PHENO-RAG: An artificial intelligence tool for guideline-informed management decisions in hepatocellular carcinoma.

Introduction
Hepatocellular carcinoma (HCC) represents the sixth most prevalent cancer and the third leading cause of cancer mortality, imposing a major health burden worldwide.1 Unlike other malignancies, challenges associated with the management of HCC are related to several factors: 1) it develops primarily in the context of chronic liver disease, mainly cirrhosis; 2) diagnosis is usually radiological, rather than histopathological; 3) the co-existence with underlying cirrhosis influences prognosis and treatment outcomes, due to competing risks between oncological outcomes and liver-related events;2 4) the availability of multiple treatment options – including liver transplantation, surgical resection, ablative therapies, transarterial treatments, and systemic therapies, representing a challenge in risk stratification and identification of best treatment sequence. All in all, this leads to a complex decision-making landscape for clinicians.3
Evidence from retrospective studies consistently demonstrates that multidisciplinary team (MDT) management is significantly associated with patient outcomes in HCC.4,5 The establishment and implementation of MDT care was shown to be significantly associated with improved survival across different healthcare settings.[6], [7], [8], [9] The reasons for the survival benefit associated with MDT management are not fully understood and may include improved personalization of care through multiparametric decision-making, enhanced access to clinical trials, and shorter time from diagnosis to treatment initiation.
However, the implementation of universal MDT discussion for all patients with HCC presents significant practical challenges. There is a lack of standardization in the definition of the core MDT and its procedures. Moreover, the process is resource-intensive and time-consuming, and not all cases may require the same level of multidisciplinary expertise. Some treatment decisions might be relatively straightforward and can be made by single specialists, while others demand complex multi-modal approaches. This underscores the need for innovative tools that can effectively stratify patients based on their clinical complexity, identifying those who would derive the greatest benefit from a comprehensive MDT approach.
Recent advances in artificial intelligence, particularly large language models (LLMs), have shown promising potential in clinical decision support and guideline interpretation.10,11 These transformer-based neural networks, trained on large datasets from unstructured and structured sources, can process and analyse complex clinical information while accounting for multiple variables simultaneously. Recent studies have demonstrated that LLMs, when enhanced with domain-specific knowledge through techniques such as retrieval augmented generation (RAG) or supervised fine-tuning, can achieve high accuracy rates in guideline-based decision making.12,13 Conversely, foundational LLMs have demonstrated unacceptably low accuracy in gastroenterology and hepatology clinical tasks, potentially leading to suboptimal decision making, when used by patients or non-specialists.14 Therefore, these implementations require careful framework design, including structured guideline reformatting and advanced prompt engineering to ensure reliable outputs.
Despite these advances, there exists a significant knowledge gap in the application of LLMs to HCC decision-making. While studies have explored LLM applications in different medical specialties, none have specifically addressed their potential role in stratifying patients with HCC by clinical complexity or predicting optimal treatment decisions using real-world clinical data. The unstructured nature of real-life clinical reports presents distinct challenges compared to artificially constructed clinical cases. Navigating unstructured clinical real-world data requires LLMs to first extract relevant clinical concepts before applying decision frameworks – a two-step process with the potential for compounding errors. Therefore, we sought to leverage LLM capability for clinical concept extraction from unstructured notes to create structured clinical cases containing information relevant for HCC management.
To address this gap, we developed PHENO-RAG, a modular ensemble LLM framework that combines patient phenotyping through automated clinical information extraction with treatment decision support guided by RAG, based on EASL (European Association for the Study of the Liver) guidelines15 and the multiparametric therapeutic hierarchy concept.3 By integrating unstructured real-world clinical notes with current HCC management guidelines, PHENO-RAG performs sequential reasoning across two core tasks: the structured extraction of clinically relevant parameters to enable patient stratification, and the generation of guideline-informed recommendations for treatment allocation and multidisciplinary team (MDT) discussion. To minimize error propagation through these two sequential tasks, the extraction step was systematically optimized with validation against human-labelled data, and each component was independently evaluated to ensure that structured outputs meet high accuracy thresholds before being used for downstream decision-making. The framework was evaluated for its ability to predict therapeutic decisions, assess clinical complexity, and identify patients likely to benefit from MDT evaluation.

