Machine learning outperforms large language models for survival prediction in advanced hepatocellular carcinoma: a multicenter study Authors: Jiacheng Tan, Yangyang Li, Fengtao Zhang, Kai Sun, Qingkang Zheng, Huanzhang Niu Published: 2026 ABSTRACT AbstractAccurate prognostic prediction remains a critical unmet need in advanced hepatocellular carcinoma (HCC). While machine learning (ML) models have demonstrated value in outcome prediction, the ability of large language models (LLMs) to perform structured clinical prognostic tasks remains unclear. In this multicenter retrospective study, 1031 patients with HCC receiving interventional therapy combined with targeted treatment were analyzed and randomly divided into training (n = 717) and test (n = 314) cohorts. Six ML algorithms were developed and compared with two LLMs (ChatGPT-4o and DeepSeek-v3). Seven variables, including age, comorbidities, albumin–bilirubin grade, tumor burden, portal vein tumor thrombus, and alpha-fetoprotein level, were identified as key predictors. In the test cohort, SVM achieved the highest performance (AUC = 0.658), followed by XGBoost (AUC = 0.654), whereas LLMs showed limited discriminative ability (AUC = 0.590–0.591; P < 0.05 vs ML). ML models effectively stratified patients into risk groups with significantly different 1-year survival, while LLM-based predictions failed to distinguish outcomes. These findings indicate that ML models outperform current LLMs in structured prognostic prediction and provide more reliable support for risk stratification and clinical decision-making in advanced HCC.Keywords: Hepatocellular carcinoma, Large language models, Machine learningSubject terms: Biomarkers, Cancer, Computational biology and bioinformatics, Oncology, Risk factors Introduction Hepatocellular carcinoma (HCC) is one of the most prevalent malignancies worldwide and remains a major cause of cancer-related mortality. Over the past decades, both the incidence and mortality of HCC have continued to rise. According to global cancer statistics, approximately 900,000 new cases of HCC are diagnosed annually, ranking sixth among all cancers in incidence1. Meanwhile, HCC accounts for nearly 750,000 deaths each year, making it the third leading cause of cancer-related mortality worldwide2. China bears the greatest burden of HCC, contributing to more than half of newly diagnosed cases globally3. Despite continuous advances in diagnostic techniques and therapeutic strategies, the prognosis of patients with HCC remains poor. A major challenge in HCC management is that early-stage disease is frequently asymptomatic. In combination with limited public awareness and insufficient screening programs, this results in many patients being diagnosed at advanced stages, when curative surgical resection is no longer feasible4. For patients with advanced HCC, interventional therapy combined with targeted agents has become an important treatment strategy in clinical practice. Although this combined regimen has improved outcomes for a subset of patients, overall survival remains unsatisfactory, reflecting considerable heterogeneity in treatment response5. Therefore, accurately identifying patients who are most likely to benefit from specific therapeutic strategies is essential for optimizing individualized treatment and improving clinical outcomes. Recent advances in artificial intelligence (AI) have created new opportunities for improving clinical decision-making. By integrating and analyzing multidimensional medical data, AI systems can identify complex patterns across clinical characteristics, imaging findings, and pathological information6. Among AI approaches, machine learning has demonstrated strong performance in tasks such as disease classification, risk stratification, and prognosis prediction through iterative model training and parameter optimization. With the continuous growth of clinical datasets and improvements in computational power, machine learning models have shown increasing potential for clinical applications7. In the field of HCC, machine learning–based predictive models have been explored for diagnosis, treatment response assessment, and survival prediction. For example, Wei et al.8 applied machine learning methods to predict disease progression in patients with intermediate-stage HCC following transarterial chemoembolization (TACE), demonstrating that the CatBoost model outperformed