The organization of information about genes, genetic variants, and associated diseases from scientific literature texts can facilitate progress in precision medicine. However, the vast scale of this literature demands development of automated strategies for identifying and extracting this information. Generative large language models (LLMs) represent a promising avenue for such automation. We systematically evaluate the performance of LLMs on the extraction of information relating to impacts of genetic variation on disease from the biomedical literature, considering the challenge of genetic and disease named entity recognition in Spanishlanguage scientific abstracts, and experimenting with a range of instruction strategies over a dataset known as GenoVarDis. We evaluate cross-linguistic prompting, and zero- and few-shot strategies, along with the optional provision of an annotation guideline and variations in the requested output format. A key finding is that the natural language of the prompt had only a limited impact on the model's performance on NER, demonstrating the feasibility of cross-linguistic information extraction. Overall, optimal results were obtained with few-shot prompting. However, we identify that generative LLMs failed to adhere to the instructions provided, leading to the over-generation and fabrication (hallucinations) of entities not appearing in the texts. We find that adding examples to the prompts and providing an overview of the expected output structure reduces hallucinated entities. Lastly, we explore the limitations of the prompting strategies and demonstrate the value of grounding generated outputs in the original texts. Overall, LLMs do not reach the accuracy of task-specific models, but we gain insight into effective strategies for their use.

‘Silent’ evaluation refers to the prospective, non-interventional testing of artificial intelligence (AI) model performance in the intended clinical setting without affecting patient care or institutional operations. The silent evaluation phase has received less attention than in silico algorithm development or formal clinical evaluations, despite increasing recognition of this type of evaluation as a critical phase in an effective translation process for healthcare AI tools. There are currently no formal guidelines for conducting silent AI evaluations in health settings. We undertook a scoping review to identify silent AI evaluations described in the literature, aiming to summarize current practices for the conduct of silent evaluations. We screened PubMed, Web of Science, and Scopus databases for articles fitting our criteria for silent AI evaluations, or ‘silent trials’, published from 2015 to 2025. A total of 891 articles were identified, and 75 met the criteria for inclusion into the final review. We found wide variance in terminology, description, and rationale for silent evaluations; this led to substantial heterogeneity in what was reported. Overwhelmingly, papers reported measurement of AUC, precision/recall, positive and negative predictive values and similar technical performance metrics. Far fewer studies reported the verification of outputs against an in-situ clinical ground truth, and, when reported, the comprehensiveness of such verification was highly variable. We noted relatively less discussion of sociotechnical components such as stakeholder engagement and human-computer interaction elements. We conclude that there is an opportunity to bring together diverse evaluative practices (e.g., from data science, human factors, and other fields) if the silent evaluation phase is to be maximally effective as a translational mechanism these gaps mirror challenges in effective translation of AI tools from “computer to bedside” and identify opportunities to improve silent evaluation protocols that address key translational needs. This is important as healthcare organizations and regulatory bodies worldwide seek guidance for gathering meaningful evidence of the impact of AI tools on clinical practice.

Graph transformers are a recent advancement in machine learning, offering a new class of neural network models for graph-structured data. The synergy between transformers and graph learning demonstrates strong performance and versatility across various graph-related tasks. This survey provides an in-depth review of recent progress and challenges in graph transformer research. We begin with foundational concepts of graphs and transformers. We then explore design perspectives of graph transformers, focusing on how they integrate graph inductive biases and graph attention mechanisms into the transformer architecture. Furthermore, we propose a taxonomy classifying graph transformers based on depth, scalability, and pre-training strategies, summarizing key principles for effective development of graph transformer models. Beyond technical analysis, we discuss the applications of graph transformer models for node-level, edge-level, and graph-level tasks, exploring their potential in other application scenarios as well. Finally, we identify remaining challenges in the field, such as scalability and efficiency, generalization and robustness, interpretability and explainability, dynamic and complex graphs, as well as data quality and diversity, charting future directions for graph transformer research.