Patients and methods

Study population and data collection
We conducted a retrospective analysis of clinical reports from patients consecutively evaluated at University Hospital of Palermo (Italy), one of the hub centres of the Sicily Network for Liver Cancer,16 between September 1, 2020, and November 1, 2024. Clinical data were systematically collected through a dedicated web-based platform (https://www.sintesihepatology.org/study/PSN), which serves as the central database for the network. The platform facilitates structured data entry including: demographic information, clinical parameters related to chronic liver disease and HCC, laboratory values, radiological findings, treatment history and response assessment, multidisciplinary board discussion details and outcomes. All clinical reports were anonymized prior to analysis. A key feature of the platform is its capability to automatically generate standardized clinical reports in Italian language, available in both Word and PDF formats, which clinicians can subsequently modify as needed.

Clinical reports selection
For patients with multiple clinical reports during the study period, we selected reports corresponding to specific therapeutic decision points. For patients maintaining complete response throughout the recruitment period and requiring only clinical-radiological follow-up, we included only the latest available report. For all other patients, therapeutic decisions were categorized as: liver transplantation, surgical resection, ablation, transarterial treatment, systemic therapy, best supportive care (BSC).

Clinical decision criteria
Clinical decisions were based on multiple parameters3 including: age, Eastern Cooperative Oncology Group - performance status, Child-Pugh class, model for end-stage liver disease (MELD) score, presence of esophageal varices as an indicator of clinically significant portal hypertension, presence of severe extrahepatic comorbidities (adapted from comorbidities included in Charlson comorbidity index. Further details are reported in the supplementary materials), previous HCC treatments, response to previous HCC treatments, alpha-fetoprotein (AFP) levels, maximum diameter of the main nodule, number of nodules, vascular invasion and presence of extrahepatic disease.

Clinical complexity assessment and multidisciplinary discussion
Clinical complexity was retrospectively assessed by two clinicians (Ci.C, and G.C.) who reviewed all clinical reports and graded complexity as low, moderate or high. In cases of disagreement between the two physicians, a third senior physician (Ca.C.) acted as an adjudicator to resolve discrepancies. For each patient, we documented whether the case was discussed by the multidisciplinary board. The board comprises hepatologists, diagnostic and interventional radiologists, resection and transplant surgeons, and oncologists. The center’s standard criteria for determining which patients require multidisciplinary discussion are detailed in the supplementary materials.

PHENO-RAG framework
As illustrated in Fig. 1, the PHENO-RAG framework was designed to replicate the cognitive workflow of clinicians navigating complex HCC cases, by sequentially integrating clinical concept extraction, patient phenotyping, and evidence-based reasoning. Starting from unstructured data embedded in clinical and radiological reports, PHENO-RAG first employs LLMs and natural language processing techniques to extract key clinical variables and construct structured case profiles. Then management decisions are supported by relevant clinical guidelines with an integrated RAG system. This modular architecture enables the framework to support three essential decision-making tasks: assigning treatment recommendations, assessing clinical complexity, and identifying patients most likely to benefit from MDT discussion.

