A Noninvasive Predictive Model of Portal Hypertension in Patients with HCC Based on Clinical Features and Intra- and Peritumoral Pre-Fusion Radiomic Features

Introduction

Portal hypertension (PHT) is a common and severe complication of liver cirrhosis and other chronic liver diseases, which significantly impacts patients’ quality of life and prognosis.^1–3 Clinically Significant Portal Hypertension (CSPH) is defined as a critical predictor of post-hepatectomy liver decompensation and mortality.⁴ Furthermore, emerging evidence suggests that PHT may also be associated with an increased risk of hepatocellular carcinoma (HCC) recurrence following surgery.⁵ Consequently, accurate prediction of PHT is paramount for optimizing clinical decision-making and improving patient outcomes. Currently, the measurement of the hepatic venous pressure gradient (HVPG) is regarded as the gold standard for assessing PHT.⁶ However, its invasive nature, procedural complexity, and cost limit its clinical utility in China.^7,8 Additionally, HVPG reflects sinusoidal pressure rather than direct portal venous pressure and does not fully represent the pressure in the main portal vein.⁹ Although intraoperative direct puncture of the portal system provides reliable free portal pressure (FPP) measurements, it is exclusively feasible during surgery. This underscores the urgent need to develop effective non-invasive screening techniques.

In recent years, significant progress has been made in various non-invasive methods for predicting portal hypertension. Multiparameter models based on spleen stiffness measurement (SSM), such as the NICER model, have demonstrated high diagnostic accuracy in identifying clinically significant portal hypertension and provide a reliable basis for clinical decision-making through defined thresholds (SSM <21 kPa to rule out, >50 kPa to rule in).¹⁰ Novel functional imaging techniques, represented by harmonic-assisted pressure estimation (SHAPE), have achieved a transition from “surrogate markers” to “direct pressure measurement,” showing good correlation with the gold standard hepatic venous pressure gradient (HVPG).¹¹ Nevertheless, these emerging technologies have not yet achieved widespread clinical translation and adoption in China.

Advances in medical imaging provide indispensable tools for disease assessment. Radiomics, as a high-throughput quantitative analysis method, enables the extraction of vast amounts of feature information from routine medical images such as CT and MRI.^12–15 These features can reflect the phenotypic and biological heterogeneity of tumors,^16–18 and have been successfully applied for high-accuracy prediction of tumor-related prognoses.^19,20 Compared to traditional radiological diagnostics, radiomics demonstrates significant potential for improving the accuracy of disease prediction and diagnosis,^21,22 particularly in the field of liver tumors, where its features have been proven to correlate closely with clinical outcomes such as patient survival and treatment response.^23–25 Methodologically, LASSO regression has become widely used for dimensionality reduction in high-dimensional radiomic data, effectively preventing overfitting by penalizing irrelevant features.^26,27 Feature fusion strategies, ranging from early fusion to late fusion, have emerged as powerful approaches for integrating complementary information from different regions of interest.²⁸ Despite these methodological advances, their application to PHT prediction remains limited.

The development of PHT is closely linked to structural remodeling of the liver that increases intrahepatic vascular resistance.^4,29,30 These pathological changes, including fibrosis, microvascular distortion, and altered hemodynamics, can be captured as quantifiable radiomic features on routine CT scans.^31,32 In HCC, the peritumoral region represents a dynamic interface where tumor cells interact with adjacent liver parenchyma, inducing stromal reactions, inflammation, and neoangiogenesis that contribute to local hemodynamic disturbances and elevated portal pressure.³³ Radiomic features from this zone may reflect microarchitectural disorganization, collagen deposition, and abnormal vascular patterns.³⁴ In contrast, intratumoral radiomic features primarily capture tumor-intrinsic characteristics such as cellular density, necrosis, and heterogeneity, which may reflect tumor aggressiveness and aberrant intratumoral vascular architecture, thereby indirectly influencing portal pressure.³⁵ Thus, radiomic features derived from these two regions offer complementary biological insights, yet research utilizing both to predict PHT remains relatively scarce.