OBJECTIVE:Standardizing terminology offers opportunities for improved communication and care outcomes. With increasing adoption of clinical terminologies, questions remain about whether they adequately capture the scope of musculoskeletal (MSK) primary care practice. This scoping review examines global development efforts on MSK-relevant standardized terminology and its implementation in clinical practice. METHODS:A scoping review was conducted of 6 databases to May 2025. Identified studies (n = 3668) were included (n = 60) if they addressed standardized terminology relevant to the MSK primary care professions of chiropractic, osteopathy, and physiotherapy. Data were extracted on use cases, documentation of MSK information, alignment with national interoperability standards, and implementation status. RESULTS:Global development efforts span diverse MSK domains across condition types. Five studies achieved consensus around domain-specific terms (including tendinopathies, groin pain, and weight-bearing rehabilitation); in contrast, many studies developed extensive clinical terminology sets. Most studies (82.4%) address the development of terminologies, with few yet addressing how they have been implemented into clinical practice (2.7%). Analysis revealed MSK clinicians require documentation beyond existing core interoperability data groups, including 1) function and movement, 2) pain characteristics, 3) psychosocial factors, 4) social determinants of health (environmental factors and participation barriers), 5) intervention effectiveness and clinical outcomes, and 6) person-centered factors. Multiple barriers emerged, including technical (EHR integration, cognitive burden), workflow (time requirements, clinical value), professional (training, profession-specific terminology), and knowledge gaps (impact on care quality). CONCLUSION:Extensive terminology development has begun yet gaps exist between development and clinical adoption. Terms evolve as research evolves; therefore, MSK professions should actively engage with interoperability groups to establish hierarchical ontologies that incorporate the identified data groups and balance standardization at higher conceptual levels with flexible lexicons to enable terminology growth over time. Establishing feedback mechanisms with EHR vendors to minimize clinicians' cognitive burden will accelerate adoption and maximize clinical value.

Benchmarking the capabilities of AI systems, including Large Language Models (LLMs) and Vision Models, typically ignores the impact of uncertainty in the underlying ground truth answers from experts. This ambiguity is not just limited to human preferences, but is also consequential even in safety critical domains such as medicine where uncertainty is pervasive. In this paper, we introduce a probabilistic paradigm to theoretically explain how high certainty in ground truth answers is almost always necessary for even an expert to achieve high scores, whereas in datasets with high variation in ground truth answers there may be little difference between a random labeller and an expert. This characteristic also manifests when comparing models, where uncertainty obfuscates differences between poor and high performing models. Therefore, ignoring uncertainty in ground truth evaluation data can result in the misleading conclusion that a non-expert has similar performance to that of an expert. Using the probabilistic paradigm, we thus bring forth the concepts of expected accuracy and expected F1 to estimate the score an expert human or system can achieve given ground truth answer variability. Our work leads to the recommendation that when establishing the capability of a system, results should be stratified by probability of the ground truth answer, typically measured by the agreement rate of ground truth experts. Stratification becomes critical when the overall performance drops below a threshold of 80

Epilepsy is a chronic neurological disorder requiring multi-faceted management, including seizure detection, syndrome diagnosis, prognostication, antiseizure medication recommendation, epileptogenic zone localization, and surgical outcome prediction. Although numerous deep learning approaches have been developed for individual tasks, these models are typically siloed and modality-specific (e.g., EEG for seizure detection, MRI for localization), failing to reflect the multidisciplinary nature of real-world epilepsy care, where epileptologists, neuroradiologists, neurosurgeons, neuropsychologists and neuropsychiatrists jointly interpret heterogeneous evidence to guide decisions. In this work, we propose a clinical guideline-grounded hybrid multi-agent framework for holistic epilepsy management. Heterogeneous patient data is processed through modality-specific discriminative and generative models, where textual interpretations from generative agents are combined with structured predictions from discriminative models as auxiliary guidance. This aggregated evidence is passed to a central orchestrating agent grounded in international epilepsy guidelines, which evaluates multi-modal findings within structured clinical pathways and performs iterative cross-agent coordination for evidence-informed decision-making. We evaluate our framework across two datasets spanning six epilepsy management tasks and also introduce a publicly available multi-modal, multi-task epilepsy benchmark. Results demonstrate that integrating discriminative evidence with guideline-grounded generative coordination yields more reliable and comprehensive decisions compared to conventional LLM-based and task-specific baselines. Our dataset and code is available at https://github.com/khoapham154/epi_guide.git}. ### Competing Interest Statement The authors have declared no competing interest. ### Funding Statement This study did not receive any funding ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: This study was approved by the Alfred Hospital Ethics Committee (Ethics number 437/21). The Ethics Committee is constituted according to NHMRC guidelines and reports to the Alfred Health Executive Committee, which in turn reports to the Alfred Health Board. I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes The data using for training model can be found at: https://github.com/khoapham154/epi_guide.git

