Accurately communicating the side effects of cancer treatments to cancer survivors is critical, particularly in settings such as informed consent, where clinicians must clearly and comprehensively convey potential treatment toxicities. However, this task remains challenging due to clinical knowledge deficits about adverse treatment effects and fragmentation across electronic health record (EHR) systems. Large language models (LLMs) have the potential to assist in this task, though their reliability in oncology survivorship contexts remains poorly understood. We present a deployment-oriented stress-testing framework for evaluating LLM-generated radiation side effect lists in breast cancer treatment and survivorship care. Using 21 breast cancer patient profiles, we construct paired patient clinical scenarios that differ only in radiotherapy regimens to evaluate seven instruction-tuned LLMs under multiple prompting regimes. We then compare LLM outputs to a clinician-curated reference derived from informed consent documents at two major academic medical centers and developed by a team including more than seven breast radiation oncologists. The reference maps radiation dose-fractionation, fields, and locations to associated toxicities, broken down by frequency and temporal onset. Across models, we reveal sensitivity to minor documentation changes, trade-offs between precision and recall, and systematic under-recall of rare and long-term side effects. When used alone, constraints on the number of side effects generated reduce precision, and grounding outputs in clinician-curated side effect lists substantially improves reliability and robustness. These findings highlight important limitations of LLM use in oncology and suggest practical design choices for safer and more informative survivorship-focused applications.

Large language models (LLMs) have been rapidly adopted in healthcare, but oversight strategies are lacking. We propose capability-based monitoring, motivated by the fact that LLMs are generalist systems whose overlapping internal capabilities are reused across numerous downstream tasks. This approach organizes monitoring around shared capabilities to enable cross-task detection of systemic weaknesses, long-tail errors, and emergent behaviors. We describe considerations for developers, organizational leaders, and professional societies. and policymakers.

Large language models are increasingly being used to assess and forecast research ideas, yet we lack scalable ways to evaluate the quality of models' judgments about these scientific ideas. Towards this goal, we introduce PoT, a semi-verifiable benchmarking framework that links scientific idea judgments to downstream signals that become observable later (e.g., citations and shifts in researchers' agendas). PoT freezes a pre-cutoff snapshot of evidence in an offline sandbox and asks models to forecast post-cutoff outcomes, enabling verifiable evaluation when ground truth arrives, scalable benchmarking without exhaustive expert annotation, and analysis of human-model misalignment against signals such as peer-review awards. In addition, PoT provides a controlled testbed for agent-based research judgments that evaluate scientific ideas, comparing tool-using agents to non-agent baselines under prompt ablations and budget scaling. Across 30,000+ instances spanning four benchmark domains, we find that, compared with non-agent baselines, higher interaction budgets generally improve agent performance, while the benefit of tool use is strongly task-dependent. By combining time-partitioned, future-verifiable targets with an offline sandbox for tool use, PoT supports scalable evaluation of agents on future-facing scientific idea judgment tasks.

Characterizing a large language model's (LLM's) knowledge of a given question is challenging. As a result, prior work has primarily examined LLM behavior under knowledge conflicts, where the model's internal parametric memory contradicts information in the external context. However, this does not fully reflect how well the model knows the answer to the question. In this paper, we first introduce a taxonomy of five knowledge statuses based on the consistency and correctness of LLM knowledge modes. We then propose KScope, a hierarchical framework of statistical tests that progressively refines hypotheses about knowledge modes and characterizes LLM knowledge into one of these five statuses. We apply KScope to nine LLMs across four datasets and systematically establish: (1) Supporting context narrows knowledge gaps across models. (2) Context features related to difficulty, relevance, and familiarity drive successful knowledge updates. (3) LLMs exhibit similar feature preferences when partially correct or conflicted, but diverge sharply when consistently wrong. (4) Context summarization constrained by our feature analysis, together with enhanced credibility, further improves update effectiveness and generalizes across LLMs.

