A multimodal deep learning model predicting hyperprogressive disease for PD-1 blockade in advanced hepatocellular carcinoma Authors: Yan Li, Xin Li, Xiaoqi Lin, Ziming Xu, Rui Li, Feng Duan, Chao An, Huijun Chen Published: 2026 ABSTRACT Immune checkpoint inhibitors, particularly antibodies targeting programmed cell death 1 (PD-1), are increasingly used for advanced hepatocellular carcinoma (Ad-HCC), but treatment responses remain heterogeneous. Hyperprogressive disease (HPD) is an especially concerning pattern of rapid progression after PD-1 therapy, and reliable pre-treatment tools to identify high-risk patients are still lacking. In this multicenter retrospective study of 665 patients with Ad-HCC receiving PD-1 inhibitor-based triple therapy, we developed a transformer-based multimodal model, Hyperprogression Oncological Predictive Enhanced-model (HOPE), integrating arterial- and portal-phase computed tomography with structured clinical factors. HOPE achieved an area under the receiver operating characteristic curve of 0.801 in the internal validation cohort and 0.687 in the external validation cohort, and outperformed clinical-only and imaging-only baseline models. Ablation analyses supported the value of multimodal integration. HOPE was further supported by prespecified subgroup analyses, survival risk stratification, and Gradient-weighted Class Activation Mapping (Grad-CAM) assessments. HOPE may serve as a clinically interpretable pretreatment decision-support tool for HPD risk stratification in patients with Ad-HCC receiving PD-1 inhibitor-based triple therapy, with potential utility for closer monitoring and risk-adapted management of patients predicted to be at high risk. Introduction Immune checkpoint inhibitors (ICIs), as novel immunotherapeutic agents, have seen increasing usage in treating advanced hepatocellular carcinoma (Ad-HCC)1–3. As a common type of ICI agents, antibodies targeting programmed cell death 1 (PD-1) were highlighted for their efficacy and safety in treating patients with Ad-HCC4,5. Notably, the atezolizumab plus bevacizumab regimen is recommended as first-line therapy for Ad-HCC6. In clinical practice, PD-1 based therapy is increasingly combined with anti-angiogenic agents and locoregional treatments, forming PD-1 inhibitor-based triple therapy aimed at improving tumor control7–9. However, according to a series of phase II and randomized phase III trials10–12, PD-1 only had 15–20% response rate to HCC. This limited efficacy might be attributable to the intratumoral and intertumoral heterogeneity in HCC, leading to heterogeneous progression patterns after PD-1 treatment13. Among the various responses to PD-1 blockade, hyperprogressive disease (HPD)14–16 represents the most concerning progression pattern, with potentially fatal consequences. Moreover, HPD can deprive opportunities for subsequent therapies. Recent studies have shown that the prevalence of HPD in patients treated with PD-1 inhibitors varies across different cancer types, ranging from 13.8% to 29.4%17–20. Kim et al. recently reported that 12.7% of patients with Ad-HCC who received nivolumab monotherapy had HPD21. Another report from Japan initially demonstrated 10.2% of the occurrence rate of HPD in patients with Ad-HCC who got Atez/Bev in either the first or later lines22. Recent advances in artificial intelligence (AI) have accelerated the adoption of transformer-based deep learning models in medical imaging and multimodal data integration23–26. Leveraging attention mechanisms, transformer can model nonlinear dependencies between imaging and clinical factors, which is particularly relevant when tumor behavior is shaped by both anatomical features and patient-level factors27. Several studies have either explored risk factors or developed predictive models for HPD, however, they mostly rely on single modality analysis28,29 and frequently employed conventional machine learning approaches30,31. These conventional machine learning models often rely on handcrafted features and have limited ability to capture complex cross-modal interactions, which compromises robustness and generalizability in heterogeneous cohorts. To date, few models have been developed to predict HPD in patients with Ad-HCC receiving PD-1 inhibitors. Therefore, this study aims to develop a transformer-based multimodal deep learning model, HOPE (Hyperprogression Oncological Predictive Enhanced-model), to predict HPD risk in Ad-HCC patients treated with PD-1 blockade in order to support individualized therapeutic decision-making. HOPE is, to our knowledge, the first transformer-based multimodal model for HPD prediction in Ad-HCC under PD-1 blockade, leveraging multiphase computed tomography (CT) inputs and bidirectional multimodal attention to explicitly capture imaging and clinical interactions. Results Baseline characteristics A total of 1124 patients were initially screened for eligibility, of whom 459 were excluded based on predefined criteria, 665 patients were included for analyses in this study. The median follow-up duration was 14.5 months (interquartile range [IQR] 8.9–23.1) for the entire cohort, 16.6 months (IQR 10.2–24.6) in the training cohort (n = 403), 17.6 months (IQR 10.9–29.0) in the internal validation cohort (n = 101), and 9.7 months (IQR 5.2–14.3) in the external validation cohort (n = 161). Baseline characteristics are summarized in Table 1. Intratumoral artery demonstrated a consistent between-group difference between HPD and non-HPD patients across cohorts, whereas other variables did not show consistent differences.Table 1Baseline characteristics of patients with hyperprogression or non-hyperprogressionVariablesTraining cohort (n = 403)Internal validation cohort (n = 101)External validation cohort (n = 161)HPD (n = 69)Non-HPD (n = 334)pHPD (n = 15)Non-HPD (n = 86)pHPD (n = 25)Non-HPD (n = 136)pGender0.7850.6870.116Male61 (88.4)299 (89.5)14 (93.3)73 (84.9)19 (76.0)120 (88.2)Female8 (11.6)35 (10.5)1 (6.7)13 (15.1)6 (24.0)16 (11.8)Age (yr)49.0 (36.5–57.0)51.0 (43.0–61.0)0.016*42.0 (37.0–51.0)49.0 (41.8–58.0)0.05553.0 (48.5–58.0)54.0 (51.0-62.8)0.184ECOG0.002**0.6880.121054 (78.3)305 (91.3)14 (93.3)75 (87.2)22 (88.0)100 (73.5)115 (21.7)29 (8.7)1 (6.7)11 (12.8)3 (12.0)36 (26.5)Hepatic virus infection0.7490.6410.548Absence6 (8.7)31 (9.3)2 (13.3)8 (9.3)1 (4.0)11 (8.1)HBV62 (89.9)301 (90.1)13 (86.7)78 (90.7)23 (92.0)123 (90.4)HCV1 (1.4)2 (0.6)0 (0.0)0 (0.0)1 (4.0)2 (1.5)Comorbidities0.0530.6880.071Absence65 (94.2)286 (85.6)14 (93.3)75 (87.2)22 (88.0)96 (70.6)Presence4 (5.8)48 (14.4)1 (6.7)11 (12.8)3 (12.0)40 (29.4)AFP (ng/mL)10357.0 (229.2–121000.0)816.0 (38.4–20885.3)0.001**41390.0 (432.0–67310.0)1357.0 (60.5–24930.8)0.21712047.0 (1218.3–60500.0)1565.5 (56.8–23015.3)0.015*ANC (×109/L)5.1 (3.8–6.7)4.8 (3.6–6.4)0.3935.2 (4.0–6.5)4.5 (3.6–5.9)0.2043.8 (2.7–5.8)3.2 (2.4–4.6)0.324ALC (×109/L)1.6 (1.1–2.0)1.5 (1.2–1.9)0.6531.5 (1.1–1.9)1.5 (1.1–1.7)0.7821.2 (1.1–1.8)1.2 (0.9–1.5)0.028*PLT (×109/L)256.0 (179.5–330.5)242.0 (163.8–311.5)0.216301.0 (192.0–361.0)204.5 (139.3–284.0)0.074196.0 (144.0–242.5)147.5 (107.3–214.3)0.027*CRE (µmol/L)71.3 (57.6–81.5)70.8 (60.5–78.4)0.92962.1 (59.2–77.2)71.5 (61.6–80.3)0.14961.4 (50.1–68.4)61.0 (53.2–69.8)0.491ALB (g/L)40.9 (37.8–44.8)41.4 (37.4–44.0)0.56142.8 (39.3–48.7)41.4 (38.9–44.1)0.36238.3 (34.7–41.2)39.4 (36.2–42.9)0.370TBIL (µmol/L)15.9 (11.2–24.1)15.8 (11.7–21.0)0.88716.9 (11.3–24.3)15.9 (12.6–20.9)0.71717.8 (12.9–30.9)18.1 (12.9–30.9)0.913ALT (U/L)51.8 (34.0–70.4)45.6 (30.5–65.9)0.27843.5 (32.3–62.6)47.0 (29.9–79.2)0.61941.1 (28.3–54.3)39.6 (27.7–62.8)0.970AST (U/L)73.8 (46.9–130.7)68.0 (42.6–105.0)0.07376.2 (56.8–156.2)67.5 (41.0–118.2)0.42881.7 (47.9–161.9)62.0 (37.9–94.2)0.058Size of the largest tumor (cm)11.2 (9.1–13.6)10.9 (8.0–13.3)0.19611.3 (9.7–13.1)11.3 (7.4–14.0)0.77813.5 (11.0–17.3)10.5 (6.3–15.3)0.013*Tumor number0.1700.4750.720≤three26 (37.7)156 (46.7)6 (40.0)43 (50.0)13 (52.0)76 (55.9)>three43 (62.3)178 (53.3)9 (60.0)43 (50.0)12 (48.0)60 (44.1)Vascular invasion0.5180.4330.369Absence25 (36.2)135 (40.4)5 (33.3)38 (44.2)6 (24.0)45 (33.1)Presence44 (63.8)199 (59.6)10 (66.7)48 (55.8)19 (76.0)91 (66.9)Metastasis0.0600.0780.868Absence35 (50.7)210 (62.9)7 (46.7)61 (70.9)15 (60.0)84 (61.8)Presence34 (49.3)124 (37.1)8 (53.3)25 (29.1)10 (40.0)52 (38.2)BCLC staging0.0990.6110.674A5 (7.2)50 (15.5)2 (13.3)20 (23.3)1 (4.0)11 (8.1)B9 (13.0)60 (18.0)2 (13.3)14 (16.3)2 (8.0)15 (11.0)C55 (79.7)224 (67.1)11 (73.3)52 (60.5)22 (88.0)110 (80.9)Intratumoral artery<0.001***<0.001***<0.001***Absence11 (15.9)327 (97.9)4 (26.7)84 (97.7)Presence58 (84.1)7 (2.1)11 (73.3)2 (2.3)8 (32.0)7 (5.1)“Capsule” appearance0.038*0.1921.000Absence26 (37.7)85 (25.4)6 (40.0)19 (22.1)4 (16.0)23 (16.9)Presence43 (62.3)249 (74.6)9 (60.0)67 (77.9)Necrosis or severe ischemia0.5540.004**0.742Absence60 (87.0)281 (84.1)9 (60.0)79 (91.9)Presence9 (13.0)53 (15.9)6 (40.0)7 (8.1)2 (8.0)18 (13.2)Fat in mass, more than liver0.6940.1801.000Absence52 (75.4)259 (77.5)14 (93.3)65 (75.6)Presence17 (24.6)75 (22.5)1 (6.7)21 (24.4)3 (12.0)17 (12.5)Blood products in mass0.5340.4540.203Absence60 (87.0)299 (89.5)14 (93.3)71 (82.6)Presence9 (13.0)35 (10.5)1 (6.7)15 (17.4)5 (20.0)15 (11.0)Corona enhancement0.0781.0000.230Absence59 (85.5)253 (75.7)13 (86.7)73 (84.9)Presence10 (14.5)81 (24.3)2 (13.3)13 (15.1)6 (24.0)19 (14.0)Categorical variables are shown as n (%), and continuous variables as median (IQR). The chi-square or Fisher’s exact test was used for categorical variables, and the Kruskal–Wallis test for continuous variables.p < 0.05 was considered statistically significant. *p < 0.05; **p < 0.01; ***p < 0.001. Table 1Baseline characteristics