Intelligent Head and Neck CTA Report Quality Detection with Large Language Models.

PurposeLarge language models (LLMs) like Generative Pre-trained Transformer 4 (GPT-4) can assist in detecting errors in radiology reports, but privacy concerns limit their clinical applicability. This study compares closed-source and privacy-compliant open-source LLMs for detecting common errors in radiology reports.Materials and methodsA total of 120 radiology reports were compiled (30 each from X-ray, ultrasound, CT, and MRI). Subsequently, 397 errors from five categories (typographical, numerical, findings-impression discrepancies, omission/insertion, interpretation) were inserted into 100 of these reports; 20 reports were left unchanged. Two open-source models (Llama 3-70b, Mixtral 8x22b) and two commercial closed-source (GPT-4, GPT-4o) were tasked with error detection using identical prompts. The Kruskall–Wallis test and paired t-test were used for statistical analysis.ResultsOpen-source LLMs required less processing time per radiology report than closed-source LLMs (6 ± 2 s vs. 13 ± 4 s; p < 0.001). Closed-source LLMs achieved higher error detection rates than open-source LLMs (GPT-4o: 88% [348/397; 95% CI: 86, 92], GPT-4: 83% [328/397; 95% CI: 80, 87], Llama 3-70b: 79% [311/397; 95% CI: 76, 83], Mixtral 8x22b: 73% [288/397; 95% CI: 68, 77]; p < 0.001). Numerical errors (88% [67/76; 95% CI: 82, 93]) were detected significantly more often than typographical errors (75% [65/86; 95% CI: 68, 82]; p = 0.02), discrepancies between findings and impression (73% [73/101; 95% CI: 67, 80]; p < 0.01), and interpretation errors (70% [50/71; 95% CI: 62, 78]; p = 0.001).ConclusionOpen-source LLMs demonstrated effective error detection, albeit with comparatively lower accuracy than commercial closed-source models, and have potential for clinical applications when deployed via privacy-compliant local hosting solutions.Key PointsQuestionCan privacy-compliant open-source large language models (LLMs) match the error-detection performance of commercial non-privacy-compliant closed-source models in radiology reports?FindingsClosed-source LLMs achieved slightly higher accuracy in detecting radiology report errors than open-source models, with Llama 3-70b yielding the best results among the open-source models.Clinical relevanceOpen-source LLMs offer a privacy-compliant alternative for automated error detection in radiology reports, improving clinical workflow efficiency while ensuring patient data confidentiality. Further refinement could enhance their accuracy, contributing to better diagnosis and patient care.Graphical

This study aims to evaluate whether large language models (LLMs) can accurately predict the urgency and severity of radiology reports. Based on the recommendations of the Academy of Royal Colleges, we defined radiology reports that include unexpected findings of high urgency or severity as "high-priority (HP) radiology reports." Overall, 1906 radiology reports were used as the training set, and 176 radiology reports were used as the test set, with a balanced ratio of HP to non-HP radiology reports (1:1) in both sets. Four types of LLMs (Llama2 7B, Llama3 8B, Llama3 Elyza 8B, and Llama 3.1 8B) were fine-tuned using four different input settings: (1) findings only, (2) findings + referring department, (3) findings + referring department + clinical diagnosis before examination, and (4) findings + referring department + clinical diagnosis before examination + details of examination request. The fine-tuned LLMs predicted whether each radiology report was HP or not. Among the four LLMs, Llama3 Elyza 8B, with inputs comprising findings and the referring department, demonstrated the best performance, achieving PRAUC = 0.962, ROCAUC = 0.968, accuracy = 0.915, sensitivity/recall = 0.932, specificity = 0.898, and F1 = 0.916. Adding a clinical diagnosis before the examination and details of examination requests did not necessarily lead to performance improvement. The fine-tuned LLMs accurately predicted HP radiology reports, suggesting their potential utility in supporting communication regarding radiology reports with high urgency or severity. Question This study aims to evaluate whether large language models (LLMs) can accurately predict the high-priority (HP) radiology reports. Findings The fine-tuned best LLM accurately HP radiology reports, achieving PRAUC of 0.962 and ROCAUC of 0.968. Clinical relevance This study demonstrates that fine-tuned LLMs can accurately identify HP radiology reports, potentially improving timely clinical decision-making and enhancing patient safety through faster communication of critical findings.