Patient phenotyping: Extraction of clinical concepts from medical reports
To identify the most effective method for clinical concept extraction, we conducted an ablation study comparing five different approaches: regular expression (REGEX)-based extraction alone, LLM-based extraction using LLMs separately, and combined REGEX-LLM pipelines for both model variants.
To validate our clinical concept extraction approach, we randomly selected 100 full medical reports from our database for manual labelling. Two experienced physicians independently reviewed these clinical notes to extract numerical values and determine the presence/absence of key clinical features (i.e., vascular invasion, severe comorbidities, metastatic disease, previous HCC treatment, complete response to last treatment, esophageal varices) or numerical values (i.e., AFP levels, serum albumin, performance status, Child-Pugh score, MELD score, and maximum nodule diameter). In cases of disagreement between the two physicians, a third senior physician acted as an adjudicator to resolve discrepancies. These manually labelled clinical notes served as the gold standard for evaluating our automated extraction system's performance.
REGEX is a standardized pattern-matching language used to identify and extract specific text patterns within documents. REGEX operates by defining search patterns that can match both exact text and more complex patterns like numbers, dates, or specific word combinations. For example, a simple pattern might look for the word "nodule" followed by a number and the unit "mm", while more complex patterns can account for variations in terminology, spacing, and surrounding context. To systematically extract relevant clinical information from the reports, we developed a comprehensive REGEX-based approach. For instance, to identify vascular invasion, the system would look for patterns like "portal[∖s]∗vein[∖s]∗thrombosis", where [∖s] represents optional spaces between words. Regular expressions were employed to identify and extract both categorical features (presence/absence) and numerical parameters from the clinical narratives. For categorical features (vascular invasion, severe comorbidities, metastatic disease, previous HCC treatment, complete response to last treatment, and esophageal varices), we created an extensive dictionary of relevant terms and their variations (detailed in the supplementary materials). The extraction algorithm incorporated common negation phrases (such as "no", "absent", "without") and evaluated their presence within a three-word window before and after the target keywords to accurately determine the presence or absence of each condition. For numerical parameters (AFP, serum albumin, performance status, Child-Pugh score, MELD score), the algorithm searched for numerical values immediately preceding or following predefined keyword markers. Special consideration was given to the extraction of nodule measurements, where the algorithm first identified the most recent imaging report based on the documented date (year), then searched for specific dimensional keywords within that imaging section to determine the maximum nodule diameter.
For the LLM-based extraction approach, we implemented a system using the following locally hosted, privacy-preserving language models: Meta’s Llama-3-8B and Llama-3-70B, OpenAI’s GPT-oss 20B and GPT-oss-120B, Qwen’s Qwen3-8B and Qwen3-Next-80B-A3B-Instruct, TII’s Falcon-7B and Falcon-40B. These specific models were selected based on their demonstrated strong performance in processing Italian language text as reported in their technical documentation, which was crucial as all clinical reports in our dataset were written in Italian.[17], [18], [19], [20], [21], [22], [23], [24] All prompts were formulated in Italian, and model outputs were generated and evaluated in the same language to ensure linguistic alignment and accuracy. For extraction, the models were configured with deterministic settings (temperature = 0, max tokens = 50) to ensure consistent outputs and minimize hallucination risks. Our prompt design (reported in the supplementary materials) was developed through an iterative optimization process, where multiple prompt variations were tested and evaluated for extraction accuracy. After several iterations and performance comparisons, we selected the best-performing prompt structure, which incorporated both chain-of-thought reasoning and few-shot learning principles. The final prompt structure began with a clear specification of the extraction task, listing all clinical parameters to be identified and their expected formats. This was followed by ten carefully selected clinical report-extraction pairs, representing diverse cases including reports with multiple findings, cases with negated conditions, various formats of numerical values, and complex scenarios requiring inference from context. Each example pair included both the input clinical report and its corresponding structured extraction output. The prompt also included explicit instructions for the model to identify relevant sections in the clinical report, extract specific values or conditions, validate extracted information against given criteria, and format the output in a consistent structure.
For the combined REGEX-LLM approach, we implemented a sequential processing pipeline where each clinical report was first analyzed by the LLM system, which attempted to extract all relevant parameters. For each parameter, the LLM would output either a specific value (numerical or categorical) or indicate "not reported" if the information could not be confidently extracted. In cases where the LLM indicated "not reported", the REGEX system was automatically triggered as a backup extraction method, providing an additional opportunity to identify and extract the missing information. This complementary approach leveraged the strengths of both methods, with the LLM's ability to understand context and complex relationships serving as the primary extraction tool, while the REGEX system acted as a safety net for cases where the LLM could not make a definitive determination.
We evaluated the performance of each extraction method against the human-labelled dataset using eight different LLMs as mentioned above. Each model was tested individually and in combination with REGEX. For clinical concepts – defined as categorical variables including vascular invasion, severe comorbidities, metastatic disease, previous HCC treatment, complete response to last treatment, and esophageal varices – we assessed the F1-score, reporting the median F1-score across all categorical variables. For numerical parameters – comprising AFP, serum albumin, performance status, Child-Pugh score, MELD score, and nodule maximum diameter – we evaluated the agreement using the intraclass correlation coefficient (ICC, two-way mixed effects), reporting the median ICC across all numerical variables. For numerical parameters, agreement with the human-labelled reference was defined as an exact match up to the first decimal place. Values differing beyond one decimal precision were classified as incorrect to ensure strict comparability with the reference annotations. For categorical variables, LLM outputs were evaluated using exact matching against human labels. Only binary responses (present/absent) matching the reference standard were considered correct, while 'not reported' or ambiguous outputs were classified as incorrect.