of patients with hyperprogression or non-hyperprogressionVariablesTraining cohort (n = 403)Internal validation cohort (n = 101)External validation cohort (n = 161)HPD (n = 69)Non-HPD (n = 334)pHPD (n = 15)Non-HPD (n = 86)pHPD (n = 25)Non-HPD (n = 136)pGender0.7850.6870.116Male61 (88.4)299 (89.5)14 (93.3)73 (84.9)19 (76.0)120 (88.2)Female8 (11.6)35 (10.5)1 (6.7)13 (15.1)6 (24.0)16 (11.8)Age (yr)49.0 (36.5–57.0)51.0 (43.0–61.0)0.016*42.0 (37.0–51.0)49.0 (41.8–58.0)0.05553.0 (48.5–58.0)54.0 (51.0-62.8)0.184ECOG0.002**0.6880.121054 (78.3)305 (91.3)14 (93.3)75 (87.2)22 (88.0)100 (73.5)115 (21.7)29 (8.7)1 (6.7)11 (12.8)3 (12.0)36 (26.5)Hepatic virus infection0.7490.6410.548Absence6 (8.7)31 (9.3)2 (13.3)8 (9.3)1 (4.0)11 (8.1)HBV62 (89.9)301 (90.1)13 (86.7)78 (90.7)23 (92.0)123 (90.4)HCV1 (1.4)2 (0.6)0 (0.0)0 (0.0)1 (4.0)2 (1.5)Comorbidities0.0530.6880.071Absence65 (94.2)286 (85.6)14 (93.3)75 (87.2)22 (88.0)96 (70.6)Presence4 (5.8)48 (14.4)1 (6.7)11 (12.8)3 (12.0)40 (29.4)AFP (ng/mL)10357.0 (229.2–121000.0)816.0 (38.4–20885.3)0.001**41390.0 (432.0–67310.0)1357.0 (60.5–24930.8)0.21712047.0 (1218.3–60500.0)1565.5 (56.8–23015.3)0.015*ANC (×109/L)5.1 (3.8–6.7)4.8 (3.6–6.4)0.3935.2 (4.0–6.5)4.5 (3.6–5.9)0.2043.8 (2.7–5.8)3.2 (2.4–4.6)0.324ALC (×109/L)1.6 (1.1–2.0)1.5 (1.2–1.9)0.6531.5 (1.1–1.9)1.5 (1.1–1.7)0.7821.2 (1.1–1.8)1.2 (0.9–1.5)0.028*PLT (×109/L)256.0 (179.5–330.5)242.0 (163.8–311.5)0.216301.0 (192.0–361.0)204.5 (139.3–284.0)0.074196.0 (144.0–242.5)147.5 (107.3–214.3)0.027*CRE (µmol/L)71.3 (57.6–81.5)70.8 (60.5–78.4)0.92962.1 (59.2–77.2)71.5 (61.6–80.3)0.14961.4 (50.1–68.4)61.0 (53.2–69.8)0.491ALB (g/L)40.9 (37.8–44.8)41.4 (37.4–44.0)0.56142.8 (39.3–48.7)41.4 (38.9–44.1)0.36238.3 (34.7–41.2)39.4 (36.2–42.9)0.370TBIL (µmol/L)15.9 (11.2–24.1)15.8 (11.7–21.0)0.88716.9 (11.3–24.3)15.9 (12.6–20.9)0.71717.8 (12.9–30.9)18.1 (12.9–30.9)0.913ALT (U/L)51.8 (34.0–70.4)45.6 (30.5–65.9)0.27843.5 (32.3–62.6)47.0 (29.9–79.2)0.61941.1 (28.3–54.3)39.6 (27.7–62.8)0.970AST (U/L)73.8 (46.9–130.7)68.0 (42.6–105.0)0.07376.2 (56.8–156.2)67.5 (41.0–118.2)0.42881.7 (47.9–161.9)62.0 (37.9–94.2)0.058Size of the largest tumor (cm)11.2 (9.1–13.6)10.9 (8.0–13.3)0.19611.3 (9.7–13.1)11.3 (7.4–14.0)0.77813.5 (11.0–17.3)10.5 (6.3–15.3)0.013*Tumor number0.1700.4750.720≤three26 (37.7)156 (46.7)6 (40.0)43 (50.0)13 (52.0)76 (55.9)>three43 (62.3)178 (53.3)9 (60.0)43 (50.0)12 (48.0)60 (44.1)Vascular invasion0.5180.4330.369Absence25 (36.2)135 (40.4)5 (33.3)38 (44.2)6 (24.0)45 (33.1)Presence44 (63.8)199 (59.6)10 (66.7)48 (55.8)19 (76.0)91 (66.9)Metastasis0.0600.0780.868Absence35 (50.7)210 (62.9)7 (46.7)61 (70.9)15 (60.0)84 (61.8)Presence34 (49.3)124 (37.1)8 (53.3)25 (29.1)10 (40.0)52 (38.2)BCLC staging0.0990.6110.674A5 (7.2)50 (15.5)2 (13.3)20 (23.3)1 (4.0)11 (8.1)B9 (13.0)60 (18.0)2 (13.3)14 (16.3)2 (8.0)15 (11.0)C55 (79.7)224 (67.1)11 (73.3)52 (60.5)22 (88.0)110 (80.9)Intratumoral artery<0.001***<0.001***<0.001***Absence11 (15.9)327 (97.9)4 (26.7)84 (97.7)Presence58 (84.1)7 (2.1)11 (73.3)2 (2.3)8 (32.0)7 (5.1)“Capsule” appearance0.038*0.1921.000Absence26 (37.7)85 (25.4)6 (40.0)19 (22.1)4 (16.0)23 (16.9)Presence43 (62.3)249 (74.6)9 (60.0)67 (77.9)Necrosis or severe ischemia0.5540.004**0.742Absence60 (87.0)281 (84.1)9 (60.0)79 (91.9)Presence9 (13.0)53 (15.9)6 (40.0)7 (8.1)2 (8.0)18 (13.2)Fat in mass, more than liver0.6940.1801.000Absence52 (75.4)259 (77.5)14 (93.3)65 (75.6)Presence17 (24.6)75 (22.5)1 (6.7)21 (24.4)3 (12.0)17 (12.5)Blood products in mass0.5340.4540.203Absence60 (87.0)299 (89.5)14 (93.3)71 (82.6)Presence9 (13.0)35 (10.5)1 (6.7)15 (17.4)5 (20.0)15 (11.0)Corona enhancement0.0781.0000.230Absence59 (85.5)253 (75.7)13 (86.7)73 (84.9)Presence10 (14.5)81 (24.3)2 (13.3)13 (15.1)6 (24.0)19 (14.0)Categorical variables are shown as n (%), and continuous variables as median (IQR). The chi-square or Fisher’s exact test was used for categorical variables, and the Kruskal–Wallis test for continuous variables.p < 0.05 was considered statistically significant. *p < 0.05; **p < 0.01; ***p < 0.001. Baseline characteristics of patients with hyperprogression or non-hyperprogression Categorical variables are shown as n (%), and continuous variables as median (IQR). The chi-square or Fisher’s exact test was used for categorical variables, and the Kruskal–Wallis test for continuous variables. p < 0.05 was considered statistically significant. *p < 0.05; **p < 0.01; ***p < 0.001. HPD incidence was comparable across cohorts: 17.1% (69/403) in training, 14.9% (15/101) in internal validation, and 15.5% (25/161) in external validation, consistent with prior reports in immunotherapy-treated populations15,32. Baseline characteristics by cohort and clinical characteristics stratified by HOPE-predicted risk groups are provided in Supplementary Tables 1, 2. The distribution of specific PD-1 inhibitors across cohorts is summarized in Supplementary Table 3. Performance comparison between various models In this study, two senior radiologists manually refined the automatic segmentation outputs, requiring ~6 h in total. A total of 1330 contrast-enhanced computed tomography (CECT) slices from two centers were used for the development and validation of HOPE. The predictive performance of HOPE is illustrated in Fig. 1. Across all datasets, HOPE demonstrated robust and consistent discrimination for identifying patients who developed HPD. In the receiver operating characteristic (ROC) analysis (Fig. 1A), HOPE achieved an area under the receiver operating characteristic curve (AUC) of 0.801 (95% confidence interval [CI], 0.710–0.874) in the internal validation cohort and moderate yet clinically relevant discrimination in the external validation cohort (AUC 0.687 [95% CI, 0.610–0.758]). Confusion matrices (Fig. 1B) provide clinically critical insights into HOPE’s behavior. HOPE achieved very high sensitivity for HPD in both the training and internal validation cohorts, resulting in an extremely low false negative rate. This is particularly important given the association of HPD with rapid tumor acceleration, early treatment failure, and profoundly shortened survival33, making missed HPD predictions potentially detrimental for patients being considered for immunotherapy. The ability of HOPE to minimize false negative predictions therefore directly supports its potential utility in pre-treatment risk stratification. HOPE-derived risk score distributions (Fig. 1C) further validated the model’s robustness, demonstrating clear separation (all p < 0.001), and the preservation of this separation in the external validation cohort supports the cross-institutional generalizability of HOPE. To further characterize HOPE performance, we additionally evaluated calibration and clinical utility. As shown in Supplementary Figure 1, the calibration curve demonstrated acceptable agreement between predicted and observed probabilities, with a Brier score of 0.127, and decision curve analysis (DCA) showed that HOPE provided a positive net benefit across a clinically relevant range of threshold probabilities. In addition, sensitivity analyses using alternative commonly adopted HPD definitions showed that HOPE retained predictive value across different endpoint operationalizations, while the prespecified study definition yielded the most favorable overall performance (Supplementary Table 4).Fig. 1Performance of different models and assessment of hyperprogressive disease identification of HOPE.A The receiver operating characteristic curves for predicting hyperprogressive disease (HPD) in the training cohort, internal validation cohort, and external validation cohort. B Confusion matrices of HOPE’s prediction results for the training cohort, internal validation cohort, and external validation cohort. C Distribution of the risk scores predicted by HOPE in the training cohort, internal validation cohort, and external validation cohort. Performance of different models and assessment of hyperprogressive disease identification of HOPE. A The receiver operating characteristic curves for predicting hyperprogressive disease (HPD) in the training cohort, internal validation cohort, and external validation cohort. B Confusion matrices of HOPE’s prediction results for the training cohort, internal validation cohort, and external validation cohort. C Distribution of the risk scores predicted by HOPE in the training cohort, internal validation cohort, and external validation cohort. Table 2 presents a detailed comparison of HOPE with clinical-only and image-only baseline models. Among the clinical models, clinical model using clinical factors alone (Clin model) showed limited discriminative ability, with AUCs ranging from 0.613 to 0.629 across cohorts. Incorporating radiological features (ClinRad model) improved performance in the training cohort (AUC 0.711 [95% CI, 0.664–0.755]), though this enhancement did not generalize well to external validation (AUC 0.559 [95% CI, 0.479–0.637]). Among image-only deep learning models, 3D ResNet-18 demonstrated moderate discrimination in the training and internal validation cohorts (AUCs 0.739 [95% CI, 0.694–0.782] and 0.716 [95% CI, 0.618–0.802], respectively). However, the AUC in the external validation cohort was 0.561 (95% CI, 0.481–0.639), suggesting limited regularization ability. Despite a high training AUC (0.996 [95% CI, 0.984–1.000]), Vision Transformer (ViT) yielded lower AUCs in the internal and external validation cohorts (AUC 0.703 [95% CI, 0.604–0.790] and 0.556 [95% CI, 0.476–0.634]), indicating reduced robustness when applied to independent cohorts. Among all evaluated models, HOPE achieved the highest and the most stable overall performance. DeLong