conventional staging systems in predicting treatment outcomes. Similarly, a systematic review by Feng et al.9 highlighted the growing value of machine learning in improving diagnostic accuracy and prognostic evaluation in HCC. Nevertheless, despite these promising advances, several challenges remain for the practical implementation of machine learning models. Issues related to model complexity, interpretability, and clinical applicability continue to limit their widespread adoption in routine practice. Consequently, there is a need for more efficient and accessible predictive approaches that can support personalized treatment decision-making. Currently, the emergence of large language models (LLMs) has introduced a new paradigm for medical data analysis. LLMs possess powerful capabilities in semantic understanding and contextual reasoning, enabling them to extract latent information from complex clinical data and integrate diverse sources of medical knowledge10,11. In addition to their advantages in processing unstructured clinical text, LLM-based systems provide intuitive natural-language interfaces that substantially reduce the technical barriers associated with traditional modeling approaches. These features may facilitate broader clinical adoption, particularly in healthcare settings with limited computational expertise12. Preliminary studies have begun to explore the clinical potential of LLMs. For instance, Xu et al.13 reported that LLMs could outperform physicians in predicting immunotherapy response in patients with HCC. However, whether LLMs can effectively model structured clinical data and achieve predictive performance comparable to or exceeding that of traditional machine learning algorithms remains largely unknown. Therefore, the present study aimed to investigate the potential of LLMs for prognostic modeling in patients with HCC. We developed an LLM-based model to predict one-year survival in patients with HCC undergoing interventional therapy combined with targeted treatment. By systematically comparing its predictive performance with multiple traditional machine learning algorithms, we sought to evaluate whether general-purpose LLMs could serve as an alternative approach for structured clinical data–based risk prediction. Our findings may provide new insights into the role of LLMs in clinical prognostic modeling and offer a practical framework for personalized outcome prediction in advanced HCC. Materials and methods Patients selection This multicenter retrospective study included patients with advanced hepatocellular carcinoma (HCC) who received interventional therapy combined with targeted treatment at two tertiary hospitals in China between January 2019 and June 2022. A total of 1,168 eligible patients were enrolled. All patients underwent either transarterial chemoembolization (TACE) combined with targeted therapy or hepatic artery infusion chemotherapy (HAIC) combined with targeted therapy. The inclusion criteria were as follows: (1) histologically confirmed HCC or radiologically diagnosed HCC based on contrast-enhanced computed tomography (CT) or magnetic resonance imaging (MRI) according to the American Association for the Study of Liver Diseases (AASLD) guidelines; (2) age ≥ 18 years; (3) Eastern Cooperative Oncology Group performance status (ECOG PS) score of 0–1; and (4) Barcelona Clinic Liver Cancer (BCLC) stage B or C. The exclusion criteria were: (1) loss to follow-up; (2) incomplete clinical data; (3) prior anticancer treatment; and (4) concomitant malignancies other than HCC. All patients underwent contrast-enhanced CT or MRI within one week prior to treatment initiation for baseline imaging assessment. Patients with incomplete key clinical variables were excluded prior to model development. The study was designed based on the Helsinki Declaration and authorized by the Ethics Committee of the First Affiliated Hospital of Henan University of Science and Technology. As the study used de-identified retrospective data, the requirement for informed consent was waived by the ethics committees. Clinical data collection Clinical variables were extracted from electronic medical records and included demographic characteristics, clinical status, laboratory parameters, tumor burden, and treatment information. The variables analyzed were age, sex, ECOG performance status, comorbidities, hepatitis history, alanine aminotransferase (ALT), aspartate