LLM-enhanced clinical decision support system
To evaluate the impact of transforming unstructured clinical notes into structured formats and to systematically assess the effect of different prompt design strategies and the integration of external domain knowledge, we implemented a second ablation study focused on clinical decision generation. This study aimed to isolate the contribution of input structure (unstructured vs. structured clinical notes), prompting approach (zero-shot vs. few-shot), and the addition of evidence retrieval through RAG. We tested eight different LLMs as mentioned above – across six distinct experimental configurations: zero-shot (unstructured note), zero-shot (structured note), few-shot (unstructured note), few-shot (structured note), few-shot+RAG (unstructured note), and few-shot+RAG (structured note) as depicted in Fig. 2. For each configuration, we evaluated model performance across three core clinical tasks: treatment allocation, clinical complexity assessment, and identification of cases requiring MDT discussion. This framework allowed us to quantify the individual and combined benefits of structured input, prompt exemplars, and guideline-informed retrieval in optimizing clinical recommendation quality.
We designed a systematic evaluation framework comparing different prompting strategies: zero-shot, few-shot, and few-shot with RAG. Each strategy was tested using both raw clinical notes and structured notes containing key elements extracted through our previously described pipeline. Structured notes were generated by employing the best performing framework for note extraction (REGEX-GPT-oss-120B). This approach allowed us to assess the impact of information structure on recommendation quality.
For the few-shot experiments, we randomly selected 10 clinical cases from our database to serve as examples. These cases were presented in note-outcome pairs, maintaining consistency with the input format (structured or unstructured) used in each experimental condition, in order to provide comprehensive context for the models compatible with the model context window.
For the RAG-enhanced approach, we incorporated two key reference documents: the EASL clinical practice guidelines published in 201815 and a previously published paper proposing an evidence-based framework for HCC treatment based on the concept of multiparametric therapeutic hierarchy.3 The whole document preprocessing followed our established protocols already reported in detail elsewhere12 and involved chunking the text at the paragraph level to create meaningful, self-contained units of information.25,26 We generated vector representations of these chunks using the all-MiniLM-L6-v2 embedding model, chosen for its robust performance in capturing medical domain semantics. During inference, our RAG system dynamically retrieved the five most relevant chunks from each reference document based on highest cosine similarity scores to the input clinical case. These retrieved chunks were then incorporated into the prompt structure to provide evidence-based context for the model's recommendations.
For each experimental condition (zero-shot, few-shot, and RAG), we developed specific prompt templates optimized for the task requirements. The zero-shot prompts contained clear task descriptions and expected output formats. Few-shot prompts included the ten example cases with their corresponding decisions, while RAG-enhanced prompts incorporated both examples and relevant retrieved passages from the reference documents.
The prompt structure for each condition was designed to maintain consistency while accommodating the specific requirements of each approach: zero-shot (direct instruction with clear specification of the three decision tasks); (2) few-shot (incorporation of the 10 example cases followed by the target case); RAG (integration of retrieved reference chunks alongside few-shot examples). This methodological framework enabled systematic evaluation of the models' capabilities across different levels of context and input structure, providing insights into the optimal approach for clinical recommendation generation.
All model interactions were configured with a temperature of 0.8 and max_tokens = 500. These hyperparameters were selected based on our prior experiments in clinical decision-support tasks, which identified this configuration as providing optimal text comprehension while minimizing the risk of hallucinations.26,27 The maximum output length was set to accommodate comprehensive recommendations while maintaining computational efficiency.
For experiments using structured notes, we utilized the output from our previous extraction pipeline, which provided standardized representations of key clinical parameters. For unstructured experiments, we used the raw clinical notes as direct input to the models, allowing them to process the natural language content in its original form.
For model evaluation in treatment allocation, clinical complexity assessment, and MDT discussion recommendation, we computed accuracy, defined as exact match rate between LLM recommendations and real-life recommendations across all three items.