tests further confirmed that HOPE significantly outperformed both clinical models and image-only models in multiple comparisons, particularly within the training and internal validation cohorts. Details of model parameters and input variables are summarized in Supplementary Tables 5, 6.Table 2The performance of different modelsModelsCohortsAUC (95% CI)Accuracy (95% CI)Sensitivity (95% CI)Specificity (95% CI)PPV (95% CI)NPV (95% CI)LR+ (95% CI)LR- (95% CI)p-valueClin model (CLI)Training0.629 (0.580–0.676)0.556 (0.506–0.605)0.681 (0.558–0.788)0.681 (0.558–0.788)0.230 (0.197–0.267)0.889 (0.849–0.920)1.449 (1.189–1.766)0.602 (0.420–0.862)<0.001Internal validation0.613 (0.511–0.708)0.614 (0.512–0.709)0.667 (0.384–0.882)0.605 (0.493–0.708)0.227 (0.159–0.314)0.912 (0.833–0.956)1.686 (1.083–2.627)0.551 (0.264–1.151)0.033External validation0.616 (0.536–0.691)0.596 (0.516–0.673)0.640 (0.425–0.820)0.588 (0.501–0.672)0.222 (0.167–0.290)0.899 (0.838–0.939)1.554 (1.089–2.219)0.612 (0.356–1.052)0.362ClinRad model (CLI + RAD)Training0.711 (0.664–0.755)0.660 (0.612–0.706)0.870 (0.767–0.939)0.617 (0.562–0.669)0.319 (0.285–0.356)0.958 (0.925–0.977)2.269 (1.926–2.673)0.211 (0.114–0.391)<0.001Internal validation0.626 (0.524–0.720)0.604 (0.502–0.700)0.867 (0.595–0.983)0.558 (0.447–0.665)0.255 (0.201–0.318)0.960 (0.867–0.989)1.961 (1.439–2.673)0.239 (0.065–0.880)0.039External validation0.559 (0.479–0.637)0.665 (0.586–0.737)0.480 (0.278–0.687)0.699 (0.614–0.774)0.226 (0.153–0.321)0.880 (0.832–0.915)1.592 (0.984–2.577)0.744 (0.503–1.102)0.1003D ResNet-18 (IMG)Training0.739 (0.694–0.782)0.799 (0.757–0.837)0.493 (0.370–0.616)0.862 (0.821–0.897)0.425 (0.340–0.514)0.892 (0.867–0.912)3.578 (2.497–5.126)0.588 (0.464–0.745)<0.001Internal validation0.716 (0.618–0.802)0.554 (0.452–0.653)1.000 (0.782–1.000)0.477 (0.368–0.587)0.250 (0.214–0.290)1.000 (1.000–1.000)1.911 (1.562–2.338)0.000 (0.000–0.000)0.299External validation0.561 (0.481–0.639)0.851 (0.786–0.902)0.240 (0.094–0.451)0.963 (0.916–0.988)0.545 (0.284–0.784)0.873 (0.847–0.896)6.528 (2.157–19.760)0.789 (0.631–0.986)0.185ViT (IMG)Training0.996 (0.984–1.000)0.980 (0.961–0.991)0.918 (0.830–0.969)0.994 (0.978–0.999)0.971 (0.894–0.993)0.982 (0.962–0.992)151.438 (37.968–604.019)0.083 (0.038–0.178)0.177Internal validation0.703 (0.604–0.790)0.711 (0.606–0.802)0.600 (0.323–0.837)0.733 (0.619–0.829)0.310 (0.205–0.440)0.902 (0.829–0.945)2.250 (1.288–3.932)0.545 (0.289–1.029)0.252External validation0.556 (0.476–0.634)0.752 (0.677 – 0.816)0.400 (0.211–0.613)0.816 (0.741–0.877)0.286 (0.181–0.421)0.881 (0.842–0.911)2.176 (1.198–3.951)0.735 (0.529–1.022)0.042HOPE (CLI + RAD + IMG)Training0.999 (0.989–1.000)0.953 (0.927–0.971)1.000 (0.948–1.000)0.943 (0.913–0.965)0.784 (0.701–0.849)1.000 (1.000–1.000)17.579 (11.359–27.204)0.000 (0.000–0.000)–Internal validation0.801 (0.710–0.874)0.713 (0.614–0.799)0.733 (0.449–0.922)0.709 (0.601–0.802)0.306 (0.219–0.408)0.938 (0.867–0.973)2.523 (1.609–3.955)0.376 (0.161–0.880)–External validation0.687 (0.610–0.758)0.708 (0.631–0.777)0.640 (0.425–0.820)0.721 (0.637–0.794)0.296 (0.220–0.386)0.916 (0.865–0.949)2.291 (1.537–3.414)0.500 (0.293–0.851)–Significance was tested by the DeLong test between the other models and HOPE. Table 2The performance of different modelsModelsCohortsAUC (95% CI)Accuracy (95% CI)Sensitivity (95% CI)Specificity (95% CI)PPV (95% CI)NPV (95% CI)LR+ (95% CI)LR- (95% CI)p-valueClin model (CLI)Training0.629 (0.580–0.676)0.556 (0.506–0.605)0.681 (0.558–0.788)0.681 (0.558–0.788)0.230 (0.197–0.267)0.889 (0.849–0.920)1.449 (1.189–1.766)0.602 (0.420–0.862)<0.001Internal validation0.613 (0.511–0.708)0.614 (0.512–0.709)0.667 (0.384–0.882)0.605 (0.493–0.708)0.227 (0.159–0.314)0.912 (0.833–0.956)1.686 (1.083–2.627)0.551 (0.264–1.151)0.033External validation0.616 (0.536–0.691)0.596 (0.516–0.673)0.640 (0.425–0.820)0.588 (0.501–0.672)0.222 (0.167–0.290)0.899 (0.838–0.939)1.554 (1.089–2.219)0.612 (0.356–1.052)0.362ClinRad model (CLI + RAD)Training0.711 (0.664–0.755)0.660 (0.612–0.706)0.870 (0.767–0.939)0.617 (0.562–0.669)0.319 (0.285–0.356)0.958 (0.925–0.977)2.269 (1.926–2.673)0.211 (0.114–0.391)<0.001Internal validation0.626 (0.524–0.720)0.604 (0.502–0.700)0.867 (0.595–0.983)0.558 (0.447–0.665)0.255 (0.201–0.318)0.960 (0.867–0.989)1.961 (1.439–2.673)0.239 (0.065–0.880)0.039External validation0.559 (0.479–0.637)0.665 (0.586–0.737)0.480 (0.278–0.687)0.699 (0.614–0.774)0.226 (0.153–0.321)0.880 (0.832–0.915)1.592 (0.984–2.577)0.744 (0.503–1.102)0.1003D ResNet-18 (IMG)Training0.739 (0.694–0.782)0.799 (0.757–0.837)0.493 (0.370–0.616)0.862 (0.821–0.897)0.425 (0.340–0.514)0.892 (0.867–0.912)3.578 (2.497–5.126)0.588 (0.464–0.745)<0.001Internal validation0.716 (0.618–0.802)0.554 (0.452–0.653)1.000 (0.782–1.000)0.477 (0.368–0.587)0.250 (0.214–0.290)1.000 (1.000–1.000)1.911 (1.562–2.338)0.000 (0.000–0.000)0.299External validation0.561 (0.481–0.639)0.851 (0.786–0.902)0.240 (0.094–0.451)0.963 (0.916–0.988)0.545 (0.284–0.784)0.873 (0.847–0.896)6.528 (2.157–19.760)0.789 (0.631–0.986)0.185ViT (IMG)Training0.996 (0.984–1.000)0.980 (0.961–0.991)0.918 (0.830–0.969)0.994 (0.978–0.999)0.971 (0.894–0.993)0.982 (0.962–0.992)151.438 (37.968–604.019)0.083 (0.038–0.178)0.177Internal validation0.703 (0.604–0.790)0.711 (0.606–0.802)0.600 (0.323–0.837)0.733 (0.619–0.829)0.310 (0.205–0.440)0.902 (0.829–0.945)2.250 (1.288–3.932)0.545 (0.289–1.029)0.252External validation0.556 (0.476–0.634)0.752 (0.677 – 0.816)0.400 (0.211–0.613)0.816 (0.741–0.877)0.286 (0.181–0.421)0.881 (0.842–0.911)2.176 (1.198–3.951)0.735 (0.529–1.022)0.042HOPE (CLI + RAD + IMG)Training0.999 (0.989–1.000)0.953 (0.927–0.971)1.000 (0.948–1.000)0.943 (0.913–0.965)0.784 (0.701–0.849)1.000 (1.000–1.000)17.579 (11.359–27.204)0.000 (0.000–0.000)–Internal validation0.801 (0.710–0.874)0.713 (0.614–0.799)0.733 (0.449–0.922)0.709 (0.601–0.802)0.306 (0.219–0.408)0.938 (0.867–0.973)2.523 (1.609–3.955)0.376 (0.161–0.880)–External validation0.687 (0.610–0.758)0.708 (0.631–0.777)0.640 (0.425–0.820)0.721 (0.637–0.794)0.296 (0.220–0.386)0.916 (0.865–0.949)2.291 (1.537–3.414)0.500 (0.293–0.851)–Significance was tested by the DeLong test between the other models and HOPE. The performance of different models Significance was tested by the DeLong test between the other models and HOPE. Sensitivity analysis of slice representation To further assess whether limited through-plane context improved prediction, we additionally implemented a 2.5D variant of HOPE centered on the largest tumor slice by incorporating neighboring slices at offsets ±1, ±2, and ±4 under the same multimodal framework. Compared with 2.5D model, the 2D formulation showed better discrimination in both the internal validation cohort (AUC, 0.801 vs 0.757) and the external validation cohort (AUC, 0.687 vs 0.562), with generally higher accuracy and specificity, although the 2.5D model showed higher sensitivity in the internal validation cohort. These findings suggest that, in the present dataset, adding limited through-plane context did not improve and may have reduced generalizability, supporting the use of the simpler largest-slice representation in the main analysis (Supplementary Table 7). Ablation analysis of multimodal integration To assess the incremental value of multimodal integration in HOPE, we evaluated the model across different input configurations, including image-only, clinical-only, and multimodal ablation settings in which one category of clinical factors was removed at a time (Table 3). HOPE, integrating CT images with demographic, tumor morphological, and laboratory factors, achieved the highest discriminative performance in both the internal and external validation cohorts, with AUCs of 0.801 (95% CI, 0.710–0.874) and 0.687 (95% CI, 0.610–0.758), respectively. In contrast, the image-only model yielded AUCs of 0.753 (95% CI, 0.657–0.833) in the internal validation cohort and 0.584 (95% CI, 0.503–0.661) in the external validation cohort, whereas the clinical-only model yielded AUCs of 0.665 (95% CI, 0.564–0.756) and 0.669 (95% CI, 0.591–0.741), respectively. Together, these findings indicate that neither imaging nor clinical factors alone was sufficient to recapitulate the performance of the full multimodal model.Table 3Ablation analysis of HOPE under different input configurationsImagesDemographicTumor morphologicalLaboratoryInternal validation AUC (95% CI)External validation AUC (95% CI)HOPE√√√√0.801 (0.710–0.874)0.687 (0.610–0.758)Image-only√0.753 (0.657–0.833)0.584 (0.503–0.661)Clinical-only√√√0.665 (0.564–0.756)0.669 (0.591–0.741)w/o Demographic√√√0.704 (0.605–0.791)0.613 (0.533–0.689)w/o Tumor morphological√√√0.724 (0.626–0.808)0.611 (0.531–0.687)w/o Laboratory√√√0.753 (0.657–0.833)0.623 (0.544–0.698) Table 3Ablation analysis of HOPE under different input configurationsImagesDemographicTumor morphologicalLaboratoryInternal validation AUC (95% CI)External validation AUC (95% CI)HOPE√√√√0.801 (0.710–0.874)0.687 (0.610–0.758)Image-only√0.753 (0.657–0.833)0.584 (0.503–0.661)Clinical-only√√√0.665 (0.564–0.756)0.669 (0.591–0.741)w/o Demographic√√√0.704 (0.605–0.791)0.613 (0.533–0.689)w/o Tumor morphological√√√0.724 (0.626–0.808)0.611 (0.531–0.687)w/o Laboratory√√√0.753 (0.657–0.833)0.623 (0.544–0.698) Ablation analysis of HOPE under different input configurations Under matched unimodal settings, HOPE-derived models also outperformed the corresponding baseline models. HOPE-derived image-only model achieved higher AUCs than both 3D ResNet-18 (0.716 internally and 0.561 externally) and ViT (0.703 internally and 0.556 externally). Similarly, HOPE-derived clinical-only model outperformed the ClinRad model (0.626 internally and 0.559 externally). These