Notably, the portal venous phase was selected for radiomic analysis in this study because it provides optimal contrast between the tumor and surrounding liver parenchyma, minimizes motion artifacts, and most accurately reflects hemodynamic alterations associated with PHT. In contrast, the arterial phase is more susceptible to variations in cardiac output and bolus timing, while the delayed phase offers less distinct tumor-liver delineation.^34,36

To more fully exploit imaging information, this study introduces pre-fusion radiomic features extracted from portal venous phase CT images, covering both intra- and peritumoral regions, to develop a multimodal model for non-invasive PHT prediction in combination with clinical features. Pre-fusion radiomic features are defined as an early fusion strategy that concatenates normalized radiomic features from the two regions before dimension reduction or classifier input, differing from standard feature-level fusion and late fusion. We hypothesize that the multimodal model integrating clinical features with intra- and peritumoral radiomic features can achieve optimal predictive performance, surpassing any single-feature model. The primary objectives of this study were to develop and validate a non-invasive PHT prediction model integrating clinical factors with radiomic features, and to compare its predictive performance against models based solely on clinical features or radiomic features alone.

Materials and Methods

Study Population

Based on initial screening and application of predefined inclusion and exclusion criteria, a total of 884 patients were enrolled in this retrospective study. Ethical approval was granted by the Ethics Committee of the First Affiliated Hospital of Guangxi Medical University (Approval No.: 2023-E488-01). Portal hypertension (PHT) was defined as a portal venous pressure (PVP) level ≥ 25 cmH₂O.³⁷ Patients were randomly divided into training and validation cohorts in an 8:2 ratio (training: n=707, 89 with PHT; validation: n=177, 23 with PHT).

Inclusion and Exclusion Criteria

Inclusion criteria: (1) having undergone partial hepatectomy with direct intraoperative PVP measurement at the First Affiliated Hospital of Guangxi Medical University between January 2013 and January 2020; (2) aged ≥ 18 years and < 70 years; and (3) postoperative pathological confirmation of hepatocellular carcinoma (HCC).

Exclusion criteria: (1) history of postoperative HCC recurrence, tumor rupture with hemorrhage, previous upper abdominal surgery, any preoperative interventional therapy (such as Transarterial Chemoembolization [TACE] or Hepatic Arterial Infusion Chemotherapy [HAIC]), portal vein ligation, splenectomy, and/or splenic artery ligation; (2) presence of portal vein tumor thrombus, splenic vein tumor thrombus, any macrovascular invasion, or arteriovenous fistula; (3) absence of a complete preoperative contrast-enhanced CT examination; or (4) lack of essential preoperative laboratory tests, including complete blood count, liver function tests, and coagulation profile (Figures 1 and 2).

Figure 1 Flow chart for patient selection.

Figure 2 Flow chart for technical roadmap. (A) Subgroups of study population; (B) radiomics feature selection; (C) Clinical feature selection; (D) Nomogram.

Clinical Data Collection and Follow-Up

Patient clinical data were collected retrospectively using the hospital’s electronic medical record system, outpatient/imaging systems, and telephone follow-up. The follow-up concluded on December 31, 2023. Cases with missing key variables were excluded. All missing data were assessed for potential impact and handled appropriately to minimize bias in the analysis.

Pre-Fusion Radiomic Features

We define intra- and peritumoral pre-fusion radiomic features as the concatenation of radiomic features extracted separately from the intratumoral ROI and the peritumoral ROI after z-score normalization, forming a single fused feature vector representing both regions. This constitutes an early fusion strategy, in contrast to standard feature-level fusion by simple post-extraction concatenation without joint normalization, and to model-level late fusion that combines separate models’ outputs. The fused features are then subjected to the same screening pipeline (t-test, correlation threshold, and LASSO) and used for classifier training.

Contrast-Enhanced CT Scanning Protocol

CT examinations were performed using GE LightSpeed VCT (64-slice) and Siemens dual-source CT scanners. Non-ionic iodinated contrast agent (300 mgI/mL) was administered intravenously via an antecubital vein using a power injector at a flow rate of 2.5 mL/s, with a bolus dose of 3 mL/kg body weight. Sequential helical scans were acquired with an 8-mm slice thickness at the following phases: arterial phase (22–30s delay), portal venous phase (50–55s delay), and delayed phase (180s delay). Scanning parameters included a tube voltage of 120 kV, beam current of 280 mAs, and gantry rotation speed of 0.5 seconds per rotation. Images were stored in Digital Imaging and Communications in Medicine format within the Picture Archiving and Communication System.