Statistical analysis
Continuous variables were summarized as mean ± standard deviation, and categorical variables as frequencies and percentages. For clinical concept extraction evaluation, we used two complementary metrics: F1 score and ICC. The F1 score was calculated for categorical variables as the harmonic mean of precision (positive predictive value) and recall (sensitivity), ranging from 0 (worst performance) to 1 (perfect performance), with higher values indicating better agreement. The ICC was computed for numerical parameters using a two-way mixed-effects model for absolute agreement, with values interpreted as: <0.50 = poor, 0.50-0.75 = moderate, 0.75-0.90 = good, and >0.90 = excellent reliability. In the main text, aggregated summaries of model agreement are reported using median, first quartile (Q1), and third quartile (Q3) across all variables within each domain. Full item-level performance metrics for every variable and model configuration are provided in the supplementary materials to enable detailed examination of extraction performance.
For clinical decision support tasks (treatment allocation, clinical complexity, and MDT recommendation), performance was evaluated using accuracy, defined as the exact match rate between LLM-generated recommendations and real-world clinical decisions, expressed as a percentage.
Confidence intervals (95% CIs) for F1 scores and ICC values were calculated using bootstrap resampling with 1,000 iterations. For each iteration, we randomly sampled with replacement from the original dataset, computed the metric, and derived CIs from the 2.5th and 97.5th percentiles of the bootstrap distribution.

Computational environment
All analyses were performed in Python 3.11 using a Linux-based high-performance computing environment with GPU acceleration. LLMs were executed locally via the Hugging Face transformers and accelerate libraries, ensuring on-premise inference and data confidentiality. Model weights for Llama-3, GPT-oss, Qwen-3, and Falcon architectures were loaded from official Hugging Face repositories and run on NVIDIA A100 and H100 GPUs. Data preprocessing and evaluation were implemented with pandas (v2.2.3), numpy (v1.26.4), and scikit-learn (v1.5.1), while the retrieval component relied on FAISS-based vector indexes and embeddings generated with the all-MiniLM-L6-v2 model. This environment provided sufficient computational capacity to support inference and benchmarking of models ranging from 8B to 120B parameters.

Ethical considerations
This study was conducted in accordance with the Declaration of Helsinki and approved by Ethics Committee of University Hospital of Palermo (PSN 2017-Action 6.28). All clinical data were fully anonymized prior to analysis. Data confidentiality was maintained throughout the study, with access restricted to authorized researchers only.

Results
During study period, 489 clinical reports from 424 individual patients (mean age 71.2 ± 10.6 years, 73.3% male) were included. Clinical decisions included liver transplantation in 20 cases (4.1%), surgical resection in 41 (8.4%), ablation in 42 (8.6%), transarterial treatments in 49 (10.0%), systemic therapy in 151 (30.9%), BSC in 95 (19.4%) and follow-up continuation in 91 (18.6%).
Case complexity was judged as low in 113 cases (23.1%), moderate in 272 (55.6%) and high in 104 (21.3%). The distribution of clinical decisions according to case complexity is reported in Table 1.
Multidisciplinary discussion was undertaken in 63 cases (14.9%) (mean age 68.3 ± 11.9 years, 75.0% male). Among them, clinical decisions included liver transplantation in 10 cases (15.9%), surgical resection in 18 (28.6%), ablation in 12 (19.0%), transarterial treatments in 4 (6.3%), systemic therapy in 15 (23.8%), BSC in 2 (3.2%) and follow-up continuation in 2 (3.2%).

Accuracy analysis of the clinical concept extraction from clinical reports
The composition of 100 cases used for validating the clinical concept extraction approach is reported in the supplementary materials. Accuracy metrics for both categorical clinical concepts (evaluated with the F1 score) and continuous numerical parameters (evaluated with the ICC) are summarized in Table 2. Overall, LLMs outperformed the baseline REGEX approach across both domains. The best performance was achieved by GPT-oss-120B combined with REGEX, which reached a median F1 score of 0.92 (95% CI 0.85–0.95) for clinical concepts and an ICC of 0.93 (95% CI 0.85–0.94) for numerical parameters. Llama-3-70B+REGEX also demonstrated consistently strong results (F1 0.90, 95% CI 0.79–0.91; ICC 0.90, 95% CI 0.80–0.92), while Qwen-3-80B+REGEX achieved intermediate performance (F1 0.88, 95% CI 0.80–0.93; ICC 0.82, 95% CI 0.70–0.83). In contrast, Falcon-based models showed lower accuracy overall, although their performance improved modestly with REGEX integration. Model-specific performance metrics for extraction of individual clinical concepts and numerical parameters is reported in Tables S2 and S3, respectively. GPT-oss-120B combined with REGEX achieved the highest overall performance, with the best extraction accuracy for previous HCC treatment (F1 = 0.98) and AFP levels (ICC = 0.99), whereas its lowest scores were observed for metastatic disease (F1 = 0.80) and nodule diameter (ICC = 0.78). Per-parameter extraction performance of GPT-oss-120B+REGEX is reported in Fig. 3.