Current ophthalmology clinical workflows are plagued by over-referrals, long waits, and complex and heterogeneous medical records. Large language models (LLMs) present a promising solution to automate various procedures such as triaging, preliminary tests like visual acuity assessment, and report summaries. However, LLMs have demonstrated significantly varied performance across different languages in natural language question-answering tasks, potentially exacerbating healthcare disparities in Low and Middle-Income Countries (LMICs). This study introduces the first multilingual ophthalmological question-answering benchmark with manually curated questions parallel across languages, allowing for direct cross-lingual comparisons. Our evaluation of 6 popular LLMs across 7 different languages reveals substantial bias across different languages, highlighting risks for clinical deployment of LLMs in LMICs. Existing debiasing methods such as Translation Chain-of-Thought or Retrieval-augmented generation (RAG) by themselves fall short of closing this performance gap, often failing to improve performance across all languages and lacking specificity for the medical domain. To address this issue, We propose CLARA (Cross-Lingual Reflective Agentic system), a novel inference time de-biasing method leveraging retrieval augmented generation and self-verification. Our approach not only improves performance across all languages but also significantly reduces the multilingual bias gap, facilitating equitable LLM application across the globe.

Large language models (LLMs) exhibit a vulnerability arising from being trained to be helpful: a tendency to comply with illogical requests that would generate false information, even when they have the knowledge to identify the request as illogical. This study investigated this vulnerability in the medical domain, evaluating five frontier LLMs using prompts that misrepresent equivalent drug relationships. We tested baseline sycophancy, the impact of prompts allowing rejection and emphasizing factual recall, and the effects of fine-tuning on a dataset of illogical requests, including out-of-distribution generalization. Results showed high initial compliance (up to 100%) across all models, prioritizing helpfulness over logical consistency. Prompt engineering and fine-tuning improved performance, improving rejection rates on illogical requests while maintaining general benchmark performance. This demonstrates that prioritizing logical consistency through targeted training and prompting is crucial for mitigating the risk of generating false medical information and ensuring the safe deployment of LLMs in healthcare.

Objectives: The application of natural language processing (NLP) in the clinical domain is important due to the rich unstructured information in clinical documents, which often remains inaccessible in structured data. When applying NLP methods to a certain domain, the role of benchmark datasets is crucial as benchmark datasets not only guide the selection of best-performing models but also enable the assessment of the reliability of the generated outputs. Despite the recent availability of language models capable of longer context, benchmark datasets targeting long clinical document classification tasks are absent. Materials and Methods: To address this issue, we propose Long Clinical Document (LCD) benchmark, a benchmark for the task of predicting 30-day out-of-hospital mortality using discharge notes of Medical Information Mart for Intensive Care IV and statewide death data. We evaluated this benchmark dataset using baseline models, from bag-of-words and convolutional neural network to instruction-tuned large language models. Additionally, we provide a comprehensive analysis of the model outputs, including manual review and visualization of model weights, to offer insights into their predictive capabilities and limitations. Results: Baseline models showed 28.9% for best-performing supervised models and 32.2% for GPT-4 in F1 metrics. Notes in our dataset have a median word count of 1687. Discussion: Our analysis of the model outputs showed that our dataset is challenging for both models and human experts, but the models can find meaningful signals from the text. Conclusion: We expect our LCD benchmark to be a resource for the development of advanced supervised models, or prompting methods, tailored for clinical text.

BACKGROUND:Adequate patient awareness and understanding of cancer clinical trials is essential for trial recruitment, informed decision making, and protocol adherence. Although large language models (LLMs) have shown promise for patient education, their role in enhancing patient awareness of clinical trials remains unexplored. This study explored the performance and risks of LLMs in generating trial-specific educational content for potential participants. METHODS:Generative Pretrained Transformer 4 (GPT4) was prompted to generate short clinical trial summaries and multiple-choice question-answer pairs from informed consent forms from ClinicalTrials.gov. Zero-shot learning was used for summaries, using a direct summarization, sequential extraction, and summarization approach. One-shot learning was used for question-answer pairs development. We evaluated performance through patient surveys of summary effectiveness and crowdsourced annotation of question-answer pair accuracy, using held-out cancer trial informed consent forms not used in prompt development. RESULTS:For summaries, both prompting approaches achieved comparable results for readability and core content. Patients found summaries to be understandable and to improve clinical trial comprehension and interest in learning more about trials. The generated multiple-choice questions achieved high accuracy and agreement with crowdsourced annotators. For both summaries and multiple-choice questions, GPT4 was most likely to include inaccurate information when prompted to provide information that was not adequately described in the informed consent forms. CONCLUSIONS:LLMs such as GPT4 show promise in generating patient-friendly educational content for clinical trials with minimal trial-specific engineering. The findings serve as a proof of concept for the role of LLMs in improving patient education and engagement in clinical trials, as well as the need for ongoing human oversight.