A common strategy in transfer learning is few shot fine-tuning, but its success is highly dependent on the quality of samples selected as training examples. Active learning methods such as uncertainty sampling and diversity sampling can select useful samples. However, under extremely low-resource and class-imbalanced conditions, they often favor outliers rather than truly informative samples, resulting in degraded performance. In this paper, we introduce RADS (Reinforcement Adaptive Domain Sampling), a robust sample selection strategy using reinforcement learning (RL) to identify the most informative samples. Experimental evaluations on several real world clinical datasets show our sample selection strategy enhances model transferability while maintaining robust performance under extreme class imbalance compared to traditional methods.

Optical Chemical Structure Recognition (OCSR) aims to convert two-dimensional molecular images into machine-readable formats such as SMILES strings. Deep learning has substantially improved OCSR performance, yet most methods rely on synthetic training data and struggle to generalize to real-world inputs, especially hand-drawn diagrams, where stroke width, geometry, and drawing conventions vary widely across individuals. In this work, we propose an image-to-graph model AdaptMol that enables effective transfer from synthetic to real-world data without requiring manual graph annotations in the target domains. AdaptMol is an integrated pipeline that starts with training a base model on synthetic data, and then refines model representations through unsupervised domain adaptation and self-training. Our key insight is that bond features are domain-invariant in nature; they encode structural relationships between atoms that are independent of visual variations across domains. Thus, during domain adaptation, we align bond-level feature distributions via class-conditional Maximum Mean Discrepancy (MMD) to enforce cross-domain consistency. We also design a comprehensive data augmentation strategy to enhance the robustness of the base model, facilitating stable self-training on unlabeled target samples. On hand-drawn molecular images, our model achieves 82.6% accuracy and outperforms the best prior method by 10.7 points, while maintaining competitive performance across four benchmarks comprising molecular images from scientific literature and patent documents.Scientific contributionWe propose AdaptMol, an image-to-graph model that predicts molecular structures as graphs of atoms and bonds, achieving effective transfer from synthetic to hand-drawn molecular images without requiring target domain graph annotations. We combine class-conditional Maximum Mean Discrepancy to align bond features across domains with comprehensive data augmentation to increase training data variation, jointly improving base model accuracy sufficiently for self-training and addressing the critical failure mode of prior approaches that begin with insufficient accuracy. We further introduce a dual position representation that supervises atom positions through both discrete coordinate tokens and continuous spatial heatmaps to reduce false positives in atom localization.

Pregnancy complications are a leading cause of maternal and neonatal mortality worldwide. Understanding their underlying mechanisms is hindered by dispersed evidence across thousands of studies and complex biological interactions. We present PregBase, a comprehensive knowledge base for pregnancy research comprising automated extraction, validation, and exploration. PregBase was constructed by utilising large language models (LLMs) to extract relationships from literature; 13 LLMs were benchmarked across seven prompting strategies, with ensemble shuffle prompting outperforming single-strategy alternatives. A three-tier validation pipeline combining ontology mapping, graph neural networks, and statistical analysis produced PregKG, a knowledge graph containing 64,087 associations between 13,303 biomedical entities spanning 155 semantic types and 50 vocabularies across 8 relationship types from 8420 articles. Link prediction validated PregBase’s inference capability beyond extracted knowledge, recovering established clinical interventions, reconstructing canonical hormonal pathways across maternal-placental-fetal compartments, and identifying novel biomarker candidates for preterm birth, gestational diabetes, and preeclampsia, supported by genetic and expression databases. An interactive web interface (https://pregknowledgebase.com) has been created to enable further exploration via conversational queries. This work provides a scalable foundation for systematic discovery in maternal health research.

