].

Despite a lack of evidence-based protocols, existing data suggest that closer surveillance improves outcomes and access to curative therapies[36]. The ILTS recommends abdominal/chest imaging and AFP every 3 months for the first 3 years[35], while the National Comprehensive Cancer Network suggests imaging and AFP every 3-6 months for 2 years, then every 6-12 months[37].

Importantly, patients transplanted for HCC may have distinct immunologic profiles, influencing both rejection risk and recurrence[38]. Gene expression signatures hold promise for personalized post-LT management[38,39].

This meta-analysis has limitations. All included studies were retrospective in nature, introducing inherent risks of selection and reporting biases-particularly with respect to patient inclusion criteria, data completeness, and follow-up duration. These limitations constrain the generalizability of our findings and underscore the need for prospective, multicenter validation cohorts to confirm the performance of these models under standardized conditions. Differences in patient characteristics, inclusion criteria, and outcome definitions may have contributed to heterogeneity. The overall recurrence rate (7%) is lower than in other large cohorts, potentially due to underreporting or short follow-up. None of the studies included patients undergoing living donor LT; however, a systematic review found no significant recurrence rate differences based on donor type[40]. We could not include promising models such as COMBO-MORAL[21], R3-AFP[22], and RELAPSE[23] due to insufficient data, which limits the scope of our conclusions and highlights the need for future validation in diverse populations.

Despite these limitations, consistent methodology strengthens the validity of our findings. This review provides clinically relevant insights into recurrence risk stratification using explant-derived models.

CONCLUSION

Patients with higher scores based on explant data exhibit a markedly elevated risk of HCC recurrence within five years post-transplant. Although current guidelines do not endorse a specific model for surveillance, our findings support their utility in guiding follow-up. Deep learning model has been applied in this context. While AUROC comparisons do not establish the clear superiority of one model, high-risk thresholds-RETREAT ≥ 5 (37% recurrence), Decaens ≥ 4 (50%), and PCRS ≥ 3 (66%)-clearly identify patients who require intensive surveillance. These scores may enable better resource allocation, enhancing cost-effectiveness and early recurrence detection. Low-risk patients may benefit from less frequent follow-up. Our review also highlights the need to integrate histopathological, immunological, and treatment-related variables into future predictive models, ideally in large prospective studies using artificial intelligence to improve stratification and personalize care.

ACKNOWLEDGEMENTS

The authors thank Hospital Vozandes Quito-HVQ, Quito, Ecuador, for their support in the publication of this article.