While large language models (LLMs) have shown promise in medical text analysis, their application in automated medical billing code extraction remains underexplored, particularly for the German medical fee schedule system (GOÄ). Therefore, an LLM was fine-tuned to perform multi-label classification of GOÄ codes from radiology reports automatically, and its performance was compared with state-of-the-art commercial and open-source LLMs. Following ethics committee approval, we analyzed 499,601 radiology reports from 124,497 patients, containing 1,799,971 manually identified GOÄ codes as ground truth. The MediPhi-Instruct 4B model was fine-tuned using five-fold cross-validation. Performance was evaluated on the hold-out test set and compared against GPT-5, GPT-4.1, GPT-oss, Kimi-K2, Deepseek-R1, Deepseek-V3, Gemini 2.5, Llama-70B, and Qwen-3 LLMs on a subset of 500 anonymized and 350 cleaned reports using zero-shot and few-shot prompting techniques. The fine-tuned model achieved an accuracy of 77.15% ± 0.47% and a micro-average F1-score of 87.79% ± 0.31% on the hold-out test set. On a subset of 500 real-world samples, our models outperformed the best-performing LLM, Gemini 2.5 Flash, with an F1-score of 70.32% ± 1.54% compared to 58.22% ± 1.50% (p < 0.001). For the cleaned dataset of 350 samples, GPT-5 achieved the best F1-score of 89.51 ± 1.52% and outperformed the fine-tuned models (p < 0.001). Fine-tuned LLMs can effectively automate GOÄ code classification from radiology reports, with the potential of outperforming commercial LLMs. This approach shows promise for improving billing efficiency and accuracy in healthcare settings, though manual verification is still recommended. Question LLMs with high parameters possess medical knowledge, but how effective are they at predicting billing codes from radiology reports compared to smaller, fine-tuned models? Finidngs A fine-tuned ensemble model achieved competitive results and can outperform larger, proprietary LLMs. Clinical relevance Smaller, fine-tuned models offer an efficient alternative to proprietary LLMs in generating billing codes and can be integrated to assist clinical coding. This technology has the potential to transform clinical billing procedures, but its use should be overseen by qualified professional personnel.

Patient safety research in human medicine has identified the causes and common types of medical error and subsequently informed the development of interventions which mitigate harm, such as the WHO's...

Background The complex medical terminology of radiology reports may cause confusion or anxiety for patients, especially given increased access to electronic health records. Large language models (LLMs) can potentially simplify radiology report readability. Purpose To compare the performance of four publicly available LLMs (ChatGPT-3.5 and ChatGPT-4, Bard [now known as Gemini], and Bing) in producing simplified radiology report impressions. Materials and Methods In this retrospective comparative analysis of the four LLMs (accessed July 23 to July 26, 2023), the Medical Information Mart for Intensive Care (MIMIC)-IV database was used to gather 750 anonymized radiology report impressions covering a range of imaging modalities (MRI, CT, US, radiography, mammography) and anatomic regions. Three distinct prompts were employed to assess the LLMs' ability to simplify report impressions. The first prompt (prompt 1) was "Simplify this radiology report." The second prompt (prompt 2) was "I am a patient. Simplify this radiology report." The last prompt (prompt 3) was "Simplify this radiology report at the 7th grade level." Each prompt was followed by the radiology report impression and was queried once. The primary outcome was simplification as assessed by readability score. Readability was assessed using the average of four established readability indexes. The nonparametric Wilcoxon signed-rank test was applied to compare reading grade levels across LLM output. Results All four LLMs simplified radiology report impressions across all prompts tested (P < .001). Within prompts, differences were found between LLMs. Providing the context of being a patient or requesting simplification at the seventh-grade level reduced the reading grade level of output for all models and prompts (except prompt 1 to prompt 2 for ChatGPT-4) (P < .001). Conclusion Although the success of each LLM varied depending on the specific prompt wording, all four models simplified radiology report impressions across all modalities and prompts tested. © RSNA, 2024 Supplemental material is available for this article. See also the editorial by Rahsepar in this issue.