Data quality (DQ) and transparency of secondary data are critical factors that delay the adoption of clinical AI models and affect clinician trust in them. Many DQ studies fail to clarify where, along the lifecycle, quality checks occur, leading to uncertainty about provenance and fitness for reuse. This study develops a framework for transparent reporting of DQ assessments across the clinical electronic health record (EHR) data lifecycle. The reporting framework was developed through iterative analysis to identify actors and phases of the clinical data lifecycle. The framework distinguishes between data-generating organisations and data-receiving organisations to allow users to map DQ parameters to stages across the data lifecycle. The framework defines five key lifecycle phases and multiple actors. When applied to the real-world dataset, the framework demonstrated applicability in revealing where DQ issues may originate. The framework provides a structured approach for reporting DQ assessments, which can enhance transparency regarding data fitness for reuse, supporting reliable clinical research, AI model development, and internal organisational governance. This work provides practical guidance for researchers to understand data provenance and for organisations to target DQ improvement efforts across the data lifecycle.

Large Language Models (LLMs) are increasingly being explored for clinical question answering and decision support, yet safe deployment critically requires reliable handling of patient measurements in heterogeneous clinical notes. Existing evaluations of LLMs for clinical numerical reasoning provide limited operation-level coverage, restricted primarily to arithmetic computation, and rarely assess the robustness of numerical understanding across clinical note formats. We introduce ClinicNumRobBench, a benchmark of 1,624 context-question instances with ground-truth answers that evaluates four main types of clinical numeracy: value retrieval, arithmetic computation, relational comparison, and aggregation. To stress-test robustness, ClinicNumRobBench presents longitudinal MIMIC-IV vital-sign records in three semantically equivalent representations, including a real-world note-style variant derived from the Open Patients dataset, and instantiates queries using 42 question templates. Experiments on 14 LLMs show that value retrieval is generally strong, with most models exceeding 85

Background Electronic health records (EHRs) collect real-world clinical data that can objectively measure clinical outcomes and deliver personalized learning opportunities for clinicians when analyzed. Current safety principles and the learning health system framework recognize that proactive analysis of routine practice data, rather than only monitoring adverse events, is fundamental to the continuous improvement of patient and clinician outcomes. The aim was to assess musculoskeletal clinicians’ use of an EHR and identify areas of potential clinical improvement by comparing data entry variability with clinical outcomes. Methods A retrospective observational study was conducted on EHR data from an Australian occupational healthcare organization employing chiropractors, physiotherapists, and osteopaths from 2014–2021. Data were restricted to food manufacturing industries to control for patient population variability, and clinicians with fewer than 100 complaints were excluded. The primary outcome was treatments to release (TTR). Variables were selected based on a literature review of predictors of work-related musculoskeletal disorder outcomes. Results Data from 46,539 complaints and 52 clinicians were analyzed. Meaningful variation was identified across all variables examined, consistent with the presence of learning opportunities within routine clinical data. Mean TTR ranged from 1.02 to 5.16 across clinicians, with body region and mechanism of injury associated with variation in recovery length. Obstacles to recovery documentation varied widely across clinicians, from 1.9% to 85.4%, while psychosocial factors ranged from 0% to 20.4%. Prognosis accuracy was consistent across all experience levels, with clinicians consistently overestimating recovery length, offering insights into opportunities for clinician development. Conclusions Clinical EHR data contains meaningful information that supports the identification of learning opportunities for clinicians, to inform feedback discussions on clinician skills, biases, and EHR training. Clinical quality measurement as part of routine clinical governance is a significant safety improvement for patients, who deserve care from clinicians who have access to objective EHR data, not solely self-assessed competence.