results suggest that the observed performance gain was not explained solely by the inclusion of additional input features, but was also consistent with a contribution from the HOPE fusion framework itself. We next assessed the relative contribution of individual clinical factor groups within the multimodal framework. Removing any clinical subgroup led to a reduction in performance in both validation cohorts. Specifically, excluding demographic factors reduced the AUC to 0.704 (95% CI, 0.605–0.791) internally and 0.613 (95% CI, 0.533–0.689) externally; excluding tumor morphological factors reduced the AUC to 0.724 (95% CI, 0.626–0.808) and 0.611 (95% CI, 0.531–0.687), respectively; and excluding laboratory factors reduced the AUC to 0.753 (95% CI, 0.657–0.833) and 0.623 (95% CI, 0.544–0.698), respectively. Notably, the largest performance decrease in the internal validation cohort was observed after removal of demographic factors, whereas in the external validation cohort the largest decrease was observed after removal of tumor morphological factors, with demographic and laboratory factors also providing consistent incremental benefit. These findings support that the predictive advantage of HOPE arose from the complementary integration of imaging and multidomain clinical factors, rather than from any single modality alone. Subgroup analysis Subgroup analyses across key clinical and tumor-related factors demonstrated that HOPE maintained relatively stable discriminative performance across heterogeneous patient subsets (Fig. 2; Supplementary Table 8). Representative results showed that HOPE achieved an AUC of 0.699 (95% CI, 0.551–0.847) in patients with α-fetoprotein (AFP) ≥ 400 ng/mL and 0.682 (95% CI, 0.510–0.853) in tumors ≥10 cm, indicating maintained performance in biologically aggressive disease. Consistent discrimination was also observed between patients with and without metastasis (AUC 0.737 vs 0.620) and vascular invasion (AUC 0.848 vs 0.622), supporting the robustness of HOPE across clinically high-risk phenotypes.Fig. 2Subgroup analyses of discrimination performance for HOPE and baseline models across key clinical and tumor-related factors.AUCs with 95% confidence intervals are shown for AFP level, vascular invasion, size of the largest tumor, corona enhancement, metastasis, and “capsule” appearance. Subgroup analyses of discrimination performance for HOPE and baseline models across key clinical and tumor-related factors. AUCs with 95% confidence intervals are shown for AFP level, vascular invasion, size of the largest tumor, corona enhancement, metastasis, and “capsule” appearance. Across all prespecified subgroups, DeLong tests revealed no statistically significant differences in HOPE AUCs between subgroups within any factor (all p ≥ 0.22), indicating consistent discrimination across clinical subgroups. In contrast, baseline models exhibited subgroup-dependent variability. For example, ClinRad model showed marked performance imbalance between patients without and with metastasis (AUC 0.419 vs 0.797; p = 0.0018), while 3D ResNet-18 demonstrated significantly reduced discrimination in tumors ≥10 cm compared with smaller lesions (AUC 0.524 vs 0.833; p = 0.0188). To further determine whether this performance reflected HPD-related prediction rather than baseline tumor burden alone, we performed a multivariable logistic regression analysis with HPD as the outcome, including HOPE score, AFP, size of the largest tumor, vascular invasion, metastasis, corona enhancement and “capsule” appearance as covariates (Supplementary Table 9). The HOPE score remained independently associated with HPD after adjustment; for each 0.01 increase in score, the odds of HPD increased by 1.792-fold (95% CI, 1.534–2.094; p < 0.001). In contrast, AFP, size of the largest tumor, vascular invasion, metastasis, corona enhancement and “capsule” appearance were not independently associated with HPD in the same model (all p > 0.05). Moreover, a model based on baseline tumor burden factors alone showed substantially lower discrimination than the full model incorporating HOPE score (AUC, 0.644 vs 0.898), and the addition of HOPE score significantly improved model fit (likelihood ratio test, p < 0.001). Collectively, these findings indicate that the predictive value of HOPE is not fully explained by conventional baseline clinical and radiological factors alone, and that HOPE provides substantial incremental value for HPD risk prediction beyond these established variables. Although all patients received the same triple-therapy backbone consisting of conventional transarterial chemoembolization (TACE), lenvatinib, and a PD-1 inhibitor, the specific PD-1 inhibitor was not uniformly distributed across the three cohorts (p < 0.001; Supplementary Table 3). To account for this potential treatment heterogeneity, we further performed a subgroup analysis according to the specific PD-1 inhibitor used. The results showed that HOPE remained predictive of HPD across the four PD-1 inhibitor subgroups (Supplementary Fig. 2), supporting that its predictive value was not attributable to any single PD-1 inhibitor regimen. Together, these findings suggest that HOPE preserved its utility despite regimen-level variation across participating institutions. Interpretation for HOPE Figure 3 presents Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations for representative patients in the external validation cohort, selected according to HOPE-predicted risk scores. In high-risk cases, the model showed prominent activation at the tumor-liver interface, with attention frequently localized to regions characterized by irregular tumor margins, heterogeneous arterial phase enhancement, and relative hypo-enhancement in the portal phase. In contrast, low-risk cases displayed more diffuse and lower-intensity activation, with limited emphasis on the tumor boundary and a more homogeneous distribution of attention within the surrounding liver parenchyma. Overall, the Grad-CAM patterns differed between risk groups in a visually interpretable manner and were consistent with imaging features commonly used to describe tumor heterogeneity and boundary characteristics, supporting the interpretability of HOPE. To quantitatively support these visual observations, we further calculated tumor-to-liver Grad-CAM ratios in arterial and portal phases in the external validation cohort (Supplementary Fig. 3). Compared with non-HPD cases, HPD cases showed significantly higher arterial-phase and portal-phase tumor-to-liver ratios (p = 0.042 and p = 0.003, respectively), indicating greater concentration of model attribution within the tumor relative to the surrounding liver parenchyma. Together, these findings support that HOPE preferentially focuses on tumor-centered and peritumoral imaging patterns associated with HPD risk.Fig. 3Histogram of HOPE-computed scores in the external validation cohort (middle), accompanied by the Grad-CAM visualizations of HOPE for predicted high-risk cases (left) and predicted high-risk cases (right).Red/yellow highlighted areas indicate the regions of interest emphasized by the model. Histogram of HOPE-computed scores in the external validation cohort (middle), accompanied by the Grad-CAM visualizations of HOPE for predicted high-risk cases (left) and predicted high-risk cases (right). Red/yellow highlighted areas indicate the regions of interest emphasized by the model. Prognostic risk stratification Under the prespecified sensitivity-constrained threshold, the confusion matrix in the external validation cohort indicated that HOPE was more effective for ruling out future HPD than for confirming it, as reflected by its high negative predictive value (NPV) (91.6%) despite a modest positive predictive value (PPV) (29.6%) (Fig. 1B). Calibration analysis in the external cohort showed acceptable agreement between predicted and observed event probabilities (Supplementary Fig. 1A), with a Brier score of 0.127, supporting the use of HOPE probabilities for downstream risk stratification. The prognostic value of HOPE-derived risk stratification is shown in Fig. 4. In both the internal and external validation cohorts, patients categorized as high-risk experienced significantly worse HPD-free survival (HFS) compared with those categorized as low-risk. In the internal validation cohort, the high-risk group had an approximately five-fold higher hazard of HPD (hazard ratio [HR] 5.10; 95% CI 1.62–16.02; p = 0.005), with clear early separation of the survival curves (Fig. 4A). In the external validation cohort, the high-risk group similarly showed an approximately four-fold higher hazard of HPD (HR 3.89; 95% CI 1.71–8.81; p = 0.001; Fig. 4B), demonstrating the reproducibility of HOPE-based risk stratification across institutions. Beyond HPD risk stratification, HOPE showed a modest association with time-to-event outcomes. Patients classified as high-risk by HOPE tended to have shorter disease-free survival and overall survival (Supplementary Figs. 4, 5), broadly paralleling the separation between the actual HPD and non-HPD groups. However, in multivariable Cox regression adjusting for actual HPD status and clinicopathological factors (Supplementary Tables 10, 11), the HOPE score was no longer independently associated with disease-free survival (DFS) (p = 0.250) or overall survival (OS) (p = 0.980), whereas actual HPD remained independently associated with both DFS (p < 0.001) and OS (p < 0.001). These findings suggest that the survival separation across HOPE-derived risk groups was largely driven by enrichment of patients with true HPD, rather than independent prognostic