Picture description tasks, such as the Cookie Theft task, are widely used in neuropsychological assessments to detect cognitive impairment. However, manual scoring is time-consuming, requires specialized training, and is subject to interrater variability. Recent advancements in natural language processing, particularly large language models (LLMs), offer a promising solution to automate and standardize the scoring process. This study evaluated the performance and reliability of five LLMs (GPT-4 Turbo, GPT-4o, Claude 3 Opus, Claude 3 Sonnet, and Llama 3 70b) in scoring the Cookie Theft picture description task. A subset of 25 participants were selected from the DementiaBank corpus. The LLMs were tasked with scoring 22 content units in the participants' responses using various prompt strategies, including few-shot learning, prompt chaining, and self-consistency. LLM performance was compared to the consensus score of three human raters. LLMs demonstrated comparable accuracy to human raters in scoring the Cookie Theft task, with no significant differences in mean absolute error (MAE) between the best performing models and human raters. Few-shot learning significantly improved LLM performance, while prompt chaining and self-consistency showed limited benefits. Claude 3 Opus and GPT-4o exhibited the highest accuracy and reliability. Notably, LLMs showed significantly higher interrater reliability compared to human raters. The findings demonstrate the potential of LLMs to accurately and reliably score picture description tasks, offering a promising approach to streamline and standardize neuropsychological assessments. By automating the scoring process, clinicians and researchers can benefit from increased efficiency, reduced subjectivity, and improved scalability in evaluating cognitive functions.

Background/Objectives. Abdominal computed tomography (CT) radiology reports contain critical information for abdominal aortic aneurysm (AAA) management, including aneurysm presence, size, rupture status, and prior repair. However, this information is often embedded within lengthy, heterogeneous reports, making manual extraction inefficient. We evaluated the performance of multiple large language models (LLMs) for automated extraction of AAA-related findings from radiology reports. Methods. We retrospectively analyzed 500 abdominal CT reports mentioning AAA from an urban academic health system (2020-2024). Ground truth labels were established by manual review. Four open-source LLMs (Qwen2.5-7B-Instruct, Llama3-Med42-8B, GPT-OSS-20B, and MedGemma-27B-text-it) were evaluated for extraction of aneurysm presence, size, morphology, rupture status, impending rupture, and prior aortic repair. Model outputs were compared with ground truth using exact-match accuracy, and inter-model agreement was assessed using Fleiss' kappa. Reasoning traces were examined to characterize correct and incorrect model behavior. Results. Accuracy for identifying AAA presence ranged from 0.90 to 0.95 (κ = 0.851), and prior aortic repair from 0.90 to 0.97 (κ = 0.793). Accuracy for aneurysm size ranged from 0.67 to 0.88 (κ = 0.340), with low κ's due to class imbalance or dimension misselection. Rupture and impending rupture were identified with accuracies exceeding 0.90 across models, though agreement was lower (κ = 0.485 and 0.589), reflecting low event prevalence. Larger models (GPT-OSS-20B, MedGemma-27B) generally outperformed smaller models. Reasoning analysis revealed strengths in measurement prioritization but recurrent errors, including dimension misselection, over-inference of prior repair, and conservative classification of rupture-related findings. Conclusions. LLMs can accurately extract clinically relevant AAA information from radiology reports with interpretable reasoning, with larger and medically trained models outperforming smaller or general-purpose models. Performance varies by task and model, underscoring the need for careful validation and human-in-the-loop deployment in clinical settings.

Evaluating gpt-4 for zero-shot classification of bleeding and clotting events: Can large language models serve as second reviewers?