CT Image Analysis and Measurement

Preoperative CT images were retrieved for analysis. Portal vein diameter (PVD) was measured at the confluence of the left and right portal branches, and splenic vein diameter (SVD) was measured at the proximal origin of the splenic vein; the average of two independent measurements was recorded for each. Liver volume (LV) and spleen volume (SV) were calculated using the Total Segmentation plugin within the 3D-Slicer software (www.slicer.org). Regions of interest (ROIs) encompassing liver tumors and non-tumorous liver parenchyma were manually delineated using ITK-SNAP software. The volumes of these ROIs were computed, and the percentages of tumor volume and non-tumorous liver volume relative to the total actual liver volume were subsequently determined.

Portal Venous Pressure (PVP) Measurement

Following induction of general anesthesia and within the first hour of laparotomy, a segment of polyvinyl chloride catheter was introduced into a right gastroepiploic vein via the greater omentum. The catheter was connected to a saline-filled manometer, with the external zero reference level positioned at the mid-axillary line. Direct and reliable PVP readings were obtained once the fluid column stabilized. Portal hypertension (PHT) was defined as a PVP level ≥ 25 cmH₂O.³⁷

Radiomics Feature Extraction of Intratumor and Peritumor

Initial image preprocessing was performed using Python. Regions of interest (ROIs) encompassing liver tumors were manually delineated on portal venous phase CT images using ITK-SNAP software (version 4.0; http://www.itksnap.org/pmwiki/pmwiki.php). Intratumoral ROIs were defined by manually tracing the tumor boundary on each slice, with careful exclusion of large vessels, necrosis, and adjacent non‑tumor structures to ensure precise delineation of tumor parenchyma. For peritumoral ROIs, a semi‑automated approach was adopted: the intratumoral ROI was dilated outward by 5 mm using morphological operations to generate an annular region around the tumor. Following automated dilation, all peritumoral ROIs were manually inspected and corrected by an experienced radiologist; any portions extending beyond the liver capsule into non‑hepatic tissues were meticulously erased using ITK‑SNAP’s editing tools, ensuring that only hepatic tissue within the peritumoral region was retained. This hybrid approach combining automated dilation with manual refinement ensured both procedural efficiency and anatomical accuracy. After ROI segmentation, a comprehensive set of radiomic features, including shape, intensity, texture, and wavelet features, was extracted using the PyRadiomics library (version 3.0.1; https://pyradiomics.readthedocs.io/en/latest/), converting visual information into quantifiable data.^20,38,39

To ensure methodological rigor, intraclass correlation coefficients (ICCs) were used to quantify intra‑ and inter‑observer variability. Two senior abdominal radiologists (with 8 and 10 years of experience, respectively) independently segmented ROIs for 30 randomly selected CT cases. The first radiologist repeated the segmentation after a 2‑week interval to assess intra‑observer agreement, while the second radiologist independently segmented the same cases for inter‑observer agreement. Reliability was calculated using a two‑way random‑effects absolute agreement model (ICC(2,1)) with the R package irr. The analysis yielded a median intra‑observer ICC of 0.92 (range: 0.81–0.98) and a median inter‑observer ICC of 0.88 (range: 0.75–0.96), with shape features (ICC = 0.95) and first‑order statistical features (ICC = 0.93) demonstrating the highest reproducibility (Table S1). Only radiomic features with an ICC greater than 0.8 (indicating good to excellent reproducibility) were retained for subsequent analysis.⁴⁰

All radiological operations, including preprocessing, delineation, and feature extraction, were performed on portal venous phase CT images with consistently applied standardized window settings (width: 400–500 HU, level: 20–50 HU).^17,41,42 All extracted features underwent Z‑score normalization to ensure data consistency.⁴³ This standardized workflow, compliant with radiomics quality assurance guidelines, effectively minimized operator‑dependent variability. Because CT images were acquired from two different scanner platforms (GE LightSpeed VCT and Siemens dual‑source CT), we applied ComBat harmonization to correct for batch effects and scanner‑related variability.⁴⁴ Specifically, after feature extraction and before feature screening, the ComBat algorithm was implemented using the neuroCombat R package, with scanner type treated as the batch covariate. This step ensured that subsequent feature selection and model building were based on harmonized radiomic features, thereby improving the robustness and generalizability of the model.