value beyond HPD status.Fig. 4The Kaplan–Meier curves according to the HOPE predicted hyperprogressive disease.HPD-free survival analysis based on HOPE predictions in the internal (A) and external validation cohorts (B). The Kaplan–Meier curves according to the HOPE predicted hyperprogressive disease. HPD-free survival analysis based on HOPE predictions in the internal (A) and external validation cohorts (B). Discussion HPD represents a particularly adverse pattern of treatment failure, characterized by rapid tumor acceleration and profoundly shortened survival, yet remains difficult to anticipate with conventional clinical assessment alone. In this multicenter, retrospective study, we developed and validated an HPD prediction model, namely HOPE, using transformer-based multimodal deep learning framework for prognostic risk stratification of patients with Ad-HCC undergoing PD-1 inhibitor-based triple therapy. By jointly integrating arterial and portal phase CT images with demographic, laboratory, tumor morphological, and radiological factors through a bidirectional multimodal attention architecture, HOPE achieved promising and interpretable prediction of HPD risk. To our knowledge, HOPE is the first deep learning model specifically designed to predict HPD in Ad-HCC using a multimodal fusion strategy grounded in transformer architectures, addressing a critical gap in precision immuno-oncology. HOPE showed good discrimination in the internal validation cohort (AUC = 0.801), whereas performance declined in the external validation cohort (AUC = 0.687). This reduction likely reflects multiple sources of inter-center heterogeneity. Differences in CT acquisition and reconstruction, including scanner vendors, imaging parameters, and slice thickness, may have altered the imaging phenotypes available to the model, which is particularly relevant for a multimodal framework relying on subtle cross-phase patterns. Differences in clinical case-mix across centers, such as Barcelona Clinic Liver Cancer (BCLC) staging, Eastern Cooperative Oncology Group (ECOG), and “capsule” appearance, may also have introduced dataset shift and reduced model transferability. In addition, the distribution of specific PD-1 inhibitors differed significantly across the training, internal validation, and external validation cohorts, which may have contributed modestly to treatment-related heterogeneity across centers. The near-perfect discrimination in the training cohort suggests a possible influence of center-specific patterns, particularly given the limited size of this retrospective dataset. However, HOPE retained acceptable calibration in the external cohort, and decision curve analysis showed positive net benefit across a clinically relevant range of threshold probabilities. Together, these findings suggest that HOPE may still provide clinically meaningful pre-treatment risk stratification; however, its utility and robustness will need to be further established through prospective validation and refinement in larger, more diverse multicenter cohorts, with harmonized imaging workflows, broader clinical and biological feature integration, and continued optimization of the multimodal framework. Compared with baseline models, HOPE demonstrated clear advantages over both traditional machine learning models and unimodal deep learning methods34–37. Clinical models based on clinical factors alone yielded modest discrimination (training AUC ≤ 0.711), consistent with prior reports highlighting the limited predictive value of isolated clinical factors for HPD38–40. Similarly, image-only models such as 3D ResNet-18 and ViT showed overfitting tendencies, particularly ViT, which achieved near-perfect training performance but performed poorly in external validation (AUC 0.556), highlighting the risks of relying solely on imaging without clinical factors. These findings echo broader challenges in medical AI: while deep learning excels at extracting complex patterns from images, it often lacks generalizability when deprived of biological and physiological grounding provided by structured clinical data. HOPE’s superiority appears to derive from its ability to fuse heterogeneous data modalities through a transformer-based architecture that enables dynamic cross-modal interaction. Unlike earlier approaches that concatenate features at a late stage, HOPE introduces early and intermediate multimodal attention modules that enable iterative information exchange between image-derived representations and embedded clinical variables, thereby facilitating joint modeling of complementary cues from both modalities. This design mirrors recent advances in multimodal foundation models in healthcare, such as those proposed by Zhou et al.41 and Lin et al.42, and allows HOPE to more comprehensively capture the underlying mechanisms related to systemic host factors implicated in immune evasion and tumor flare. This interpretation is further supported by the ablation analyses. The findings suggest that the value of HOPE lies not simply in incorporating more variables, but in integrating complementary information across modalities in a way that improves risk estimation. The ablation analyses indicate that both imaging and clinical inputs contributed meaningfully to model performance, and that the fusion strategy itself likely played an important role in capturing relationships that were not preserved in unimodal settings. The additional reduction in performance after removal of demographic, tumor morphological, or laboratory features further supports the importance of multidomain information, which is biologically plausible given the multifactorial nature of hyperprogressive disease. Although use of the largest tumor cross-section is a pragmatic and standardized strategy that has been adopted in prior CT-based deep learning studies43–45, it does not fully capture three-dimensional tumor heterogeneity, multifocal disease burden, or the complete tumor–liver interface. In our dataset, the representative-slice approach appeared sufficient to retain the dominant predictive signal, but this should not be interpreted as evidence that volumetric or multi-lesion approaches are unnecessary. Rather, these findings suggest that a relatively simple 2D representation may be effective in the current setting, while further work is still needed to determine whether more comprehensive spatial representations could improve robustness and generalizability in larger multicenter cohorts. Notably, HOPE builds upon, but significantly extends, prior efforts in HPD prediction across other malignancies. For instance, Vaidya et al. developed a radiomics-based model to predict HPD in non-small cell lung cancer (NSCLC) using baseline CT scans, reporting an AUC of 0.79 in a single-center cohort28. While innovative, this model relied solely on image texture features and did not incorporate key clinical variables such as neutrophil-to-lymphocyte ratio or PD-L1 expression, limiting its biological comprehensiveness. Similarly, Li et al. proposed a radiomic signature for distinguishing HPD from pseudoprogression in NSCLC, achieving AUCs around 0.85, yet again within a single modality and without external validation across institutions29. More recently, Long et al. introduced a pan-cancer HPD prediction model combining clinical and laboratory variables via machine learning, yielding AUCs between 0.70 and 0.78 across several tumor types31. However, this model excluded imaging entirely, neglecting spatial tumor architecture and microenvironmental cues now recognized as central to immunotherapy response. In contrast, HOPE integrates both domains synergistically, achieving higher or comparable performance while maintaining interpretability and external validity. Furthermore, unlike many prior models that use handcrafted radiomic features susceptible to scanner variability and poor reproducibility, HOPE leverages end-to-end deep learning to extract image representations directly from raw pixel data, reducing feature engineering bias and enhancing robustness. Beyond overall performance, HOPE showed stable discrimination across clinically relevant subgroups. The model maintained comparable AUCs across strata defined by AFP level, size of the largest tumor, vascular invasion, metastatic status, and key radiologic phenotypes (e.g., corona enhancement and capsule appearance), whereas baseline clinical and image-only models exhibited more pronounced subgroup-dependent fluctuations. This robustness suggests that HOPE captures higher-order multimodal representations that are less dependent on individual clinicopathologic features and more reflective of underlying biological aggressiveness associated with hyperprogressive disease. From a clinical perspective, such stability is particularly relevant in Ad-HCC, a disease characterized by marked heterogeneity in tumor biology and treatment response. Patients considered for immunotherapy often present with diverse combinations of tumor burden, vascular invasion, and radiological phenotypes, under which conventional risk models may lose reliability. The ability of HOPE to preserve discriminative performance across these heterogeneous subgroups supports its potential generalizability and applicability in routine clinical settings, where patient characteristics cannot be tightly controlled. However, several baseline models demonstrated pronounced subgroup-dependent variability, underscoring their limited robustness in heterogeneous populations. These findings further highlight the advantage of multimodal fusion and cross-modal attention mechanisms in mitigating over-reliance