搜索结果：Opus

BackgroundLarge language models (LLMs) increasingly generate clinical recommendations, but their ability to translate biliary guidelines into safe procedural triage remains uncertain. We evaluated next-generation LLMs for ERCP indication in suspected choledocholithiasis and tested whether errors could affect workflow.MethodsA cross-sectional in-silico diagnostic accuracy study was conducted from May 14 to May 18, 2026. One hundred locked synthetic vignettes were mapped to ASGE/ESGE-based standards: 45 ERCP-indicated and 55 nonindicated cases. GPT-5.5, Gemini 3.0 Pro, and Claude 4 Opus were queried with an identical zero-shot prompt at temperature 0.0. Outcomes included accuracy, sensitivity, specificity, kappa, error phenotype, and simulated under-triage delay.ResultsGPT-5.5 achieved the highest accuracy (96.0%; 95% CI, 90.2%-98.4%), followed by Gemini 3.0 Pro (90.0%; 95% CI, 82.6%-94.5%) and Claude 4 Opus (84.0%; 95% CI, 75.6%-89.9%). Agreement was near-perfect for GPT-5.5 (kappa = 0.92), substantial for Gemini 3.0 Pro (kappa = 0.80), and weaker for Claude 4 Opus (kappa = 0.68). GPT-5.5 outperformed Claude 4 Opus (McNemar P = .004). Claude 4 Opus produced the most under-triage errors (n = 9) and the largest simulated delay burden (163.8 hours per 100 vignettes; Kruskal-Wallis P = .007).ConclusionNext-generation LLMs can approximate guideline-based ERCP triage, but clinically meaningful differences emerge when errors are weighted by procedural delay and safety. GPT-5.5 showed the most balanced profile; conservative under-triage remains the key hazard requiring supervision.

To benchmark the diagnostic accuracy and inter-model agreement of contemporary large language models (LLMs) on text-only histopathologic narrative descriptions in oral and maxillofacial pathology (OMFP), and to assess how performance varies by diagnostic dependency and diagnostic category. In this retrospective diagnostic accuracy study, 171 de-identified OMFP cases were screened, and 155 cases were included after predefined selection. Each case comprised an edited histopathologic narrative and a single expert-verified reference diagnosis. Three general-purpose LLMs (ChatGPT-5.0, Gemini-2.5-Pro, Claude-Opus-4.1) were queried once per case using an identical zero-shot prompt requesting a single definitive diagnosis. Primary outcome was case-level accuracy; paired comparisons used McNemar's test, and inter-model agreement was assessed with Cohen's κ. Performance was stratified by histology-sufficient (HS) versus correlation-dependent (CD) lesions and by OMFP diagnostic category. Overall accuracies were 83.2% (ChatGPT-5.0), 77.4% (Gemini-2.5-Pro), and 72.3% (Claude-Opus-4.1). ChatGPT-5.0 significantly outperformed Claude-Opus-4.1, whereas the differences between ChatGPT-5.0 and Gemini-2.5-Pro and between Gemini-2.5-Pro and Claude-Opus-4.1 were not statistically significant. Agreement was highest between Gemini-2.5-Pro and Claude-Opus-4.1 (κ = 0.73). All models showed higher accuracy for HS than CD lesions, with best performance in malignant, hematolymphoid, and immune-mediated categories. Contemporary LLMs can interpret OMFP histopathologic narratives with moderate overall diagnostic accuracy, particularly for lesions with distinctive microscopic features. Performance declined for CD entities, indicating persistent reliance on clinicoradiologic context. Although not suitable as stand-alone diagnosticians, these findings provide a controlled benchmark of text-based diagnostic performance and suggest potential supportive value in OMFP education and exploratory pathology informatics applications under expert supervision.