Radiomics Feature Screening

All radiomic feature selection steps were performed exclusively within the training cohort to prevent data leakage. Before statistical screening, all radiomic features were standardized using z‑score normalization based on the mean and standard deviation of the training cohort alone, ensuring that features with different scales contributed equally to subsequent analyses. The same normalization parameters were then applied to the validation cohort. The screening pipeline proceeded as follows. First, univariate analysis using independent t-tests identified features with significant differences between the PHT and non-PHT groups (P < 0.05), reducing the feature space by eliminating features with low discriminative power. Second, pairwise correlation analysis was performed to reduce redundancy: features with a Pearson correlation coefficient > 0.9 were considered highly correlated, and in each such pair, the feature with the smaller P-value (stronger univariate association) was retained while the other was excluded. Finally, LASSO regression was applied to the training set. LASSO imposes an L1 penalty on coefficients, shrinking irrelevant ones to zero for simultaneous feature selection and regularization. The regularization parameter λ controls the penalty strength. To identify the optimal λ, ten-fold cross-validation was performed within the training cohort. The λ that minimized the mean cross-validation error (binomial deviance) was selected, and for model parsimony, we adopted the one-standard-error rule, choosing the largest λ within one standard error of the minimum error. This yielded a compact feature set without statistically significant loss of predictive performance. Features with non-zero coefficients were retained.^45,46

Machine Learning Predictive Model Building

Following feature selection, the predictive models were constructed in a step-by-step chronological order: first, we trained three machine learning models—Logistic Regression (LR), k-Nearest Neighbors (KNN), and Random Forest (RF)—using the selected radiomic features and clinical features exclusively on the training cohort; second, we performed hyperparameter tuning using cross-validation within the training cohort; third, we evaluated the performance of the trained models on the independent validation cohort (which had not been used in any feature selection or model training step). To assess the additional value of radiomic features in clinical risk factor prediction, we combined these features with clinical factors and compared them using the same machine learning models in logistic regression analysis. The performance of the models was evaluated by ROC curves, calibration curves and clinical decision curves on the training and validation sets to measure the accuracy, calibration and clinical applicability of the models, respectively.

Statistical Analysis

Continuous variables are expressed as mean ± SD for normally distributed data, or as median (IQR) for non-normally distributed data, based on the Shapiro–Wilk test. Group comparisons for continuous variables were performed using independent samples t-tests (normally distributed data) or Mann–Whitney U-tests (non-normally distributed data), with corresponding t-values or U-values reported. Categorical variables were presented as frequencies (percentages), and group comparisons were conducted using the Chi-square test or Fisher’s exact test, as appropriate. All statistical analyses were performed using R (version 4.2.0) and Python (version 3.9.12). A two-sided P-value < 0.05 was considered statistically significant, with corresponding χ²-values reported.

To ensure adequate statistical power, we performed a sample size estimation before analysis. Following the “events per variable” (EPV) guideline, which recommends at least 10 events per candidate predictor to avoid overfitting,⁴⁷ we noted that the training cohort contained 89 PHT patients. The final logistic regression model included three clinical predictors (portal vein diameter, Child-Pugh score, and FIB-4 score) along with a radiomic signature derived from LASSO-selected features, with the effective number of parameters constrained by regularization. Even under a conservative scenario assuming up to 10 candidate predictors, the EPV ratio exceeded 8.9, approaching the recommended threshold of 10. Moreover, studies suggest that regularization techniques like LASSO can relax EPV requirements, as they inherently reduce model complexity and mitigate overfitting.⁴⁸ Nevertheless, we acknowledge that the limited number of PHT events (n=89) may restrict the complexity of reliably developed models, which guided our decision to prioritize parsimonious models with few clinical predictors and a single composite radiomic score rather than fitting models with many individual radiomic features.

To handle class imbalance in the training cohort (PHT prevalence 12.6%), we implemented a threefold strategy: class weighting for logistic regression and random forest, SMOTE-based oversampling for KNN, and threshold optimization using the Youden index. All procedures were confined to the training cohort and performed within cross-validation.

Results

Description of Baseline Information

In this section, we compared the preoperative clinical baseline data between PHT and non-PHT groups in the training and validation sets, respectively. The analysis revealed that in both datasets, statistically significant differences (P < 0.05) between PHT and non-PHT groups were observed for spleen volume (SV), portal vein diameter (PVD), splenic vein diameter (SVD), platelet count (PLT), hepatitis B virus infection, number of liver segments invaded by the tumor, Child-Pugh score, FIB-4 score, and tumor stage. Additionally, in the training set, statistically significant differences (P < 0.05) were also found for International Normalized Ratio (INR), albumin (Alb), total bilirubin (Tbil), and aspartate aminotransferase (AST). No statistically significant differences were found for the remaining general preoperative clinical characteristics (Table 1).