on single-domain features. Subgroup consistency alone does not demonstrate that HOPE predicts HPD independently of baseline tumor burden. However, our additional multivariable logistic regression analyses suggest that the association between HOPE and HPD is unlikely to be explained solely by conventional indicators of tumor aggressiveness. This supports the interpretation that HOPE captures risk-relevant information beyond baseline tumor burden alone. At a biological level, although HPD is ultimately a post-treatment dynamic phenomenon, pretreatment imaging and clinical features may still encode susceptibility to HPD by capturing baseline tumor heterogeneity, vascular phenotype, host condition, and aspects of the tumor immune microenvironment46–49. Recent studies in hepatocellular carcinoma have suggested that baseline imaging features can noninvasively reflect immune-related tumor characteristics46,47 and may also be associated with immunotherapy response48,49. In parallel, radiomics-based pretreatment imaging biomarkers have also been explored for identifying atypical response patterns, including hyperprogression, in other tumor types50. Moreover, emerging mechanistic studies of HPD have implicated pre-existing immunosuppressive contexts and aberrant immune regulation, including regulatory T-cell- and macrophage-related mechanisms, as potential contributors to paradoxical acceleration after immune checkpoint blockade51,52. Within this framework, HOPE should be interpreted as a pretreatment risk-stratification model for HPD susceptibility, rather than a tool that directly quantifies post-treatment growth acceleration itself. Nevertheless, because HOPE relies exclusively on baseline static data, it cannot fully disentangle HPD susceptibility from baseline aggressive tumor biology, nor can it replace longitudinal assessment of tumor growth kinetics after treatment initiation. Therefore, the present model should be viewed as a pretherapeutic tool for identifying patients at increased risk of HPD occurrence, rather than as a direct measure of the acceleration process itself. Future studies integrating pretreatment multimodal features with early on-treatment imaging dynamics, longitudinal tumor growth information, and molecular biomarkers may help better distinguish true HPD-specific susceptibility from baseline disease aggressiveness and further improve biological interpretability51,52. Interpretability is a clinically meaningful strength of HOPE. Using Grad-CAM, we observed that the model’s attribution was primarily distributed over the tumor region and, notably, the tumor-liver interface. High-risk cases showed concentrated activation along the tumor boundary and peritumoral interface, often coinciding with irregular margins, heterogeneous arterial phase enhancement, and relative portal phase washout, imaging phenotypes that have been associated with aggressive pathologic correlates such as microvascular invasion, necrotic stroma, and hypoxia-related markers in Ad-HCC53,54. In contrast, low-risk cases tended to exhibit more diffuse and lower intensity activation with less emphasis on the interface and a more homogeneous distribution within the liver parenchyma. This pattern was supported quantitatively in the external validation cohort, where HPD cases had significantly higher tumor-to-liver Grad-CAM ratios than non-HPD cases in both the arterial and portal phases, indicating greater concentration of model attribution within tumor regions. These findings should be interpreted cautiously. Grad-CAM provides a post hoc visualization of image regions contributing to prediction, but does not establish causal or sufficient imaging determinants of HPD55. We therefore interpret the highlighted regions not as definitive HPD-specific imaging features, but as imaging correlates of a baseline susceptibility state. This interpretation is consistent with the view that HPD is a multifactorial process shaped by tumor-intrinsic alterations, host immune context, and treatment-related effects56, and with evidence that baseline immune status in HCC is associated with subsequent HPD during PD-1 blockade21. Accordingly, HOPE may be capturing tumor phenotype and tumor-host microenvironmental context associated with HPD risk, while Grad-CAM provides supportive, hypothesis-generating evidence that complements the model’s risk prediction. Compared with conventional radiomics, where handcrafted features can be challenging to map back to anatomy, Grad-CAM offers a more direct visual rationale for clinical review. The externally applied sensitivity-constrained threshold reflects a clinical trade-off. By prioritizing avoidance of missed HPD, HOPE was optimized to provide greater reassurance when patients were classified as low risk, as reflected by its high negative predictive value, whereas the more modest positive predictive value indicates that a proportion of patients labeled as high risk would not ultimately develop HPD. This pattern is clinically important because it supports the use of HOPE as an upfront risk-triage tool rather than a standalone determinant of treatment change. In practice, a high-risk prediction may be more appropriately used to prompt closer surveillance, earlier radiologic reassessment and multidisciplinary review, while a low-risk prediction may help identify patients less likely to experience rapid early progression. Beyond this role in early risk triage, HOPE-derived risk stratification may also have broader prognostic relevance, as high-risk patients appeared to experience a consistently poorer clinical course. However, these associations should be interpreted cautiously. Because HOPE was developed to predict HPD rather than long-term survival outcomes, the observed separation in HFS, DFS, and OS likely reflects, at least in part, its ability to identify patients predisposed to subsequent HPD and its adverse clinical trajectory. HOPE-based risk groups may therefore be better viewed as markers of early aggressive disease dynamics rather than as an independent prognostic framework beyond HPD itself. In this context, a pragmatic HOPE-guided clinical pathway may therefore be considered (Fig. 5). Patients classified as low risk may proceed with standard PD-1 inhibitor-based triple therapy, followed by routine follow-up and standard monitoring. By contrast, patients classified as high risk may warrant earlier multidisciplinary reassessment incorporating tumor burden, liver function, and overall clinical condition. In these patients, the predicted risk of HPD may support consideration of alternative therapeutic strategies, including tyrosine kinase inhibitor (TKI) monotherapy, TACE alone, TACE combined with TKIs, or other individualized regimens, together with intensified surveillance such as earlier imaging follow-up and more frequent clinical assessment. Although our retrospective design does not allow determination of whether HOPE-guided management improves outcomes, this framework provides a clinically actionable interpretation of model predictions and a practical basis for future prospective evaluation.Fig. 5A pragmatic HOPE-guided clinical pathway for advanced hepatocellular carcinoma (HCC). A pragmatic HOPE-guided clinical pathway for advanced hepatocellular carcinoma (HCC). Nonetheless, several limitations warrant consideration. First, HOPE was developed using retrospective data. Although the validity and applicability of our approach were supported by rigorous inclusion/exclusion criteria and multicenter evaluation, prospective validation in larger, diverse populations, including patients from different ethnic backgrounds and healthcare systems, is essential before widespread adoption. Second, definitions of HPD remain heterogeneous across studies; we adopted widely used criteria based on tumor growth kinetics15,21, which correlate well with clinical outcomes. In addition, sensitivity analyses using alternative commonly adopted HPD definitions showed that the overall predictive value of HOPE was broadly preserved, although performance varied depending on the specific operational definition applied. Nevertheless, further harmonization of HPD definitions across studies is still needed. Finally, HOPE was trained on baseline imaging and structured clinical data; incorporating longitudinal changes and additional modalities such as genomics, circulating biomarkers, or more granular radiomic descriptors may further enhance performance. In conclusion, HOPE represents an important step toward safer and more individualized use of immunotherapy in Ad-HCC by enabling accurate, multimodal, and interpretable prediction of HPD risk before treatment initiation. By integrating state-of-the-art deep learning with clinically relevant imaging and clinical factors, HOPE outperforms single-modality models and illustrates the potential of transformer-based multimodal architectures for risk assessment in oncology. As immune checkpoint inhibitors become increasingly embedded in treatment algorithms, tools like HOPE will be indispensable for maximizing benefit and minimizing harm. Future work should focus on prospective validation, integration into clinical workflows, and extension to other immunotherapy-treated cancers where HPD poses a significant challenge. Methods This multicenter, retrospective cohort study was approved by the institutional review board Ethical Review Committee of Sun Yat-sen University Cancer Center (approval number B2022-694), with a waiver of the requirement to obtain written informed consent due to the nature of this retrospective study. The study was conducted in accordance with the Declaration