Agreement analysis of treatment allocation
For treatment allocation, GPT-oss-120B achieved the highest performance across all experimental conditions, with exact match rates ranging from 28.8% (zero-shot on unstructured notes) to 86.5% (few-shot+RAG on structured notes). Llama-3-70B demonstrated strong performance, with exact match rates increasing from 25.8% for zero-shot on unstructured notes to 80.8% for few-shot+RAG on structured notes. In contrast, Falcon-7B showed the poorest performance, with accuracy ranging from only 12.5% to 42.3% across all experimental conditions. All models demonstrated consistent improvement with structured data input, few-shot examples, and RAG integration (Table 3).
The confusion matrix showing the distribution of treatment assignments by GPT-oss-120B with few-shot+RAG on structured notes, including misclassification patterns across all treatment categories, is reported in Fig. S1. In the few-shot setting without RAG, the model correctly classified 229 patients, with several consistent misclassification patterns. The most frequent confusion involved ablation and transarterial treatments, with 23 total bidirectional errors. Ablation remained one of the lowest-performing categories, with 18/42 correct classifications (42.9%), while systemic therapy – despite being the most prevalent category – achieved 78/151 correct classifications (51.7%). BSC performed similarly, with 45/95 correct classifications (47.4%), and follow-up continuation reached 38/91 (41.8%). The addition of RAG substantially improved model performance, increasing overall correct classifications to 425 patients – a net gain of +196 compared with few-shot alone. The most substantial improvements occurred in categories with previously low accuracy. Systemic therapy improved from 78 to 130 correct predictions (+52), ablation from 18 to 39 (+21), and transarterial treatments from 24 to 44 (+20). Follow-up continuation increased from 38 to 74 (+36), and surgical resection from 19 to 37 (+18).
Fig. 4 depicts the treatment allocation accuracy across prompting strategies for GPT-oss-120B, showing that the integration of RAG consistently improved classification accuracy across all treatment categories, with the largest gains observed for systemic therapy, ablation, and transarterial therapies.

Agreement analysis of clinical complexity
Similar patterns emerged for clinical complexity assessment. GPT-oss-120B again achieved the highest accuracy rates, ranging from 37.0% (zero-shot on unstructured notes) to 88.6% (few-shot+RAG on structured notes). Llama-3-70B showed comparable performance, with exact match rates improving from 32.9% to 83.4% across the experimental conditions. Falcon-7B consistently performed worst, with accuracy rates between 17.8% and 46.8%. The progressive enhancement from basic zero-shot prompting to the full few-shot+RAG approach with structured data was evident across all models (Table 3).
The confusion matrix showing clinical complexity assessment by GPT-oss-120B with few-shot+RAG on structured notes is depicted in Fig. S2. In the few-shot configuration, GPT-oss-120B exhibited a high rate of bidirectional misclassification, with frequent overestimation and underestimation of clinical complexity. Overestimation was the dominant pattern (110 errors), driven mainly by incorrect upgrading of low-complexity cases into the moderate or high categories and misclassification of moderate cases as high complexity. Underestimation (90 errors) was also common and largely reflected the tendency to downgrade moderate cases to low complexity and high-complexity cases to moderate or low complexity. The introduction of retrieval (few-shot+RAG) substantially altered the model’s error profile. Overall misclassification dropped from 200 to 56 cases, and – critically – the directionality of errors became balanced, with underestimation and overestimation occurring at identical frequencies (28 vs. 28 cases). Underestimation primarily involved moderate cases predicted as low complexity and high cases predicted as moderate, whereas overestimation stemmed mostly from isolated upgrades of low-complexity notes into the moderate or high category.

Agreement analysis of decisions for multidisciplinary discussion
MDT recommendation proved the most challenging task for all models. GPT-oss-120B achieved the best performance with exact match rates ranging from 18.2% (zero-shot on unstructured notes) to 66.9% (few-shot+RAG on structured notes). Llama-3-70B demonstrated similar trends, improving from 16.2% to 63.0% across experimental conditions. Falcon-7B showed the lowest accuracy, with rates between 8.0% and 33.3%. Even the best-performing models struggled with this task compared to treatment allocation and clinical complexity assessment (Table 2).