Importance: Recent breakthroughs in Large Language models (LLM) have shown impressive reasoning capabilities. Inferring disease trajectory from radiology reports is a central task in clinical oncology, and radiographic interpretation of disease state and treatment benefits in pancreas cancer can be particularly challenging. We assess the performance of LLMs (GPT3.5-turbo and GPT4) to infer disease status, response to treatment, and location of disease for patients with pancreatic adenocarcinoma from clinical radiology reports. We also assess the ability of LLMs to extrapolate from objective findings, vs. their reliance on radiologist’s summaries. Experimental Design: We used 200 deidentified radiology reports from pancreatic cancer patients at different stages of treatment and submitted these reports to the GPT3.5 and GPT4 models via an Microsoft Azure HIPAA compliant instance. We assess the reasoning capabilities of these two models in the three independent tasks in pancreatic cancer imaging. 1) disease status/response to treatment, 2) disease location, and 3) presence of indeterminate nodules requiring further follow up. Evaluation of response was compared to a medical oncologist’s interpretation of each report. A second medical oncologist served as adjudicator for discrepancies between human and GPT models and rates of data fabrication. We also assessed importance of Radiologists summaries vs Objective finding sections, as well as different prompt engineering approaches. Results: We found a maximum performance of 71% and 85% of GPT3.5 and GPT4 respectively when asked to classify radiology reports into seven categories of malignancy presence and trajectory, as compared to a medical oncologist’s interpretation. This accuracy was strongly dependent on requesting the model demonstrate its reasoning process before providing an answer; though this dependency was abrogated by providing the Radiologist’s impression section. We also found the main source of mis-classification was incongruence between reasoning and the final classification (as opposed to incorrect medical reasoning). Finally, we showed that GPT4 performs strongly on identifying specific organ sites of disease, as well as presence of indeterminate findings requiring further follow up. Conclusion: GPT algorithms accurately interpret multiple clinically relevant features in radiology reports for Pancreatic adenocarcinoma patients. Precision improved markedly with GPT 4 compared to GPT-3.5. This suggests GPT-based tools are potential valuable clinical allies in interpreting radiologist-produced text and could be valuable aids in cohort identification for large scale real-world data research and eventually clinical management support. Citation Format: Travis Ian Zack, Madhumita Sushil, Brenda Miao, Arda Demirci, Corynn Kasap, Margaret Tempero, Atul Butte, Eric Collisson. Clinical inference of location and trajectory of pancreatic cancer from radiology reports using zero-shot LLM [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Pancreatic Cancer; 2023 Sep 27-30; Boston, Massachusetts. Philadelphia (PA): AACR; Cancer Res 2024;84(2 Suppl):Abstract nr B074.

Validating Radiology Artificial Intelligence Model Performance on Photon-Counting CT Images Using Large Language Models for Ground Truth Extraction.