This study investigated role-based differences in the accuracy, readability, understandability, and internal consistency of chatbot responses to orthodontic emergencies and examined the clinical implications of patient-oriented communication. Twenty-three standardized orthodontic emergency scenarios were presented to four chatbots- ChatGPT-4o, Claude 3 Opus, Microsoft Copilot, and Gemini 2.5-using patient and orthodontist roles. Response accuracy was evaluated by expert orthodontists and research assistants using a 3-point Likert scale, while readability, understandability, and internal consistency were assessed with the Atesman index, Sonmez formula, and Cronbach's &#x3b1;. Patient-role responses were descriptively analyzed using predefined communication dimensions to contextualize quantitative findings. Significant chatbot-specific role-based differences were observed, with Claude 3 Opus (p&#x2009;=&#x2009;0.001) and Gemini 2.5 (p&#x2009;=&#x2009;0.023) showing higher accuracy in the orthodontist role. In the patient role, ChatGPT-4o and Claude 3 Opus showed the highest rates of correct information, while Claude 3 Opus had the highest rate in the orthodontist role. Patient-role responses were significantly more understandable than orthodontist-role responses (p&#x2009;<&#x2009;0.05). ChatGPT-4o (&#x3b1;&#x2009;=&#x2009;0.862) and Gemini 2.5 (&#x3b1;&#x2009;=&#x2009;0.815) showed high internal consistency. Qualitative analysis indicated that patient-oriented responses frequently adopted a reassuring tone and emphasized temporary self-care strategies, potentially influencing perceived urgency in orthodontic emergencies. Chatbot performance varied according to user role. Patient-oriented responses were more understandable despite similar readability across models and could influence perceptions of urgency and professional responsibility, highlighting the need for cautious framing of chatbot-generated information for orthodontic emergency guidance. Chatbots can provide preliminary information in orthodontic emergencies; however, due to limitations in accuracy and consistency, they should be used only as supportive tools and should not replace professional clinical judgment.

To compare the accuracy and safety of 4 large language models on cornea and external disease multiple-choice questions (MCQs). DeepSeek-V3.2, GPT-5.2 (via ChatGPT), Gemini 3 Pro, and Claude Opus 4.5 were tested on American Academy of Ophthalmology (AAO) Ophthalmic News and Education (ONE) Network Cornea/External MCQs (114 text-only) and AAO Basic and Clinical Science Course Cornea/External study questions (38 text-only). Each item was queried once per model using a standardized prompt for the primary analysis. In secondary analyses, each item was requeried 5 times per model in independent sessions. Accuracy was compared using Cochran Q and pairwise exact McNemar tests with Holm adjustment. Potentially harmful wrong answers (HarmfulWrong) were independently coded using a prespecified rubric. In the AAO ONE dataset, accuracy was 65.8% for DeepSeek-V3.2, 93.0% for GPT-5.2, 93.9% for Gemini 3 Pro, and 92.1% for Claude Opus 4.5 (Cochran Q P < 0.001). DeepSeek-V3.2 was significantly less accurate than all other models; the other 3 did not differ significantly. In the Basic and Clinical Science Course dataset, accuracy was 57.9%, 76.3%, 94.7%, and 84.2%, respectively (Cochran Q P < 0.001); DeepSeek-V3.2 was significantly less accurate than Gemini 3 Pro and Claude Opus 4.5. Five-run retesting showed similar ranking but imperfect stability, particularly for DeepSeek-V3.2. HarmfulWrong responses were infrequent, and 2 cornea specialists achieved 98.0% and 97.4% overall accuracy without HarmfulWrong responses. GPT-5.2, Gemini 3 Pro, and Claude Opus 4.5 achieved similarly high accuracy on text-only cornea/external disease MCQs, whereas DeepSeek-V3.2 underperformed. Potentially harmful errors were uncommon but support continued clinician oversight.