Discussion
HCC management presents unique challenges due to the complex interplay between tumor characteristics, underlying chronic liver disease, and patient-specific factors. The integration of LLMs into clinical decision support systems could represent a promising frontier in healthcare delivery, particularly for complex diseases like HCC, where management decisions are multiparametric, and encompass multiple factors.
In this retrospective study of 489 real-world HCC clinical notes, we developed PHENO-RAG, an AI-assisted system designed to extract clinical information from narrative reports and provide clinical recommendations via LLMs, supported by international guidelines. Across eight models tested with different prompting/knowledge conditions, the accuracy of clinical decisions dramatically improved when the models were provided with structured clinical information and access to current management guidelines. The best overall results were obtained with GPT-oss-120B, followed by Llama-3-70B, while other tested models showed lower overall performance, but still benefitted from the same methodological enhancements. These findings support the potential for LLM-assisted tools to enhance consistency and quality of HCC management decisions in clinical practice.
The first step of our approach was to assess the ability of our framework to extract key determinants of clinical decisions from unstructured clinical notes, thus obtaining structured data for subsequent clinical decision-making. This automated structuring process represents a critical advance, as it automatically transforms raw narrative clinical reports into systematically organized information that can be more reliably processed and analyzed. A key finding of our study is represented by the significant performance disparities across different models for the accurate identification of key clinical parameters influencing clinical decisions, such as tumor characteristics, liver function parameters, and response to previous HCC treatments. Combined approaches with GPT-oss-120B plus REGEX (a standardized text-searching method for identifying and extracting specific medical terms from clinical text) achieved superior clinical concept extraction, with accuracy higher than 90%, substantially outperforming smaller models. This aligns with recent evaluations showing that combined approaches are more accurate in medical applications.28,29 The integration of REGEX with LLM processing proved critical, aligning with previous findings observed in other hepatological settings,12 where guideline reformatting and structured presentation of clinical information led to similar improvements in accuracy. The performance gap between GPT-oss-120B and smaller models was substantial across all conditions, highlighting the importance of model selection in clinical applications.30 Our extraction performance is consistent with recent systematic evidence showing that natural language processing systems achieve sensitivity and specificity higher than 90% in extracting information from free-text medical reports.31
We then assessed the performance of our framework in generating clinical recommendations, that were compared with real-world specialist decisions across three clinical tasks: HCC treatment allocation, grading of clinical complexity and identification of need for MDT discussion. Our framework demonstrated high concordance with real-world decisions for treatment allocation and clinical complexity assessment, suggesting potential to support clinical workflows while maintaining quality of care. Interestingly, our framework demonstrated progressive improvement across all models with structured data preprocessing, access to representative clinical examples (few-shot learning), and guideline-informed recommendations through RAG integration.
The magnitude of improvement – from 25-30% baseline accuracy to 80-88% with full PHENO-RAG implementation – underscores the critical importance of the use of pre-processed structured and the integration of data domain-specific knowledge, as already shown in other medical AI applications.10,32 From a practical perspective, the structured data produced by this automated process enabled more accurate clinical decision-making compared to processing unstructured notes directly. This finding has important implications for clinical practice, as it suggests that automated systematic organization of clinical information may improve consistency and quality of treatment decisions. The ability to automatically convert narrative clinical reports into structured formats could therefore serve as a valuable tool for standardizing clinical workflows and supporting evidence-based decision-making. At the same time, incorporating current clinical guidelines directly into the decision-making process provided substantial performance gains. This approach, known as RAG, allows the system to automatically consult relevant sections of established treatment guidelines – similar to how clinicians reference current recommendations when making decisions. This ensures that treatment suggestions remain consistent with the latest evidence-based practices and can be updated as guidelines evolve. Our findings support evidence that LLM-RAG models can achieve high accuracy levels in specific clinical tasks when properly designed.33
Error analysis demonstrated that RAG systematically corrected the model's weakest performance areas: treatment categories with the lowest baseline accuracy (ablation, systemic therapy, follow-up) improved to 81-93% with guideline retrieval, while complexity assessment errors decreased from 200 to 56 cases with balanced directionality. This targeted improvement pattern suggests that guideline integration addresses specific knowledge gaps and systematic biases, strengthening confidence in the framework's clinical aim.
Among the three clinical tasks, MDT recommendation proved most challenging, with GPT-oss-120B achieving only 66.9% accuracy. This could reflect that current guidelines recommend universal or near-universal MDT discussion, while real-world practice requires selective prioritization due to resource constraints, creating discrepancy between guideline-based and actual decisions. Moreover, MDT referral incorporates subjective judgment and institutional factors that could not be captured in clinical notes.
Our findings have important implications for the practical implementation of our framework as a tool for clinical decision support in HCC. First, they demonstrate the feasibility of using LLMs for complex clinical decision-making when properly configured with structured data (collected from a large sample of patients managed in a real-world setting at a tertiary referral center during a relatively short period) and domain-specific knowledge. Second, they highlight the importance of maintaining human oversight, particularly for complex cases requiring multidisciplinary input. Third, they suggest a potential role for LLMs in streamlining clinical workflows by identifying cases that would benefit most from MDT discussion, according to their clinical complexity.
Several limitations of our study should be acknowledged. First, while our dataset of 489 clinical reports represents a substantial sample, it comes from a single institution and may not capture the full variety of clinical presentations and decision-making patterns across different healthcare settings. Second, the retrospective nature of our study means that the model's recommendations require prospective validation. Third, we did not evaluate the predictive performance of the framework across different treatment modalities stratified by clinical complexity. Fourth, while our framework incorporated radiological information extracted from imaging reports embedded in clinical notes, it did not perform direct analysis of radiological images themselves, which are known to be fundamental determinants in therapeutic decision-making for HCC. The absence of this information could affect the ability of the model to fully capture the nuances that influence treatment allocation in clinical practice, particularly regarding the location of the tumor. Integrating direct radiological images in future iterations would potentially enhance the framework's accuracy and clinical applicability.
Future research directions should include prospective validation of the framework in clinical practice, evaluation across multiple institutions and healthcare systems, and investigation of how the framework performs with other complex diseases requiring multidisciplinary management. Additionally, further work is needed to improve the model's performance in borderline cases and to develop more robust methods for clinical complexity stratification.
Several ethical and practical considerations warrant discussion. First, while PHENO-RAG demonstrates promising accuracy, clinical liability remains with treating physicians. Second, transparent documentation of AI involvement in decision-making processes is essential for maintaining accountability and patient trust. Third, data confidentiality must be rigorously protected, particularly when implementing such frameworks in clinical practice; our use of locally hosted models rather than cloud-based application programming interfaces represents one approach to mitigating privacy risks.
In conclusion, our evaluation demonstrates the potential of multi-model LLM frameworks to support complex clinical decision-making in HCC management when properly implemented with structured data preprocessing, domain-specific knowledge retrieval, and systematic prompt optimization. While these tools show promise in streamlining clinical workflows and supporting evidence-based practice, they should be viewed as augmentative technologies that enhance rather than replace clinical judgment and multidisciplinary expertise. Future research should focus on prospective validation, multimodal integration, and robust evaluation frameworks for complex clinical reasoning tasks.