The integration of large language models (LLMs) into electronic health records offers potential benefits but raises significant ethical, legal, and operational concerns, including unconsented data use, lack of governance, and AI-related malpractice accountability. Sycophancy, feedback loop bias, and data reuse risk amplifying errors without proper oversight. To safeguard patients, especially the vulnerable, clinicians must advocate for patient-centered education, ethical practices, and robust oversight to prevent harm.

Large language models (LLMs) are rapidly being adopted in healthcare, necessitating standardized reporting guidelines. We present transparent reporting of a multivariable model for individual prognosis or diagnosis (TRIPOD)-LLM, an extension of the TRIPOD + artificial intelligence statement, addressing the unique challenges of LLMs in biomedical applications. TRIPOD-LLM provides a comprehensive checklist of 19 main items and 50 subitems, covering key aspects from title to discussion. The guidelines introduce a modular format accommodating various LLM research designs and tasks, with 14 main items and 32 subitems applicable across all categories. Developed through an expedited Delphi process and expert consensus, TRIPOD-LLM emphasizes transparency, human oversight and task-specific performance reporting. We also introduce an interactive website ( https://tripod-llm.vercel.app/ ) facilitating easy guideline completion and PDF generation for submission. As a living document, TRIPOD-LLM will evolve with the field, aiming to enhance the quality, reproducibility and clinical applicability of LLM research in healthcare through comprehensive reporting. TRIPOD-LLM (transparent reporting of a multivariable model for individual prognosis or diagnosis–large language model) is a checklist of items considered essential for good reporting of studies that are developing or evaluating an LLM for use in healthcare settings. It is a ‘living guideline’ that emphasizes transparency, human oversight and task-specific performance reporting.

Objective To evaluate large language models (LLMs) for pre-test diagnostic probability estimation and compare their uncertainty estimation performance with a traditional machine learning classifier.Materials and Methods We assessed 2 instruction-tuned LLMs, Mistral-7B-Instruct and Llama3-70B-chat-hf, on predicting binary outcomes for Sepsis, Arrhythmia, and Congestive Heart Failure (CHF) using electronic health record (EHR) data from 660 patients. Three uncertainty estimation methods-Verbalized Confidence, Token Logits, and LLM Embedding+XGB-were compared against an eXtreme Gradient Boosting (XGB) classifier trained on raw EHR data. Performance metrics included AUROC and Pearson correlation between predicted probabilities.Results The XGB classifier outperformed the LLM-based methods across all tasks. LLM Embedding+XGB showed the closest performance to the XGB baseline, while Verbalized Confidence and Token Logits underperformed.Discussion These findings, consistent across multiple models and demographic groups, highlight the limitations of current LLMs in providing reliable pre-test probability estimations and underscore the need for improved calibration and bias mitigation strategies. Future work should explore hybrid approaches that integrate LLMs with numerical reasoning modules and calibrated embeddings to enhance diagnostic accuracy and ensure fairer predictions across diverse populations.Conclusions LLMs demonstrate potential but currently fall short in estimating diagnostic probabilities compared to traditional machine learning classifiers trained on structured EHR data. Further improvements are needed for reliable clinical use. This study explored the ability of large language models (LLMs) to predict the likelihood of 3 medical conditions-Sepsis, Arrhythmia, and Congestive Heart Failure-before diagnostic tests are conducted. The models were tested using information from 660 patients' electronic health records (EHRs) and compared to a traditional machine learning method called eXtreme Gradient Boosting (XGB). The research also investigated how well these models estimated uncertainty in their predictions using 3 different techniques. The results showed that the traditional XGB method was more accurate than the LLM-based approaches for all 3 conditions. Among the LLM methods, one technique that combined LLMs with XGB came closest to matching the XGB's performance, but other LLM-based methods were less effective. Our findings highlight that while LLMs have potential for medical prediction, they currently lag behind traditional approaches and need further development to become reliable tools in clinical settings.