aminotransferase (AST), albumin–bilirubin (ALBI) grade, platelet count, creatinine level, Child–Pugh class, maximum tumor diameter, tumor number, portal vein tumor thrombus, distant metastasis status, BCLC stage, alpha-fetoprotein (AFP) level, and treatment regimen. Patients were followed regularly after treatment to record overall survival (OS), defined as the time from treatment initiation to death from any cause or the last follow-up14. Model development and evaluation The dataset was randomly divided into a training set and a test set at a ratio of 7:3. The training set was used for model development, while the test set served as an internal validation cohort to evaluate model performance. Feature selection was performed using the Least Absolute Shrinkage and Selection Operator (LASSO) regression method. Variables with non-zero coefficients were retained for model construction. Based on the selected features, multiple machine learning models were developed, including Logistic Regression, Random Forest, K-Nearest Neighbors (KNN), eXtreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), Support Vector Machine (SVM), and Neural Network. Hyperparameter tuning was performed using grid search combined with 10-fold cross-validation within the training set. Model performance was evaluated based on the area under the receiver operating characteristic curve (AUC). To address potential class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied to the training data during cross-validation. Large language model-based prediction Two large language models (LLMs), ChatGPT-4o and DeepSeek-v3, were evaluated for survival prediction. For each patient, structured clinical variables were converted into standardized textual summaries and provided to the LLMs together with a predefined prediction query. All patient information was de-identified prior to analysis. A zero-shot prompting strategy was adopted to assess the intrinsic predictive capabilities of the LLMs without task-specific fine-tuning. Each prompt consisted of a structured summary of the patient’s clinical characteristics followed by a binary prediction question regarding 1-year survival. All structured clinical variables were converted into textual inputs using a predefined and standardized template. The order, wording, and formatting were kept identical across all patients to minimize prompt-induced bias and ensure reproducibility. All patient data were anonymized prior to analysis, with removal of identifiable information and assignment of unique study IDs. An example prompt is provided below: Patient Information: Male, 48 years old, diagnosed with HCC. ECOG PS = 1. Chronic hepatitis B infection; no other comorbidities. Laboratory findings: ALT 60 U/L, AST 86 U/L, AFP > 400 ng/mL, ALBI score − 2.68, platelets 222 × 109/L, creatinine 60.8 µmol/L, Child-Pugh class A. Imaging: largest tumor 10.6 cm, multifocal intrahepatic disease, portal vein tumor thrombus present, no distant metastasis. BCLC stage C. Treated with TACE plus targeted therapy. Question: Did this patient survive beyond one year after initiating interventional combined targeted therapy? Please respond with “Yes” or “No,” followed by a brief explanation. To improve prediction stability, each LLM independently evaluated every patient three times. The final prediction was determined using a majority voting strategy across the three responses. Model evaluation The predictive performance of all models was assessed using the test dataset. Receiver operating characteristic (ROC) curves were generated, and the area under the curve (AUC) was calculated to evaluate model discrimination ability. Additional evaluation metrics included accuracy, precision, recall, F1 score, calibration curves, and confusion matrices. Statistical analysis Continuous variables were expressed as mean ± standard deviation or median with interquartile range, depending on data distribution, and compared using t-tests or Wilcoxon rank-sum tests. Categorical variables were expressed as counts and percentages and compared using chi-square or Fisher’s exact tests. Differences in AUC values were compared using the DeLong test. Optimal cutoff values were determined using the maximum Youden index. Survival curves were estimated using the Kaplan–Meier method and compared using the log-rank test. All analyses were performed using R software (RStudio, version 4.5.1). A two-sided P value < 