Concept-based explainable models have emerged as a promising approach to enhance transparency and trust in skin cancer diagnosis by linking image features to clinical concepts. However, when concepts are imperfectly defined or correlated with confounding patterns in the data, the model may inadvertently learn spurious associations that compromise its generalizability, robustness, and interpretability. Existing approaches rarely address how to systematically unlearn such spurious concepts once they are identified. Without such correction, concept-based models may continue to rely on visual artifacts rather than pathological cues, undermining trust and limiting their reliability. We introduce an explainability-oriented unlearning strategy that targets spurious concepts within the model’s interpretability layer. Through counterfactual training, our method refines the representations within the concept space, reducing the influence of confounding patterns while preserving clinically meaningful concepts. Furthermore, we introduce a hybrid visualization framework that integrates concept importance scores with spatial saliency maps, providing intuitive explanations for model predictions. These contributions advance the development of robust and interpretable concept-based models and have the potential to improve trust and transparency in automated skin cancer diagnostics.

Artificial intelligence (AI) is a multidisciplinary scientific field that uses machines to solve real-world problems and predict outcomes. Despite the current enthusiasm about AI's potential as a clinical support tool, there is also a growing awareness and concern about the potentially harmful effects of AI. Because AI will likely impact expert-based decision-making in medicine, it is critical to consider the issues that AI raises in medical research. This paper outlines the AI guidelines of the Australian Epilepsy Project. This large-scale platform aims to democratise specialist care in epilepsy and use AI for clinical decision support based on prospective multimodal datasets (MRI, genetic, clinical, and cognitive data) from thousands of people with epilepsy. As AI develops rapidly, we focus on key areas of medical AI identified in the literature, including Trust, Responsibility and Safety. We believe AI is changing medicine, and we believe it is imperative to advance and update our AI guidelines adaptably while preparing for an era of augmented-intelligence-based medicine.

‘Silent’ evaluation refers to the prospective, non-interventional testing of artificial intelligence (AI) model performance in the intended clinical setting without affecting patient care or institutional operations. The silent evaluation phase has received less attention than in silico algorithm development or formal clinical evaluations, despite increasing recognition of this type of evaluation as a critical phase in an effective translation process for healthcare AI tools. There are currently no formal guidelines for conducting silent AI evaluations in health settings. We undertook a scoping review to identify silent AI evaluations described in the literature, aiming to summarize current practices for the conduct of silent evaluations. We screened PubMed, Web of Science, and Scopus databases for articles fitting our criteria for silent AI evaluations, or ‘silent trials’, published from 2015 to 2025. A total of 891 articles were identified, and 75 met the criteria for inclusion into the final review. We found wide variance in terminology, description, and rationale for silent evaluations; this led to substantial heterogeneity in what was reported. Overwhelmingly, papers reported measurement of AUC, precision/recall, positive and negative predictive values and similar technical performance metrics. Far fewer studies reported the verification of outputs against an in-situ clinical ground truth, and, when reported, the comprehensiveness of such verification was highly variable. We noted relatively less discussion of sociotechnical components such as stakeholder engagement and human-computer interaction elements. We conclude that there is an opportunity to bring together diverse evaluative practices (e.g., from data science, human factors, and other fields) if the silent evaluation phase is to be maximally effective as a translational mechanism these gaps mirror challenges in effective translation of AI tools from “computer to bedside” and identify opportunities to improve silent evaluation protocols that address key translational needs. This is important as healthcare organizations and regulatory bodies worldwide seek guidance for gathering meaningful evidence of the impact of AI tools on clinical practice.

Despite rapid adoption of autoregressive large language models, smaller text encoders still play an important role in text understanding tasks that require rich contextualized representations. Negation is an important semantic function that is still not properly captured by such methods, affecting many downstream applications relying on text embeddings. We propose a strategy to improve negation robustness of text encoders, by distilling data from large language models using diverse patterns of negation and hedging. We adopt a standard contrastive learning strategy to finetune a strong BERT-based model, and observe large improvement in negation understanding capabilities while maintaining competitive performance on general benchmarks. In addition, we also show that our method can be adapted to LLMs, leading to improved performance on negation benchmarks.