High-quality discharge summaries are associated with improved patient outcomes, but contribute to clinical documentation burden. Large language models (LLMs) provide an opportunity to support physicians by drafting discharge summary narratives. To determine whether LLM-generated discharge summary narratives are of comparable quality and safety to those of physicians. This cross-sectional study conducted at the University of California, San Francisco included 100 randomly selected inpatient hospital medicine encounters of 3 to 6 days' duration between 2019 and 2022. The analysis took place in July 2024. A blinded evaluation of physician- and LLM-generated narratives was performed in duplicate by 22 attending physician reviewers. Narratives were reviewed for overall quality, reviewer preference, comprehensiveness, concision, coherence, and 3 error types (inaccuracies, omissions, and hallucinations). Each error individually, and each narrative overall, were assigned potential harmfulness scores ranging from 0 to 7 on an adapted Agency for Healthcare Research and Quality scale. Across 100 encounters, LLM- and physician-generated narratives were comparable in overall quality on a Likert scale ranging from 1 to 5 (higher scores indicate higher quality; mean [SD] score, 3.67 [0.49] vs 3.77 [0.57]; P = .21) and reviewer preference (χ2 = 5.2; P = .27). LLM-generated narratives were more concise (mean [SD] score, 4.01 [0.37] vs 3.70 [0.59]; P < .001) and more coherent (mean [SD] score, 4.16 [0.39] vs 4.01 [0.53]; P = .02) than their physician-generated counterparts, but less comprehensive (mean [SD] score, 3.72 [0.58] vs 4.13 [0.58]; P < .001). LLM-generated narratives contained more unique errors (mean [SD] errors per summary, 2.91 [2.54]) than physician-generated narratives (mean [SD] errors per summary, 1.82 [1.94]). There was no significant difference in the potential for harm between LLM- and physician-generated narratives across individual errors (mean [SD] of 1.35 [1.07] vs 1.34 [1.05]; P = .99), with 6 and 5 individual errors, respectively, with scores of 4 (potential for permanent harm) or greater. Both LLM- and physician-generated narratives had low overall potential for harm (scores <1 on a scale ranging from 0-7), with LLM-generated narratives scoring higher than physician narratives (mean [SD] score of 0.84 [0.98] vs 0.36 [0.70]; P < .001) and only 1 LLM-generated narrative (compared with 0 physician-generated narratives) scoring 4 or greater. In this cross-sectional study of 100 inpatient hospital medicine encounters, LLM-generated discharge summary narratives were of comparable quality, and were preferred equally, to those generated by physicians. LLM-generated narratives were more likely to contain errors but had low overall harmfulness scores. These results suggest that, in clinical practice, using such narratives after human review may provide a viable option for hospitalists.

Large Language Models (LLMs) hold the potential for clinical task-shifting by processing unstructured clinical text, enabling tasks such as clinical concept extraction and medical question answering from electronic health records. If implemented reliably, such approaches may benefit over-burdened healthcare systems, particularly in resource-limited settings and for traditionally overlooked populations, provided that local fine-tuning is supported by appropriate clinical and technical expertise. However, this powerful technology remains largely understudied in real-world contexts, particularly in the Global South. This study aims to assess whether openly available LLMs can be used reliably for processing medical notes in real-world settings in South Asia. We used publicly available LLMs to parse de-identified clinical notes from a large electronic health records (EHR) database in Pakistan, containing hospital records for 8.2 million patients. ChatGPT (GPT-3.5) as a general-purpose LLM, and GatorTron (base), BioMegatron, BioBert and ClinicalBERT as medical LLMs were evaluated when applied to these data, after fine-tuning them with (a) publicly available clinical datasets namely Informatics for Integrating Biology & the Bedside (I2B2) and National NLP Clinical Challenges (N2C2) for medical concept extraction (MCE) and emrQA for medical question answering (MQA), and (b) the local Pakistani de-identified EHR dataset, which includes inpatient Discharge Summaries (DS) and Subjective, Objective, Assessment, and Plan (SOAP) notes, as detailed in this paper. MCE models were applied to these clinical notes using both 3-label and 9-label formats, while MQA models were applied to medical questions. Internal and external validation performance was measured for (a) and (b) using F1 score, precision, recall, and accuracy for MCE and BLEU and ROUGE-L, which measure lexical and sequence similarity, for MQA. When clinical LLMs were not fine-tuned on the local EHR dataset, their performance during external validation on local data was notably poorer compared to internal validation on the dataset used for fine-tuning, with reductions of at least 15% in F1 scores for MCE and 35% in ROUGE-L and BLEU scores for MQA tasks. This suggests potential bias and highlights the inability of the medical LLMs to reliably handle the data distribution of the local population without further fine-tuning and adaptation. This trend persisted across two distinct natural language processing tasks: concept extraction and question answering, spanning a spectrum of task complexities. However, fine-tuning the LLMs with local EHR data significantly improved model performance across both tasks, yielding a 7.5% to 15% increase in the F1 score for MCE and a 27% to 53% increase in ROUGE-L and BLEU scores for MQA. Notably, ChatGPT, as a general-purpose LLM, stood out as an exception, demonstrating superior performance across all measured metrics on the local dataset compared to the publicly available dataset, with improvements ranging from 3% to 17% on the local EHR dataset, even without fine-tuning on the local data. Publicly available LLMs, predominantly trained on data from high-income regions, were found to be unreliable when applied in a real-world clinical setting in Pakistan. Fine-tuning them with local EHR data and regional clinical contexts improved their reliability, demonstrating a feasible adaptation strategy that is substantially less resource-intensive than training large language models from scratch. Close collaboration between local clinical and technical experts to curate and leverage more representative, inclusive, and unbiased medical datasets, can play a crucial role in further ensuring reliability of LLMs for resource-limited, overburdened settings, to be used in ways that are safe, fair, and beneficial for all.