0.05 was considered statistically significant. Results Patient characteristics This multicenter retrospective study initially identified 1168 patients with hepatocellular carcinoma (HCC) who received interventional therapy combined with targeted treatment between January 2019 and June 2022. After applying the exclusion criteria, 21 patients lost to follow-up, 27 with incomplete clinical data, 53 receiving alternative treatment regimens, and 36 with concomitant malignancies were excluded. Ultimately, 1031 patients were included in the final analysis. The patient selection process is illustrated in Fig. 1. Flowchart of patient selection in this study. Among the included patients, 898 (87.6%) were male. Hepatitis virus infection was present in 969 patients (94.0%), including 957 cases (92.8%) of hepatitis B virus infection. Portal vein tumor thrombus was observed in 582 patients (56.5%), and 679 patients (65.9%) had multiple intrahepatic tumors. Distant metastasis was identified in 428 patients (41.8%). A total of 604 patients (58.6%) received TACE combined with targeted therapy, while the remaining patients received HAIC-based treatment. The dataset was randomly divided into a training set (717 patients) and a test set (314 patients). Baseline characteristics of the training set and test set are summarized in Table 1. Table 1Comparison of clinical characteristics between patients in the training set and the test set.Training setTest setDeathSurvival P DeathSurvival P n = 394n = 323n = 175n = 139Age (%)< 60292 (74.1)208 (64.4)0.006121 (69.1)85 (61.20.173>=60102 (25.9)115 (35.6)54 (30.9)54 (38.8)Gender (%)Female50 (12.7)43 (13.3)0.89321 (12.0)19 (13.7)0.787Male344 (87.3)280 (86.7)154 (88.0)120 (86.3)ECOG (%)0368 (93.4)305 (94.4)0.679158 (90.3)128 (92.1)0.721126 ( 6.6)18 ( 5.6)17 ( 9.7)11 ( 7.9)Comorbidities (%)Absence362 (91.9)284 (87.9)0.102166 (94.9)114 (82.0)0.001Presence32 ( 8.1)39 (12.1)9 ( 5.1)25 (18.0)Hepatitis (%)Absence18 ( 4.6)18 ( 5.6)0.4628 ( 4.6)18 (12.9)0.025B374 (94.9)301 (93.2)163 (93.1)119 (85.6)C2 ( 0.5)4 ( 1.2)4 ( 2.3)2 ( 1.4)ALT (mean (SD))64.76 (129.57)67.91 (101.17)0.72264.32 (69.59)63.81 (61.65)0.946AST (mean (SD))101.75 (100.07)91.33 (182.69)0.333102.44 (88.93)78.22 (77.76)0.012ALBI grade (%)1143 (36.3)183 (56.7)< 0.00164 (36.6)70 (50.4)0.0192251 (63.7)140 (43.3)111 (63.4)69 (49.6)PLT (mean (SD))232.71 (109.66)222.53 (123.81)0.244235.62 (115.68)228.85 (128.41)0.624Cre (mean (SD))70.68 (18.39)72.99 (16.35)0.07974.25 (55.93)73.94 (44.05)0.957CTP (%)A344 (87.3)292 (90.4)0.237149 (85.1)120 (86.3)0.892B50 (12.7)31 ( 9.6)26 (14.9)19 (13.7)Tumor size (mean (SD))11.10 (3.74)9.74 (3.26)< 0.00111.04 (3.80)9.74 (3.30)0.002Tumor number (%)Single118 (29.9)126 (39.0)0.01452 (29.7)56 (40.3)0.066Multiple276 (70.1)197 (61.0)123 (70.3)83 (59.7)PVTT (%)Absence150 (38.1)174 (53.9)< 0.00167 (38.3)78 (56.1)0.002Presence244 (61.9)149 (46.1)108 (61.7)61 (43.9)Metastasis (%)Absence230 (58.4)195 (60.4)0.64296 (54.9)82 (59.0)0.535Presence164 (41.6)128 (39.6)79 (45.1)57 (41.0)BCLC stage (%)B68 (17.3)82 (25.4)0.01032 (18.3)39 (28.1)0.055C326 (82.7)241 (74.6)143 (81.7)100 (71.9)AFP level (%)<=400133 (33.8)156 (48.3)< 0.00175 (42.9)65 (46.8)0.564> 400261 (66.2)167 (51.7)100 (57.1)74 (53.2)Treatment regime (%)HAIC170 (43.1)131 (40.6)0.53370 (40.0)56 (40.3)1.000TACE224 (56.9)192 (59.4)105 (60.0)83 (59.7)ALT alanine aminotransferase, AST aspartate aminotransferase, ALBI albumin-bilirubin, PLT platelets, Cre creatinine, CTP Child Turcotte Pugh, PVTT portal vein tumor thrombosis, BCLC Barcelona clinic liver cancer, AFP α fetoprotein, HAIC hepatic arterial infusion chemotherapy, TACE transcatheter arterial chemoembolization. Comparison of clinical characteristics between patients in the training set and the test set. ALT alanine aminotransferase, AST aspartate aminotransferase, ALBI albumin-bilirubin, PLT platelets, Cre creatinine, CTP Child Turcotte Pugh, PVTT portal vein tumor thrombosis, BCLC Barcelona clinic liver cancer, AFP α fetoprotein, HAIC hepatic arterial infusion chemotherapy, TACE transcatheter arterial chemoembolization. Feature selection Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied for feature selection. The optimal lambda value was selected based on the minimum cross-validation error, and seven variables with non-zero coefficients were retained as key predictors: age, comorbidities, ALBI grade, tumor size, tumor number, portal vein tumor thrombus, and