Table 1 Clinical Baseline Data of PHT and Non PHT for Training and Validation Sets

Clinical Characteristics Screening

All preoperative clinical data were analyzed by one-way logistic regression, and variables with P<0.05 were included in the multifactorial logistic regression analysis to ultimately screen for variables with OR >1 and P<0.05. Finally, PVD (OR:1.027, 95% CI [1.015,-1.04], P<0.001), Child-Pugh score (OR:1.132, 95% CI [1.05–1.221], P=0.01), FIB-4 score (OR:1.087, 95% CI [1.048–1.126], P< 0.001) were identified as independent risk factors for PHT (Table 2).

Table 2 Univariate Logistic Regression and Multivariate Logistic Regression

Screening of Radiomic Features

A total of 1834 features are extracted, and the number of each feature is (1) First-Order Statistics (FOS) features:360, (2) Gray Level Co-occurrence Matrix (GLCM) features:440, (3) Gray Level Run Length Matrix (GLRLM) features:320, (4) Gray Level Size Zone Matrix (GLSZM) features:320, (5) Neighboring Gray Tone Difference Matrix (NGTDM) features:100, (6) Gray Level Dependence Matrix (GLDM) features:280, and (7) SHAPE features:14.

The final filtered radiomic features and their corresponding weights (Figure 3 and Tables S2–S4).

Figure 3 Flow chart for screening of radiomics features. (A) Intratumoral radiomics features; (B) Peritumoral radiomics features; (C) Intra- and peritumoral pre-fusion radiomic features.

Performance Evaluation of Machine Learning Algorithmic Models

Three machine learning algorithms—Logistic Regression (LR), k-Nearest Neighbors (KNN), and Random Forest (RF)—were constructed using the selected features, and their predictive performance was comparatively evaluated. Performance metrics assessed across both training and validation sets included accuracy, area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), precision, recall, and the F1-score. Among the classifiers evaluated using the three distinct feature sets (intratumoral, peritumoral, and intra- and peritumoral pre-fusion) for predicting PHT in the training set, the KNN classifier consistently achieved the highest AUC values, demonstrating superior classification performance (AUCs: 0.897, 95% CI [0.8746–0.9195] for intratumoral; 0.87, 95% CI [0.8442–0.8956] for peritumoral; and 0.889, 95% CI [0.8669–0.9117] for the intra- and peritumoral pre-fusion radiomic features (Figure 4 and Tables S5–S7).

Figure 4 AUC of different radiomics features in different machine algorithms. (A) Intratumoral radiomics features (Training group); (B) Intratumoral radiomics features (Validation group); (C) peritumoral radiomic features (Training group); (D) peritumoral radiomic features (Validation group); (E) intra- and peritumoral pre-fusion radiomic features (Training group); (F) intra- and peritumoral pre-fusion radiomic features (Validation group).

Performance of Different Feature Sets Under KNN Classifier

The predictive performance of each feature model was evaluated in comparison to the KNN method. The predictive performance of intra- and peritumoral pre-fusion radiomic features was comparable to that of intratumor features alone and superior to peritumor alone in the training group, with an AUC of 0.889 (95% CI: 0.8669–0.9117). Furthermore, it exhibited a more favorable performance than intratumor or peritumor alone in the validation group, with an AUC of 0.684 (95% CI: 0.5721–0.7952) (Table 3 and Figure 5).

Table 3 Comparison of clinical features and radiomics feature models

Figure 5 ROC curves of each feature model. (A) Training group; (B) Validation group.

A Feature-Front Fusion Linear Model Based on Clinical Features and Intra- and Peritumoral Pre-Fusion Radiomic Features of Liver Tumors

For the final combined clinical-radiomic model, we selected Logistic Regression (LR) due to its distinct advantages for clinical application. LR produces a linear and interpretable model, allowing the contribution of each feature to be quantified through coefficients. Furthermore, it is less susceptible to overfitting with limited medical datasets compared to more complex algorithms. Finally, its structure readily supports the development of a clinically practical nomogram.