To evaluate the impact of preheating on the physical properties of bulk-fill resin composites, with a particular focus on microhardness and surface roughness. In this study, seven bulk-fill resin composites (SonicFill 3, Opus Bulk-Fill Flow, Opus Bulk-Fill, Metafil Bulk-Fill, Tetric PowerFlow, Tetric N-Ceram, and Estelite Bulk-Fill Flow) and one conventional resin composite (Filtek Z250) were tested. Specimens were divided into two main groups: Group 1 (preheated 68&#xb0;C for 10 minutes) and Group 2 (stored at room temperature). For each group, five disc-shaped resin composites (4 mm thickness x 5 mm diameter) were prepared. Each bulk-fill resin composite was applied in a single increment, whereas Filtek Z250 was placed in the molds in 2 mm layers. To complete the polymerization, the samples were kept in distilled water at 37&#xb0;C for 24 hours. Then, the surfaces of the samples were polished with Twist Dia (Clearfil) polishing discs to imitate the finishing and polishing processes. Baseline microhardness and surface roughness values were measured. Subsequently, specimens underwent artificial aging simulating 2 years of clinical use (10,000 brushing cycles + 1,200 thermal cycles). Post-aging measurements were repeated, and statistical analyses were conducted. Significant differences in microhardness values were observed among the experimental groups for all restorative materials, except for Estelite Bulk-Fill Flow (P> 0.05). The highest bottom-to-top hardness ratio was found in both preheated and non-preheated Opus Bulk-Fill Flow, while the lowest was observed in preheated Tetric PowerFlow. Microhardness generally decreased with depth, and the impact of preheating varied among materials. Post-aging surface roughness increased in all groups. Preheated Metafil Bulk-Fill exhibited the highest surface roughness, whereas Estelite Bulk-Fill Flow (both preheated and non-preheated) showed the lowest. Preheating reduced viscosity and enhanced microhardness in certain resin composites but also increased surface roughness over time. The effects of preheating are material-dependent due to compositional differences. Preheating may improve handling and selected properties of resin composites; however, its long-term effects should be considered when selecting materials for clinical use.

Background/Objectives: Accurate estimation of nutritional content from food images has important applications in dietary assessment and public health surveillance. While large language models (LLMs) have shown promise for this task, the effects of prompt design and model selection on estimation accuracy remain poorly characterized. Methods: We evaluated three Claude models (Haiku 4.5, Sonnet 4.6, Opus 4.6) for visual estimation of five mandatory nutritional components (energy, protein, fat, carbohydrate, and salt equivalent) across three datasets: NutriImage (691 Japanese meal photographs with dietitian-validated ground truth, after OCR-mask quality filtering), SNAPMe (1463 US meal photographs from a publicly available benchmark), and the Japan Branded Food Database (JBFD; 989-1000 packaged food product images). We systematically compared a default prompt and a visual estimation prompt explicitly instructing the model not to read any text or numbers visible in the image. Results: The visual estimation prompt substantially improved accuracy when paired with a sufficiently capable model (energy R2: 0.23 for Haiku to 0.60 for Sonnet, JBFD). Sonnet and Opus substantially outperformed Haiku across all datasets, while differences between Sonnet and Opus were small (MedAPE difference 1-3 percentage points). Packaged food images (JBFD) yielded higher R2 than meal photographs. Salt equivalent showed consistently poor accuracy (MedAPE 34-64%). On SNAPMe, Sonnet achieved lower energy MAE (116.9 vs. 123.0 kcal, -4.9%) and lower MAE for protein (5.9 vs. 7.9 g, -25.7%) and fat (6.6 vs. 8.7 g, -24.5%) compared with a recent ChatGPT-5 study. Conclusions: Claude Sonnet offers the best cost-performance balance for LLM-based nutritional estimation. Prompt design substantially affects accuracy, but only when paired with a sufficiently capable model; model visual recognition capability appears to be a key determinant of performance. These findings highlight the inherent difficulty of this task and provide practical guidance for dietary assessment system development.