AFP level (Fig. 2). LASSO feature selection for model construction. (A) AUC as a function of the logarithm of the lambda parameter for lasso regression. (B) The path of coefficients in a LASSO regression model as the regularization parameter - lambda. LASSO(least absolute shrinkage and selection operator). Model performance In the training set, the Random Forest model achieved the highest discriminative performance among machine learning models, with an area under the receiver operating characteristic curve (AUC) of 0.850. In contrast, the large language models (LLMs) showed substantially lower predictive performance (ChatGPT-4o: AUC 0.513; DeepSeek-v3: AUC 0.525). In the test set, the SVM and Logistic Regression models demonstrated the highest predictive performance (AUC 0.658), followed by XGBoost (AUC 0.654). The LLMs again exhibited limited discriminative ability, with AUC values of 0.520 for ChatGPT-4o and 0.535 for DeepSeek-v3 (Fig. 3). Comparison of ROC curves of the models. ROC receiver operating characteristic, AUC area under the curve, KNN K-nearest neighbors, XGBoost eXtreme gradient boosting, GBM gradient boosting machine, SVM support vector machine. Pairwise comparisons using the DeLong test showed that only the difference between XGBoost and GBM among machine learning models was statistically significant (P = 0.023), whereas no significant differences were observed among the remaining models. In contrast, the AUC values of nearly all machine learning models were significantly higher than those of both LLMs (all P < 0.05). No significant difference was observed between the two LLMs (P = 0.480). Calibration analysis showed that machine learning models demonstrated better agreement between predicted and observed outcomes than LLM-based models. On the test set, the SVM model achieved an accuracy of 0.602, precision of 0.543, recall of 0.633, and an F1 score of 0.585 (Table 2; Fig. 4). The confusion matrix for the test set is presented in Fig. 6. Overall, the SVM model demonstrated the most stable predictive performance among the evaluated approaches. Table 2Comparison of predict performance between the models.ModelAccuracyPrecisionRecallF1 ScoreAUC (95%CI)RandomForest0.6020.5630.4530.5020.633 (0.571–0.695)KNN0.5960.5440.5320.5380.622 (0.559–0.684)XGBoost0.6120.5620.550.5580.654 (0.593–0.714)GBM0.5860.5320.5400.5360.623 (0.561–0.685)SVM0.6020.5430.6330.5850.658 (0.598–0.718)NeuralNet0.5960.5450.5250.5350.614 (0.551–0.678)Logistic0.6150.5600.6040.5810.658 (0.597–0.718)Chatgpt-4o0.6020.6030.3000.3960.590 (0.529–0.652)Deepseek-v30.5920.5450.4820.5120.591 (0.528–0.653) Comparison of predict performance between the models. Calibration curves for each model. (A) Random forest; (B) KNN; (C) XGBoost; (D) GBM; (E) SVM; (F) neural network; (G) logistic regression; (H) ChatGPT-4o; (I) DeepSeek-v3. The confusion matrix of each model predicts the one-year survival rate in the test set. (A) Random forest; (B) KNN; (C) XGBoost; (D) GBM; (E) SVM; (F) neural network; (G) logistic regression; (H) ChatGPT-4o; (I) DeepSeek-v3. Risk stratification and survival analysis Based on the optimal cutoff values derived from ROC analysis, patients were stratified into high-risk and low-risk groups according to predicted survival probabilities.Kaplan–Meier survival analysis demonstrated significant differences in 1-year survival between the high-risk and low-risk groups defined by all machine learning models, with substantially lower survival rates observed in the high-risk subgroup. In contrast, no significant differences in 1-year survival were observed between risk groups defined by the LLM-based models (ChatGPT-4o: P = 0.688; DeepSeek-v3: P = 0.196) (Figs. 5, 6). Kaplan–Meier curves for different risk stratifications of each model. (A) Random forest; (B) KNN; (C) XGBoost; (D) GBM; (E) SVM; (F) neural network; (G) logistic regression; (H) ChatGPT-4o; (I) DeepSeek-v3. These findings indicate that machine learning models provide superior discriminative performance and risk stratification ability compared with LLM-based approaches in predicting 1-year survival in patients with HCC receiving interventional therapy combined with targeted treatment. Discussion This study systematically evaluated the predictive performance of traditional machine learning (ML) algorithms and large language models (LLMs) for 1-year survival prediction in patients with advanced hepatocellular carcinoma (HCC) undergoing interventional therapy combined with targeted treatment. Using multicenter data from 1,031 patients, we constructed a predictive framework incorporating six ML algorithms and two LLMs. Our findings demonstrate that ML models significantly outperformed LLMs in both predictive discrimination and risk stratification. Notably, ML-based models were able to effectively separate patients into high- and low-risk groups with significantly different survival outcomes, whereas LLM-based predictions failed to achieve meaningful survival discrimination. To our knowledge, this is among the first studies directly comparing LLMs and ML algorithms for survival prediction in HCC patients receiving interventional combined targeted therapy. Machine learning models have demonstrated increasing value in clinical decision-making across multiple domains of oncology. By learning patterns from high-dimensional clinical data, ML algorithms can identify complex relationships between patient characteristics and outcomes that may not be easily captured by traditional statistical approaches15. In the present study, several ML models showed moderate predictive performance, with the SVM model achieving the most stable results in the test cohort. The moderate predictive performance observed reflects the inherent heterogeneity of advanced HCC and the limitation of using structured clinical variables alone. Its performance in our study is consistent with previous research demonstrating the utility of ML approaches in predicting oncological outcomes16,17. Recent advances in artificial intelligence have generated considerable interest in the potential clinical applications of large language models. Owing to their strong capabilities in semantic understanding and contextual reasoning, LLMs can process complex textual information and generate flexible responses to medical queries18,19. Their natural-language interfaces and adaptability through large-scale pretraining have led to expectations that they may support clinical decision-making. Indeed, previous studies have shown that LLMs can achieve strong performance in certain medical knowledge tasks and even reach passing thresholds on medical licensing examinations20,21. These findings highlight their potential utility in knowledge-based reasoning and information retrieval within healthcare systems. However, the results of the present study indicate that these advantages do not necessarily translate into superior performance in structured clinical prognostic prediction. In our cohort, LLMs showed limited discriminative ability for predicting 1-year survival and were unable to effectively stratify patients according to survival risk. One possible explanation is that current LLMs are primarily designed for language understanding and generation rather than quantitative outcome prediction. In contrast, traditional ML algorithms are specifically optimized to learn statistical relationships between structured clinical variables and target outcomes. Consequently, ML models may be better suited to capturing the complex numerical patterns underlying survival prediction tasks22–24. But Importantly, the LLMs evaluated in this study were general-purpose models and were not specifically trained on biomedical or clinical datasets. Recent advances in domain-adapted LLMs, such as those fine-tuned on large-scale medical corpora, have demonstrated improved performance in certain clinical tasks25. Therefore, the relatively limited performance observed in our study may not fully reflect the potential of specialized medical LLMs for structured prognostic prediction. In addition, the use of a zero-shot prompting strategy in this study may have contributed to the relatively limited performance of LLMs. This approach was intentionally adopted to evaluate the intrinsic reasoning capability of LLMs without task-specific fine-tuning or example-based guidance. However, prior studies have demonstrated that LLM performance can be significantly improved through prompt engineering, few-shot learning, or domain-specific fine-tuning26,27. Therefore, the results observed in this study may underestimate the full potential of LLMs in clinical prediction tasks. Another important observation was the variability of LLM predictions. Even when identical clinical information was provided repeatedly, the generated responses occasionally differed, reflecting the stochastic nature