Abbreviations
AFP, alpha-fetoprotein; BSC, best supportive care; HCC, hepatocellular carcinoma; ICC, intraclass correlation coefficient; LLM, large language model; MELD, model for end-stage liver disease; MDT, multidisciplinary team; REGEX, regular expression; RAG, retrieval-augmented generation.

Authors’ contributions
Ci.C., and M.G. were involved in the conception of the study, the analysis and interpretation of the results and wrote the first draft of the manuscript. M.G., G.D.M, and S.K. were involved in formal analysis. G.C. and Ca.C.. were involved in the interpretation of the results and contributed to the discussion. R.C., M.S., A.G., S.R., G.G., A.Q., G.C., A.S., and V.G were involved in data collection. Calogero Cammà/Ca. C. is the guarantor. All authors edited, reviewed, and approved the final version of the manuscript.

Data availability
All supporting data are available within the article (and in the supplementary materials).

Code availability
Code can be provided based on personal requests. Please contact the corresponding author.

Financial support
The research leading to these results has received funding from the 10.13039/501100000780European Union - NextGenerationEU through the Italian Ministry of University and Research under PNRR M4C2I1.3 project PE_00000019 “HEAL ITALIA” CUPB73C22001250006 to Ca. C. M.G. is supported by the 10.13039/100005410American-Italian Cancer Foundation Post-Doctoral Research Fellowship (Year 2024/2025).

Conflict of interest
The authors declare no conflicts of interest pertaining to this manuscript.
Please refer to the accompanying ICMJE disclosure forms for further details.