of Helsinki and the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis guidelines57. Study population HCC diagnosis was established either pathologically or based on the clinical criteria used by the American Association for the Study of Liver Disease (AASLD) guideline58. Patients who had a good performance status, preserved liver function and had vascular invasion were considered “recommended” or “ideal” candidates for TACE combined with lenvatinib plus a PD-1 inhibitor. The patients with Ad-HCC were identified according to the criteria of BCLC staging systems59. From June 2019 to October 2022, a total of 1124 consecutive patients with Ad-HCC who underwent initial TACE combined with lenvatinib plus PD-1 inhibitor (PD-1 inhibitor-based triple therapy) were retrospectively reviewed at Sun Yat-sen University Cancer Center (Center 1) and Shandong Cancer Center (Center 2). The inclusion criteria were: (a) age 18–75 years; (b) ECOG performance status 1-2; (c) Child-Pugh class A liver function; (d) underwent preoperative CECT within two weeks before the initial TACE of PD-1 inhibitor-based triple therapy. Patients were excluded if they: (a) received additional systemic treatment before PD-1 inhibitor-based triple therapy; (b) had any current or prior malignancies other than HCC; (c) decompensated liver cirrhosis; (d) missing CECT data and (e) without any follow-up information. Figure 6 demonstrates the enrollment pathways of Ad-HCC patients. The distribution of enrolled patients across participating centers is summarized in Supplementary Table 12.Fig. 6Flow of participants through the study according to the inclusion and exclusion criteria. Flow of participants through the study according to the inclusion and exclusion criteria. Treatment procedure All patients received triple therapy consisting of conventional TACE, lenvatinib, and a PD-1 inhibitor. Treatment decisions were made through multidisciplinary discussion at each participating center based on patient performance status, liver function, tumor burden and distribution, and overall clinical condition. The TACE procedure was performed as previously described60,61. In brief, digital subtraction angiography (Philips FD 20, Amsterdam, Netherlands) was used. An arterial sheath was inserted into the femoral artery using the modified Seldinger technique. A 5-Fr Yashiro catheter (Terumo, Tokyo, Japan) was advanced into the celiac trunk and superior mesenteric artery to identify the tumor-feeding arteries. A 2.7-Fr microcatheter (Terumo, Tokyo, Japan) was then superselectively advanced into the feeding arteries as distally as possible. An emulsion, consisting of 10–20 mL of lipiodol, 30–50 mg of platinum-based chemotherapeutic agents, and 20–40 mg of epirubicin, was slowly injected until stasis of flow was achieved in the target vessel. If necessary, additional embolization was performed using gelatin sponge particles mixed with contrast medium until repeat angiography confirmed the absence of residual tumor staining. Across participating centers, TACE was performed according to the same core procedural principles, including superselective catheterization whenever feasible and embolization to angiographic stasis. Repeat TACE was considered according to follow-up imaging findings of residual viable tumor or intrahepatic progression, together with liver function, performance status, and multidisciplinary evaluation. Details of lenvatinib plus PD-1 inhibitor administration are provided in Supplementary Note 1. Data collection Baseline clinical variables were collected before initiation of PD-1 inhibitor-based triple therapy and pretreatment CECT examinations at each participating center. These variables included: (1) demographic and clinical history variables, including age, sex, body mass index, ECOG, pathological differentiation, smoking history, drinking history, comorbidities (i.e., hypertension, diabetes, heart disease, renal disease an esophageal gastric varices, etc.), Hepatic virus infection status (i.e. hepatitis B virus [HBV], hepatitis C virus [HCV]), cirrhosis, Child-Pugh class, ascites, and BCLC stage; (2) laboratory variables, including AFP, des-γ-carboxy prothrombin, albumin (ALB), total bilirubin (TBIL), platelet (PLT), prothrombin time, international normalized ratio, aspartate aminotransferase (AST), alanine aminotransferase (ALT), absolute lymphocyte count (ALC), absolute neutrophil count (ANC), C reactive protein, creatinine (CRE), neutrophils, lymphocyte; and (3) tumor burden and pretreatment CT imaging variables, including tumor number, size of the largest tumor, “capsule” appearance, fat in mass, blood products in mass, necrosis or severe ischemia, corona enhancement, and intratumoral artery. Detailed definitions of CT imaging features are provided in Supplementary Table 13. For patients with multiple tumors, imaging features were assessed on the largest lesion by two fellowship-trained abdominal radiologists independently, with disagreements resolved by a senior radiologist at each center. Not all collected variables were used as model inputs. Variables with substantial missing data were excluded from model development, and only variables with sufficient completeness were considered for model construction. CECT acquisition protocols are provided in Supplementary Note 2. Follow-up and definitions of HPD All patients underwent regular follow up after initial TACE at one month and every three to six months thereafter, each follow up consisting serum AFP and imaging (contrast-enhanced ultrasound, CT or magnetic resonance imaging) collection until death or the last follow-up date (October 31, 2024). The prediction window was the time from initial TACE to HPD, and HPD was defined as a flare-up of tumor growth upon ICI treatment15,56,62. HPD was assessed according to previously published literature using tumor growth dynamics before and after treatment, including tumor growth kinetics (TGK) and tumor growth rate (TGR), with target lesions measured according to RECIST version 1.121,30,63–65. Specifically, TGK was defined as the monthly change in the sum of the longest diameters of target lesions, whereas TGR was defined as the monthly log-scale calibrated change in the sum of the target lesion volumes. The TGK ratio and TGR ratio were calculated as the fold changes in TGK and TGR, respectively, in the experimental period relative to the reference period. The reference period was defined as the time interval between 12 weeks prior to treatment and baseline, and the experimental period was defined as the time interval between treatment initiation and 12 weeks after treatment. The baseline was defined as the time immediately before treatment initiation. HPD was classified using 4-fold increases in TGK and TGR ratios and a 40% increase in TGR as cut-off values21. In addition, for sensitivity analysis, alternative commonly adopted HPD definitions based on 1) TGK ratio >2 and >50% increase in tumor burden65 and 2) TGR ratio >2 and time to treatment failure (TTF) < 2 months30 were also evaluated. To improve the objectivity of outcome assessment, radiological images were retrospectively reviewed by at least two physicians and one independent radiologist according to the prespecified HPD criteria. In cases of discrepant assessments, the final HPD classification was determined by consensus after joint review. During adjudication, post-TACE treatment-related imaging changes were distinguished from true HPD. Expected post-TACE changes included iodized oil deposition, tissue necrosis and absence of perfusion in the arterial phase. By contrast, imaging findings suggestive of HPD included an increase in tumor size, an expansion of the arterial perfusion range in the arterial phase, or the appearance of new intrahepatic or extrahepatic lesions. Accordingly, treatment-related post-TACE changes were not classified as HPD unless they were accompanied by radiological evidence of accelerated tumor progression according to the prespecified HPD criteria. The endpoints were: HFS, defined as the time from initial TACE of PD-1 inhibitor-based triple therapy to first-documented HPD all-cause death; OS, defined as the time from the initiation of PD-1 inhibitor-based triple therapy to death of any cause. Data preprocessing and automatic segmentation To standardize image inputs for subsequent deep learning analyses, all CT volumes underwent a uniform preprocessing pipeline (details provided in Supplementary Note 3). After preprocessing, liver and tumor segmentation was performed simultaneously using a pretrained nnU-Net66–68, with details of the model architecture and implementation provided in Supplementary Fig. 6 and Supplementary Note 4. For each patient, the representative slice was automatically selected as the slice with the largest tumor area based on the nnU-Net segmented tumor mask. The arterial and portal phase images were independently reviewed by two senior radiologists, who verified and, when necessary, manually refined the liver region. Disagreements were resolved by consensus, with adjudication by a third senior radiologist. HOPE construction The overall study design is illustrated in Fig. 7. Patients with Ad-HCC from Center 1 were randomly divided into a training cohort and an internal validation cohort at a ratio of 4:1, whereas patients from Center 2 served as an external validation cohort. HOPE is a transformer-based multimodal deep learning framework, inspired by a recent prognostic study41. It is designed to learn cross-modal representations by converting heterogeneous medical data, including CT images and clinical factors, into a sequence of