Using this approach, clinical characteristics were combined with intra- and peritumoral pre-fusion radiomic features to develop a novel linear prediction model, designated as the nomogram. The prediction performance of each model under the KNN algorithm was subsequently evaluated. The nomogram exhibited the optimal performance in both the training and validation groups, with AUCs of 0.938 and 0.847, respectively, along with enhanced sensitivity. The calibration curve demonstrated good calibration, showing high consistency between the predicted probability and the actual observed probability, which indicates the high prediction accuracy of the model. Furthermore, clinical decision curve analysis confirmed that our model achieved exceptional performance, delivering a high net benefit across a range of threshold settings (Table 4 and Figure 6).

Table 4 Comparison of clinical characteristic model, radiomics feature model and nomogram

Figure 6 Performance evaluation of the combined clinical-radiomic nomogram model. (A) ROC curve of different models (Training group); (B) ROC curve of different models (Validation group); (C) Calibration curve of different models (Training group); (D) Calibration curve of different models (Validation group) (E) Clinical Decision Curve Analysis of different models (Training group); (F) Clinical Decision Curve Analysis curve of different models (Validation group). Note: Clinic clinical features; ALL Intra- and peritumoral pre-fusion radiomic features.

Discussion

This study constructed and validated a nomogram model based on routine clinical indicators (main portal vein diameter, Child‑Pugh score, and FIB‑4 score) combined with intra‑ and peritumoral pre‑fusion radiomic features for the noninvasive prediction of portal hypertension (PHT) risk in patients with hepatocellular carcinoma (HCC). The model achieved AUCs of 0.938 and 0.847 in the training and validation cohorts, respectively, significantly outperforming the baseline model that relied solely on clinical indicators (AUC: 0.733). This finding aligns well with current trends in the field and indicates that integrating intra‑ and peritumoral radiomic features with clinical data can enhance predictive performance and improve result stability. For instance, a study on intrahepatic cholangiocarcinoma utilized intratumoral and peritumoral magnetic resonance imaging (MRI) features to predict tumor recurrence and found that the combined model integrating radiomic features (intra- plus peritumoral) with clinical factors achieved the best predictive performance, with an AUC of 0.852 (95% CI, 0.724–0.937) for early and late recurrence.⁴⁹ Another study focusing on stage II/III gastric cancer developed a radiomic signature from contrast‑enhanced CT scans that included both intratumoral and peritumoral regions; when combined with clinical features, this signature proved to be a robust predictor of disease‑free survival and identified patients who would benefit from adjuvant chemotherapy.⁵⁰ Furthermore, a study on preoperative Lauren classification of gastric cancer using intratumoral and peritumoral radiomic analysis also showed that the model combining radiomic features with clinical factors outperformed other models, achieving an AUC of 0.745 (95% CI, 0.696‑0.795) in the training set and 0.758 (95% CI, 0.685‑0.831) in the validation set.⁵¹ Together, these results consistently demonstrate that a multi‑regional fusion strategy is more reliable than single‑region models.⁵²

A key and noteworthy finding of this study is that a model based solely on the three clinical indicators—portal vein diameter, Child‑Pugh score, and FIB‑4 score—already exhibited superior predictive performance compared to any model relying solely on radiomic features (intratumoral, peritumoral, or their combination). Rather than being a secondary observation, this finding critically underscores that portal hypertension is essentially a systemic consequence of hemodynamic and structural remodeling throughout the liver in the context of chronic liver disease, a perspective that aligns with the integrated pathophysiology of the condition. One plausible explanation is that these clinical indicators inherently capture the cumulative effects of hepatic architectural distortion and hemodynamic alterations, which may overlap substantially with the information encoded in radiomic features.⁵³ Additionally, the relatively limited sample size may also restrict the full exploitation of high-dimensional radiomic features, potentially contributing to the observed performance gap.⁵⁴ In line with this, previous studies have suggested that deep learning features tend to demonstrate their capacity for characterizing complex patterns more effectively in large-scale datasets, whereas in moderate-sized cohorts, models based on clinical variables—which are structurally simpler and more stable—often exhibit superior performance.⁵⁵ Consequently, the marginal added value of radiomics may reflect such informational redundancy and data constraints rather than a lack of biological relevance. Under this premise, the role of radiomics is not replacement but rather a valuable complement to the overall clinical assessment. Furthermore, while the intratumoral model showed a marginal advantage in the training cohort, the “intra- and peritumoral pre-fusion” strategy adopted in this study demonstrated superior stability in the validation cohort. This shift in performance across cohorts may reflect that the fusion strategy, by capturing synergistic information from the tumor core and its surrounding parenchyma, more effectively captures diffuse intrahepatic pathological changes that are inherently variable in clinical settings.^34,56 This implies that radiomic features extracted from the peritumoral region may capture microstructural changes in the liver parenchyma related to the distribution of fibrosis, microvascular abnormalities, or local inflammation, which directly participate in the regulation of intrahepatic vascular resistance.^4,30