Introduction:Patients receiving thoracic radiotherapy (RT) have an increased risk of major adverse cardiac events (MACE) posttreatment. We utilized machine learning (ML) to discover novel predictors of MACE and validated them on an external cohort. Methods:This multi-institutional retrospective study included 984 patients [n = 803 non-small cell lung cancer (NSCLC), n = 181 breast cancer] treated with radiotherapy. Extreme gradient boosting was utilized to discover novel clinical, dosimetric, and anatomical features (CT-based cardiac substructure segmentations) associated with MACE in a cohort of locally advanced NSCLC patients. Fine-Gray regression was performed with non-cardiac death as a competing risk. External validation was performed utilizing independent cohorts of NSCLC or breast cancer patients. Results:In the discovery dataset (n = 701), 70 patients experienced MACE. ML modeling (training AUC, 0.68; testing AUC, 0.71) identified right and left atrial volume indices (RAVI and LAVI, respectively) as top predictors. After adjusting for baseline cardiovascular risk and known radiotherapy predictive factors, RAVI was associated with an increased risk of MACE [subdistribution hazard ratio (sHR) 1.02/unit, 95% confidence interval (CI): 1.00-1.04; p = 0.03]. In the validation cohorts (n = 102 NSCLC; n = 181 breast cancer), RAVI was associated with an increased risk of MACE (NSCLC: sHR 1.05, 95% CI: 1.001-1.106, p = 0.04; breast cancer: sHR 1.06, 95% CI: 1.01-1.11, p = 0.03). Similar findings were found for LAVI. Discussion:ML modeling identified right and left atrial enlargement as novel radiographic predictors for increased risk of MACE following chest radiotherapy, which was validated in independent breast and lung cancer datasets. Given that echocardiography studies have demonstrated the prognostic utility of atrial volume indices across cardiovascular risk groups, these findings warrant further study to identify additional strategies for upfront cardiovascular risk profiling.

Generative artificial intelligence has brought disruptive innovations in health care but faces certain challenges. Retrieval-augmented generation (RAG) enables models to generate more reliable content by leveraging the retrieval of external knowledge. In this perspective, we analyze the possible contributions that RAG could bring to health care in equity, reliability, and personalization. Additionally, we discuss the current limitations and challenges of implementing RAG in medical scenarios.

PURPOSE To evaluate the performance and consistency of large language models (LLMs) across brand and generic oncology drug names in various clinical tasks, addressing concerns about potential fluctuations in LLM performance because of subtle phrasing differences that could affect patient care. METHODS This study evaluated three LLMs (GPT-3.5-turbo-0125, GPT-4-turbo, and GPT-4o) using drug names from HemOnc ontology. The assessment included 367 generic-to-brand and 2,516 brand-to-generic pairs, 1,000 drug-drug interaction (DDI) synthetic patient cases, and 2,438 immune-related adverse event (irAE) cases. LLMs were tested on drug name recognition, word association, DDI (DDI) detection, and irAE diagnosis using both brand and generic drug names. RESULTS LLMs demonstrated high accuracy in matching brand and generic names (GPT-4o: 97.38% for brand, 94.71% for generic, P < .01). However, they showed significant inconsistencies in word association tasks. GPT-3.5-turbo-0125 exhibited biases favoring brand names for effectiveness (odds ratio [OR], 1.43, P < .05) and being side-effect-free (OR, 1.76, P < .05). DDI detection accuracy was poor across all models (<26%), with no significant differences between brand and generic names. Sentiment analysis revealed significant differences, particularly in GPT-3.5-turbo-0125 (brand mean 0.67, generic mean 0.95, P < .01). Consistency in irAE diagnosis varied across models. CONCLUSION Despite high proficiency in name-matching, LLMs exhibit inconsistencies when processing brand versus generic drug names in more complex tasks. These findings highlight the need for increased awareness, improved robustness assessment methods, and the development of more consistent systems for handling nomenclature variations in clinical applications of LLMs.

Large language models (LLMs) are increasingly envisioned as decision-support tools in clinical practice, yet safe clinical reasoning demands integrating heterogeneous knowledge bases – trials, primary studies, regulatory documents, and cost data – under strict accuracy constraints. Existing evaluations often rely on synthetic prompts, reduce the task to single-hop factoid queries, or conflate reasoning with open-ended generation, leaving their real-world utility unclear. To close this gap, we present MedBrowseComp, the first benchmark that systematically tests an agent's ability to reliably retrieve and synthesize multi-hop medical facts from live, domain-specific knowledge bases. MedBrowseComp contains more than 1,000 human-curated questions that mirror clinical scenarios where practitioners must reconcile fragmented or conflicting information to reach an up-to-date conclusion. Applying MedBrowseComp to frontier agentic systems reveals performance shortfalls as low as ten percent, exposing a critical gap between current LLM capabilities and the rigor demanded in clinical settings. MedBrowseComp therefore offers a clear testbed for reliable medical information seeking and sets concrete goals for future model and toolchain upgrades. You can visit our project page at: https://moreirap12.github.io/mbc-browse-app/