Background: Accurate identification of cancer progression events from electronic health records (EHRs) can help enable promising oncology applications such as predicting disease trajectory, assessing treatment efficacy, and generating real-world evidence. These use cases require both large-scale and high-quality data but manual abstraction of real-world progression (rwP) is time-intensive, difficult to scale, and inherently challenging given the varied and unstructured ways it can be documented across cancer types. Large language models (LLMs) offer a scalable alternative, but their accuracy relative to expert human abstractors is unclear. We evaluated the ability of LLMs to extract rwP events and dates across 7 cancer types and assessed how using LLM-extracted data impacted real-world progression-free survival (rwPFS) estimates compared to using human-abstracted data. Methods: We applied LLM-based extraction techniques to unstructured EHR text for 7 cancer types from the Flatiron Health Research Database: bladder (N=377), breast (N=1000), colorectal (N=564), hepatocellular (N=217), renal cell (N= 229), non–small cell lung (N=1000), and small cell lung (N=955). Prompt engineering strategies including zero-shot, few-shot, and chain-of-thought were tested to optimize performance. We measured agreement between the LLM and abstractor on the presence of rwP (Y/N) and first rwP date (± 30 days) in the first-line (1L) setting. To contextualize the LLM’s ability to extract rwP relative to an expert human abstractor, we evaluated the difference in F1 scores between both curation approaches using a duplicate human-abstracted reference dataset. We also compared rwPFS calculated from LLM-curated data versus human abstractor-curated data for 1000 patients in each cancer type, indexed to 1L start date. Results: Across all cancer types, agreement between the LLM and abstractor on the presence of at least 1 rwP event ranged from 86%-90% while first rwP date agreement ranged from 80%-92%. The difference in F1 score between the LLM and human abstraction was within 3-8 points across cancer types. A comparison of rwPFS between the LLM and human abstractors showed &amp;lt;1 month difference in median rwPFS and overlapping 95% confidence intervals across all cancer types. Discussion: LLMs extracted rwP with high performance, achieving F1 scores similar to expert human abstraction. Across 7 distinct cancers, agreement with human-abstracted data aligned with published inter-abstractor reliability benchmarks, and rwPFS estimates were nearly identical across curation approaches, demonstrating both the generalizability and validity of the approach. These results highlight the potential of LLMs to extract high-quality clinical endpoints at scale, helping to advance research, enhance applications such as predictive algorithms, and ultimately supporting more personalized and effective cancer care. Citation Format: Aaron B. Cohen, Konstantin Krismer, Kelly Magee, James Gippetti, Aaron Dolor, Tori Williams, Erin Fidyk, Hank Kim, Qianyu Yuan, Melissa Estevez. Using large language models for scalable extraction of real-world progression events across multiple cancer types [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B006.