Background Adequate self-harm surveillance is a key part of suicide prevention efforts. Our prior work has demonstrated the efficacy of an artificial intelligence model for detecting self-harm in emergency department triage notes. This model was developed based on data from a single hospital, raising the question about the model’s robustness to different contexts. Here, we aim to validate the model prospectively and externally to understand its portability across hospital settings. Methods Our self-harm classification model was developed and tested using triage notes from a large metropolitan hospital in Melbourne, Australia from 2012 to 2017. The model combined extensive text pre-processing with a Gradient Boosting classifier that used 644 selected features. In this study, we assessed the portability of both model components. We performed prospective validation using 329,655 triage notes from the same hospital collected over the following four years. For external validation, we used 316,877 triage notes from 2012 to 2021 from a regional hospital located 150km outside Melbourne. Results On the initial test set, the model achieved an area under the precision-recall curve (PR AUC) of 0.86, positive predictive value (PPV) of 0.81, and sensitivity of 0.80. Prospectively, the performance remained stable with PR AUC of 0.84, PPV of 0.76, and sensitivity of 0.76. Externally, the model showed a diminished ability to discern self-harm cases with an overall classification metric PR AUC of 0.77, PPV of 0.57, and sensitivity of 0.83. The text normalisation component of the model was equally effective across the datasets. Conclusions At the metropolitan hospital, the self-harm detection model is sufficiently performant for both epidemiological and potential clinical uses. At the regional hospital, the text normalisation pipeline is effective, but the machine learning classifier may need to be re-trained locally to produce more accurate results. ### Competing Interest Statement The authors have declared no competing interest. ### Funding Statement KW is supported by a National Health and Medical Reseacrh Council (NHMRC) Emerging Leaders 1 Investigator Grant (1177787). JR is supported by a NHMRC Investigator Grant (2008460) and the University of Melbourne Dame Kate Campbell Fellowship. ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: Melbourne Health Human Research Ethics Committee (HREC; 2017.342) gave ethical approval for this work. I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes All data produced in the present study are available upon reasonable request to the authors

Invasive fungal infections (IFIs) pose significant risks to patients with weakened immune systems, requiring timely detection. To improve IFI detection from clinical reports, we explore the value of recent advances in NLP techniques for this task, including transformer-based pre-trained language models (PLMs) and generative large language models (LLMs). Experimental results show these methods are more effective for IFI detection than prior approaches, with a hybrid approach missing only one positive case over a public benchmark dataset, CHIFIR. These findings highlight the value of modern NLP methods, and the utility of combining diverse approaches.

Clinical predictive analysis is a crucial task with numerous applications and has been extensively studied using machine learning approaches. Clinical notes, a vital data source, have been employed to develop natural language processing (NLP) models for risk prediction in healthcare with robust performance. However, clinical notes vary considerably in text composition-written by diverse healthcare providers for different purposes-and the impact of these variations on NLP modeling is also underexplored. It also remains uncertain whether the recent Large Language Models (LLMs) with instruction-following capabilities can effectively handle the risk prediction task out-of-the-box, especially when using routinely collected clinical notes instead of polished text. We address these two important research questions in the context of in-hospital mortality prediction within the critical care setting. Specifically, we propose a supervised hierarchical network with note-specific modules to account for variations across different note categories, and provide a detailed comparison with strong supervised baselines and LLMs. We benchmark 34 instruction-following LLMs based on zero-shot, few-shot, and chain-of-thought prompting with diverse prompt templates. Our results demonstrate that the note-specific network delivers improved risk prediction performance compared to established supervised baselines from both measurement-based and text-based modeling. In contrast, LLMs consistently underperform on this critical task, despite their remarkable performances in other domains. This highlights important limitations and raises caution regarding the use of LLMs for risk assessment in the critical setting. Additionally, we show that the proposed model can be leveraged to select informative clinical notes to enhance the training of other models.