tokens. Through a transformer-based architecture, the model integrates these multimodal inputs while preserving their complementary semantic and biological relevance, thereby characterizing tumor behavior.Fig. 7Overview of the HOPE workflow and evaluation.A Multicenter 3D images were segmented and converted into 2D arterial (AP) and portal (PP) inputs. B HOPE integrated imaging and clinical factors via transformer-based multimodal attention to predict HPD risk. C Model analysis included ROC curves, confusion matrix, score distribution, subgroup analysis, Grad-CAM visualizations, and Kaplan–Meier survival analysis. Overview of the HOPE workflow and evaluation. A Multicenter 3D images were segmented and converted into 2D arterial (AP) and portal (PP) inputs. B HOPE integrated imaging and clinical factors via transformer-based multimodal attention to predict HPD risk. C Model analysis included ROC curves, confusion matrix, score distribution, subgroup analysis, Grad-CAM visualizations, and Kaplan–Meier survival analysis. Specifically, as shown in Fig. 7, HOPE takes paired 2D CT images from the liver region in the arterial and portal phases as two separate inputs. Each phase is first converted into a sequence of patch tokens through a dedicated embedding module, and the resulting arterial and portal token sequences are concatenated along the token dimension to form a unified image-token sequence. In parallel, structured clinical factors, including demographic, tumor morphological, laboratory, and radiological factors, are encoded via a text embedding module and concatenated to construct a clinical-token sequence. The image-token and clinical-token sequences are subsequently integrated through two bidirectional multimodal attention blocks, enabling reciprocal information exchange between image-derived and clinical tokens while outputting attention weights for interpretability analyses. Following multimodal fusion, the integrated representations were further refined through a series of twelve self-attention blocks, allowing HOPE to capture higher-order interactions and global dependencies across all tokens. The final hidden states were aggregated via mean pooling and passed through a classification head to generate patient-level risk predictions. Detailed descriptions of the bidirectional multimodal attention block and the self-attention block are provided in Supplementary Note 5. Model evaluation Discriminative performance was primarily quantified by AUC, and further characterized using accuracy, sensitivity, specificity, PPV, NPV, positive likelihood ratio (LR+), and negative likelihood ratio (LR-). Confusion matrices were generated to further examine the model’s ability to identify patients who developed HPD. The distribution of HOPE-predicted risk scores was compared across different cohorts to assess its consistency and discriminative capability. Additionally, model calibration of HOPE was assessed using calibration curves and the Brier score. To further evaluate potential clinical usefulness, DCA was performed to quantify the net benefit of the model across different threshold probabilities. The detailed procedures for calibration curve construction and DCA are described in Supplementary Note 6. As an additional sensitivity analysis, we constructed a 2.5D version of HOPE by augmenting the largest tumor slice with neighboring slices at offsets ±1, ±2, and ±4 from both arterial- and portal-phase CT, while keeping the same clinical inputs, multimodal fusion framework, preprocessing pipeline, and training strategy as the original 2D model. To assess the incremental contribution of multimodal fusion in HOPE, ablation experiments were performed. Using the same cohorts, preprocessing procedures, training pipeline, and evaluation strategy as the primary analysis, we constructed a series of HOPE-derived models with reduced input configurations. The full model incorporated CT images together with demographic, tumor morphological, and laboratory factors. To evaluate unimodal performance, we generated an image-only model and a clinical-only model. To assess the contribution of individual clinical factor groups within the multimodal framework, we further constructed three additional variants by excluding demographic, tumor morphological, or laboratory factors, respectively, while retaining image input. Discriminative performance was evaluated in the internal and external validation cohorts using AUCs with 95% confidence intervals. To improve interpretability, Grad-CAM was implemented to visualize imaging regions contributing to HOPE’s predictions. We further quantitatively compared tumor-to-liver Grad-CAM ratios between HPD and non-HPD cases in the external validation cohort, separately for the arterial and portal phases. Group differences were assessed using the Mann–Whitney U test. The clinical significance of HOPE-derived risk stratification was evaluated using survival analysis. HFS was compared between predicted high- and low-risk groups using Kaplan–Meier curves with log-rank tests. Additional survival endpoints, including DFS and OS, were analyzed similarly. In addition, univariable and multivariable Cox proportional hazards regression analyses were performed for DFS and OS to further assess whether the HOPE score provided prognostic information independent of actual HPD status and other clinical factors. The risk-stratification threshold for HOPE was determined using a sensitivity-constrained strategy69, reflecting the clinical priority of avoiding missed HPD. Thresholds from 0 to 1 were scanned at 0.001 increments, and among those achieving sensitivity ≥0.70, the threshold with the highest accuracy was chosen. If no threshold met the sensitivity requirement, the one with the highest achievable sensitivity was selected. Under the selected threshold, classification performance in the validation cohorts was summarized using confusion matrices, from which sensitivity, specificity, positive predictive value, and negative predictive value were derived. In addition, model calibration was evaluated by calibration plots comparing predicted probabilities with observed event rates, and quantified using the Brier score. To further demonstrate the performance gains of multimodal fusion, HOPE was compared with that of single modality models (clinical factors alone or imaging alone). The baseline models included: (1) a clinical model constructed using clinical factors alone (Clin model); (2) a clinical model constructed using clinical and radiological factors (ClinRad model); (3) an image-only 3D ResNet-1870 trained with the same image inputs as HOPE; and (4) an image-only ViT trained with the same image inputs as HOPE. Details of baseline model implementation are provided in Supplementary Note 7. The AUCs of different models were compared using the DeLong test. To assess the generalizability of HOPE and benchmark its performance against baseline models, prespecified subgroup analyses were conducted across key clinical and tumor-related characteristics, including AFP, vascular invasion, size of largest tumor, corona enhancement, metastasis, and “capsule” appearance, in the external validation cohort. Within each prespecified subgroup, model discrimination was quantified by the AUC. For each binary subgroup variable, differences in AUC between the two subgroup strata were assessed using a z-test for independent AUC estimates. In addition, to determine whether HOPE predicted HPD beyond baseline aggressive tumor biology, we performed a multivariable logistic regression analysis with HPD as the dependent variable. Covariates included HOPE score, AFP, size of the largest tumor, vascular invasion, metastasis, corona enhancement and “capsule” appearance. A reduced model without HOPE score and a full model including HOPE score were then compared using AUC and likelihood ratio testing to evaluate the incremental predictive value of HOPE beyond baseline tumor biology. To further account for potential treatment heterogeneity, we also performed an exploratory subgroup analysis according to the specific PD-1 inhibitor used, including Camrelizumab, Sintilimab, Tislelizumab, and Toripalimab. Within each PD-1 inhibitor subgroup, separate multivariable logistic regression models were fitted with HPD as the dependent variable and HOPE score as the main predictor, adjusting for AFP, size of the largest tumor, vascular invasion, metastasis, corona enhancement and “capsule” appearance. Statistical analysis Inter-rater agreement between the two radiologists was assessed with the intraclass correlation coefficient for tumor size and with Cohen’s kappa value for binary imaging features. Continuous variables were summarized as mean ± standard deviation or as median with IQRs. Comparisons of continuous variables were performed using the Student’s t test, analysis of variance (ANOVA), the Mann–Whitney U test or Kruskal–Wallis test. Categorical variables were presented as counts and percentages and were compared using the χ² test or Fisher’s exact test, as appropriate. AUCs of different models were compared using the DeLong test to evaluate their predictive performance for HPD. The cumulative survival outcomes of different risk groups were estimated using the Kaplan–Meier method and compared using the log-rank test. Statistical analysis was performed using SPSS version 25.0 (IBM Corp., NY, USA) and the RMS package of R software version 4.5.1 (http://www.r-project.org/). All statistical tests were two-sided, and p value <0.05 was considered statistically significant. Supplementary information Supplementary information Supplementary information