Introduction: Clinical and population decision-making relies on the systematic evaluation of extensive regulatory evidence. The FDA drug reviews provide detailed information on clinical trial design, enrollment criteria, sample size, randomization, comparators, endpoints, and indications. However, extracting these data is resource-intensive and time-consuming. Generative Artificial Intelligence large language models (LLMs) may accelerate the extraction and synthesis of such information. This study compares the performance of three LLMs, ChatGPT-4o, Gemini 2.5 Pro, and DeepSeek R1, in extracting and synthesizing regulatory and clinical information using antibiotics approved for complicated urinary tract infections (cUTIs) between 2010 and 2025 as a case study. Methods: LLM models were evaluated using general (short, direct) and detailed (structured, guidance-referencing) prompts across five domains including accuracy (precision and recall), explanation quality, error type (hallucination rate, misclassification, and omission), efficiency (response time, correct answers per second, and seconds per correct answer), and consistency with responses generated in duplicate runs. Two investigators independently reviewed outputs against FDA guidance, resolving discrepancies by consensus. Statistical analyses included χ², Wilcoxon, and Kruskal–Wallis tests with false discovery rate correction. Results: Among 324 responses, accuracy differed significantly across models (χ², p &lt;0.001) with Gemini 2.5 Pro achieving the highest accuracy (66.7%), followed by ChatGPT-4o (51.9%) and DeepSeek R1 (37.0%). General prompts outperformed detailed prompts (59.3% vs 44.4%; p =0.011). Gemini 2.5 Pro showed highest explanation quality and most consistent outputs, while ChatGPT-4o had shortest response times and highest efficiency. Hallucination was the most frequent error type across models. Conclusion: LLMs showed variable capability in extracting regulatory and clinical trial information. Gemini 2.5 Pro showed the strongest overall performance, while ChatGPT-4o was faster but less accurate, and DeepSeek R1 underperformed across most domains. These findings highlight both the promise and limitations of LLMs in regulatory science and support complementary use alongside human review to streamline evidence synthesis. Author Summary: Our research addresses a critical question in artificial intelligence for healthcare: how well do generative Generative Artificial Intelligence (GenAI) tools extract and synthesize regulatory and clinical information to inform decision-making? We assessed ChatGPT-4o, Gemini 2.5 Pro, and DeepSeek R1 performance in extracting and synthesizing information from regulatory documents and clinical trial data using all FDA approved antibiotics for the treatment of complicated urinary tract infections. We compared LLMs outputs directly with the original data sources. We assessed the models’ performance using both broad and detailed prompts across several areas, including accuracy of the information (precision and recall), quality of explanation, type of errors (hallucination, misclassification, and omission), efficiency and speed (response time, correct answers per second, and seconds per correct answer), and consistency of responses across repeated runs. The results suggest that while the models were generally fast and efficient extracting large volumes of information, they also produced errors and omissions that could limit their reliability. These findings highlight both the promise and the current limitations of GenAI, underscoring its potential value as a human supervised tool for safely supporting regulatory science and clinical decision-making.

Large Language Models (LLMs) can solve some undergraduate-level to graduate-level physics textbook problems and are proficient at coding. Combining these two capabilities could one day enable AI systems to simulate and predict the physical world. We present an evaluation of state-of-the-art (SOTA) LLMs on PhD-level to research-level computational physics problems. We condition LLM generation on the use of well-documented and widely-used packages to elicit coding capabilities in the physics and astrophysics domains. We contribute ∼50 original and challenging problems in celestial mechanics (with REBOUND), stellar physics (with MESA), 1D fluid dynamics (with Dedalus) and non-linear dynamics (with SciPy). Since our problems do not admit unique solutions, we evaluate LLM performance on several soft metrics: counts of lines that contain different types of errors (coding, physics, necessity and sufficiency) as well as a more educational’ Pass-Fail metric focused on capturing the salient physical ingredients of the problem at hand. As expected, today's SOTA LLM (GPT4) zero-shot fails most of our problems, although about 40% of the solutions could plausibly get a passing grade. About 70%–90% of the code lines produced are necessary, sufficient and correct (coding & physics). Physics and coding errors are the most common, with some unnecessary or insufficient lines. We observe significant variations across problem class and difficulty. We identify several failure modes of GPT4 in the computational physics domain, such as poor physical units handling, poor code versioning, tendency to hallucinate plausible sub-modules, lack of physical justification for global run parameters (e.g., simulation time, or upper-lower bounds for parametric exploration) and inability to define steady-state or stopping conditions reliably. Our reconnaissance work provides a snapshot of current computational capabilities in classical physics and points to obvious improvement targets if AI systems are ever to reach a basic level of autonomy in physics simulation capabilities.