The interpretability of our model benefits from the logistic regression framework. Analysis showed that portal vein diameter, as a macroscopic marker of structural changes in the portal venous system, contributed the highest weight; the Child‑Pugh score and FIB‑4 score followed closely, highlighting the central role of liver functional reserve and fibrotic burden in the pathophysiology of PHT.^57,58 The selected high‑weight radiomic features were predominantly derived from texture analysis.⁵⁹ We cautiously speculate, pending future pathological correlation studies, that intratumoral features may be associated with tumor aggressiveness and vascular invasion tendency,⁶⁰ while peritumoral features may quantify microstructural alterations in the liver parenchyma caused by fibrosis, microvascular abnormalities, or local inflammation.⁶¹ However, these mechanistic interpretations remain speculative at this stage, as we did not perform direct histopathological validation. These alterations directly participate in the regulation of intrahepatic vascular resistance, thereby providing supplementary information beyond macroscopic indicators for the model.

In terms of clinical translation, this model provides a practical tool for preoperative PHT risk stratification using routine portal venous phase CT. By categorizing patients into different risk levels, it can help identify individuals at high risk for PHT, thereby guiding more proactive preoperative preparation (such as optimizing liver function and managing ascites), influencing the choice of surgical procedure to preserve more functional liver volume, and providing early warning for managing bleeding from portosystemic collaterals during surgery. However, the clinical utility of the model warrants careful consideration in light of its performance characteristics in the validation cohort. Specifically, the combined nomogram achieved a sensitivity of 0.783 and a specificity of 0.747, yet yielded a positive predictive value (PPV) of only 0.316. This PPV indicates that, when the model predicts PHT, approximately 31.6% of those predictions are correct in the validation cohort, implying that nearly 70% of positive predictions may be false positives. Such a discrepancy carries important clinical implications: a high false-positive rate could lead to unnecessary preoperative interventions, overly cautious surgical planning, or unwarranted patient anxiety. Several factors may contribute to this modest PPV, including the relatively low prevalence of PHT in the validation cohort (13.0%), which inherently limits PPV even with reasonable sensitivity and specificity. Additionally, the model was developed on a single-center cohort with a specific disease spectrum, and potential overfitting—despite internal cross-validation—may have inflated performance estimates in the training set that did not fully generalize.⁶² The use of intraoperatively measured portal venous pressure as the reference standard, while rigorous, may also introduce variability due to anesthesia-related hemodynamic fluctuations. These observations underscore the critical need for external validation using independent, geographically distinct populations to confirm the model’s generalizability and to provide more reliable estimates of its predictive performance across diverse clinical settings. Without such validation, the current PPV estimate should be interpreted cautiously, and the model is best positioned as an adjunct to clinical judgment rather than a standalone decision-making tool.

A critical consideration for clinical translation is whether the incremental benefit of radiomic features justifies their computational complexity and institution-dependent reproducibility, particularly when simpler clinical parameters already demonstrate robust predictive performance.In the present study, the clinical-only model achieved an AUC of 0.933 in the training cohort and 0.733 in the validation cohort, suggesting that readily available clinical indicators alone provide a reasonably strong foundation for PHT risk stratification. The addition of radiomic features improved the validation AUC to 0.847, representing a modest but non-negligible gain. However, this incremental improvement must be weighed against the substantial drop in performance from training to validation (0.938 vs 0.847) and, more importantly, the critically low positive predictive value (PPV) of 0.316 in the validation cohort. This PPV indicates that, when the model predicts PHT, nearly 70% of positive predictions are false positives in this externally-like validation setting, which severely undermines the model’s reliability as a standalone clinical decision-making tool. These findings suggest that while radiomic features may capture complementary biological information not fully reflected in clinical parameters, their current contribution does not yet justify widespread clinical adoption, particularly given the challenges of standardizing radiomic extraction across different scanners and institutions. Therefore, at this stage, the model is best positioned as a supplementary tool to clinical judgment rather than a replacement for simpler clinical risk assessment, and its utility remains contingent on rigorous external validation in diverse, multicenter cohorts.