of generative language models28. Although we attempted to mitigate this limitation by applying a majority-voting strategy across repeated predictions, such variability may reduce the reliability of LLM-based predictions in clinical settings. In contrast, once trained, conventional ML models produce deterministic outputs, which improves reproducibility and may enhance their suitability for clinical decision support. Consistent with previous studies, our analysis also confirmed the prognostic value of several established clinical factors. Feature selection using LASSO regression identified age, comorbidities, ALBI grade, tumor size, tumor number, portal vein tumor thrombus, and AFP level as key predictors of 1-year survival. These findings align with existing evidence indicating that tumor burden and liver function remain critical determinants of prognosis in patients with advanced HCC29,30. Interestingly, certain commonly used indicators, including Child–Pugh classification and BCLC stage, were not retained in the final model. This may reflect differences in variable importance within specific patient populations and highlights the data-driven nature of ML-based feature selection. Our findings are broadly consistent with previous studies highlighting the prognostic value of machine learning in HCC. For example, Li et al.31 developed an ML-based model that outperformed conventional scoring systems in predicting post-transplant recurrence risk in HCC patients. Similarly, Fang et al.32 identified an anoikis–ErbB signaling-related prognostic signature using a LASSO-based ML framework, demonstrating robust predictive ability across multiple cohorts. Other studies have also applied ML models to radiomic and genomic datasets to improve prognostic stratification in cancer patients33. In contrast, the clinical utility of LLMs for outcome prediction remains uncertain. While some studies have reported that LLMs perform comparably to physicians in specific prediction tasks, other investigations have shown limited sensitivity in high-risk clinical scenarios13,34. Taken together, current evidence suggests that LLMs may be more suitable for knowledge-based reasoning and clinical documentation tasks, whereas ML algorithms remain more reliable tools for quantitative prognostic modeling. Several limitations of this study should be acknowledged. First, the retrospective design may introduce potential selection bias despite the use of multicenter data, and may limit the generalizability of the findings. Second, the overall predictive performance of the models remained moderate, suggesting that additional improvements may be achieved through larger datasets and integration of multimodal data such as radiomic or genomic features. Third, the LLMs evaluated in this study were general-purpose models and were not specifically fine-tuned for medical prediction tasks, which may have limited their performance. Finally, the absence of an independent external validation cohort may restrict the generalizability of our findings. In conclusion, traditional ML models demonstrated superior predictive performance compared with large language models for predicting 1-year survival in patients with advanced HCC undergoing interventional therapy combined with targeted treatment. ML-based models also enabled effective risk stratification, highlighting their potential value for personalized prognosis assessment and treatment planning. Future studies integrating larger multicenter datasets, multimodal clinical information, and domain-specific model optimization are warranted to further improve predictive accuracy and evaluate the clinical utility of artificial intelligence–based prognostic tools. Author contributions J. T. and Y. L. contributed equally to this work. F.Z. and Y. L. contributed to the conception and design of this research, as well as the analysis and interpretation of the data. J. T. drafted the initial version. Y. L. , K.S. and Q.Z. provided the clinical data of the patients. H.N reviewed the manuscript and gave it final approval. All authors read and approved the manuscript. Data availability All data in this study are available through contacting the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Jiacheng Tan and Yangyang Li. References Associated Data Data Availability Statement All data in this study are available through contacting the corresponding author on reasonable request.