In fact, the most robust finding of this work may be that simple, readily available clinical variables alone—specifically portal vein diameter, Child-Pugh score, and FIB-4 score—provide a reasonably strong foundation for PHT screening. The pattern of results suggests that while radiomic features may capture subtle biological signals relevant to PHT pathophysiology, their current performance, reproducibility challenges across institutions, and computational overhead render them better suited for research investigations than for routine clinical application. Hence, the primary contribution of this study may lie not in promoting an integrated model for clinical deployment, but in demonstrating that clinical variables remain the dominant predictors of PHT, with radiomic features offering incremental, yet currently insufficient, added value that requires further methodological refinement and rigorous external validation before clinical translation can be responsibly recommended.

This study has several limitations. First, external multicenter validation is mandatory before any claims of clinical utility can be made. As a single-center retrospective study with a geographically restricted population, our findings are inherently subject to institutional biases in patient selection, imaging protocols, and clinical management, which substantially limits generalizability.⁶³ Moreover, the sample included insufficient representation of patients with advanced Child-Pugh class C cirrhosis—the population at highest risk for clinically significant PHT and postoperative complications.⁶⁴ This underrepresentation precludes meaningful evaluation of model performance where accurate risk stratification is most needed. Without rigorous external validation across diverse populations and imaging protocols, the current model remains a hypothesis-generating tool rather than a clinically deployable instrument. Second, although we collected data on liver disease etiology, the potential confounding effects of etiological differences on PHT severity and radiomic features were not fully explored. Distinct etiologies—viral hepatitis versus alcohol-related liver disease—may influence fibrosis patterns, inflammation, and hemodynamic alterations through different pathophysiological mechanisms, potentially yielding distinct radiomic signatures. Furthermore, the presence of esophageal or gastric varices, a direct indicator of PHT severity, was not systematically documented or incorporated into the model. The absence of detailed stratification by variceal status and the limited granularity of etiological subgroup analysis remain limitations that may affect model generalizability. Third, radiomic features are susceptible to variations in CT acquisition parameters, and their multicenter standardization requires further improvement. Despite the use of cross-validation and ComBat harmonization, the relatively limited sample size carries a potential risk of overfitting, and residual batch effects across scanners may not have been fully eliminated.⁶⁵ Furthermore, the comparison of intratumoral, peritumoral, and pre-fusion radiomic feature models was performed without explicit correction for multiple comparisons (eg, Bonferroni), which may increase the risk of Type I error.⁶⁶ This issue should be addressed in future studies. In summary, while our model demonstrates promising discriminative ability and clinical interpretability in internal validation, its current performance—particularly the low PPV and the decline from training to validation—does not yet support standalone clinical use. The incremental benefit of integrating radiomic features must be carefully weighed against the added complexity and challenges of standardization. External validation in independent, geographically diverse populations remains indispensable before any claims of clinical utility can be made.

Conclusion

In this study, we developed a non‑invasive predictive model that integrates clinical variables with multi‑regional radiomic features. The model demonstrated moderate discriminative ability in internal validation, achieving an AUC of 0.847. However, this performance represents a substantial decline from the training AUC of 0.938, and the critically low positive predictive value of 0.316 in the validation cohort raises concerns about the model’s reliability for clinical decision‑making. Rather than promoting the integrated model as a clinical advance, the most robust finding of this study is that simple, readily available clinical variables alone—specifically portal vein diameter, Child‑Pugh score, and FIB‑4 score—provide a reasonably strong foundation for PHT screening. Radiomic features offered incremental but currently insufficient added value, which requires further methodological refinement and rigorous external validation before responsible clinical translation can be recommended. These findings suggest that although the combination of clinical and radiomic features yields incremental predictive value, the current model’s performance does not yet support its standalone use in clinical practice. It should be emphasized that clinical variables remain the dominant predictors in this model, with radiomic features serving as a complementary rather than a replacement component. This tool holds promise for assisting in the individualized assessment of PHT risk in HCC patients, thereby potentially optimizing clinical management decisions. Nevertheless, given the retrospective single‑center design, the generalizability of our findings is inherently limited; external multi‑center validation is essential to establish the model’s robustness and clinical applicability.