Prior authorization (PA) for exome or genome sequencing is a time-consuming process that impedes timely rare disease diagnosis. Large language model-based browser agents offer potential for automating these workflows, but their clinical reliability remain uncharacterized. We developed a sandbox compromising a simulated ES/GS PA submission payer portal and a synthetic EHR containing 836 patient records spanning compliant profiles and deficient profiles with different types of issues. Gemini 3 Pro, Gemini 3 Flash, and Claude Opus 4.5 were evaluated on task completion rate, form completion accuracy, and appropriate withholding for deficient profiles. Larger models achieved much higher task completion rates (Gemini 3 Pro 95.45%, Claude Opus 4.5 93.67%) compared to Gemini 3 Flash (56.05%), but nearly universally failed to withhold submission for deficient profiles whereas Gemini 3 Flash ironically demonstrated superior withholding performance (17.33%). In a non-agentic setting, Gemini 3 Pro correctly identified 91% of the issues in deficient profiles, indicating that withholding failure is attributable to the browser interaction rather than the model's reasoning limitations. Current LLM-based browser agents exhibit a systematic bias towards form submission that poses risks in PA workflows. A modular, multi-agent architecture with human supervision is necessary for a safe clinical deployment.

Background: Data extraction for systematic reviews is highly resource-intensive. This study evaluated four frontier large language models (LLMs) on complex structured metadata extraction from specialized neuroimaging artificial intelligence (AI) literature to determine their performance in automated evidence synthesis. Methods: We compared Google Gemini 3 Pro Preview, Anthropic Claude Opus 4.5, Perplexity Sonar Pro, and OpenAI GPT 5.2. Using a standardized prompt, each model extracted 22 variables from 91 peer-reviewed neuroimaging AI articles. The variables were stratified into low-, medium-, and high-complexity tiers. The performance was measured via the exact-match accuracy against a consensus-based expert ground truth. Results: The overall exact-match accuracy was moderate. Gemini 3 Pro Preview achieved the highest overall rate (56.4%), followed by Sonar Pro (52.1%), Claude Opus 4.5 (51.3%), and GPT 5.2 (46.5%). Gemini significantly outperformed all other models (p < 0.001). The performance declined dramatically as the variable complexity increased. Across models, the accuracy was 88.9-92.9% for low-complexity categorical fields, 47.0-63.3% for medium-complexity text extraction, and 2.7-15.5% for high-complexity variables requiring clinical judgment or multi-section synthesis. The most common type of error was misclassification. All four models scored 0% on the main performance metric, but this reflected a representational mismatch with the ground truth rather than extraction failure, indicating that the exact-match accuracy underestimates the true semantic performance. Conclusions: Frontier LLMs can effectively automate the retrieval of simple categorical data, but have serious difficulties with methodological variables that are complex. Although extraction can be fully automated for low-complexity fields, human review remains essential for context-dependent variables that require clinical judgment.

Large language models (LLMs) have emerged as promising tools for estimating energy and nutrient values, yet most existing evaluations focus on image-based queries rather than text. Few studies compare LLM estimates with reference databases commonly used in nutrition research. To examine agreement between LLM estimates and a research food composition database for energy and nutrients, and to determine if agreement varies by food group. We conducted a cross-sectional analysis of energy and nutrient estimates for frequently consumed food items in the United States (US). Food items were entered as text prompts into four LLMs (ChatGPT 5.2, Claude Opus 4.5, Gemini 3 Pro Preview, and Llama 4 Maverick), which provided energy and nutrient estimates. Corresponding foods were matched to the Nutrition Coordinating Center (NCC) Food and Nutrient Database, and agreement between LLM and database values was assessed using intraclass correlation coefficients (ICCs) and Bland-Altman analyses. Agreement was also evaluated within the three most frequently consumed food groups. Agreement was high for energy and macronutrients for all LLMs. Variability was observed for several micronutrients, particularly vitamin D, folate, and iron. Claude Opus 4.5 showed consistently high agreement, with no nutrients classified as poor. Other LLMs exhibited poor agreement for at least one micronutrient. Certain food categories, including condiments and mixed dishes, contributed disproportionately to variability. However, agreement remained high within the most frequently consumed broader food groups. LLMs show promise for estimating energy and macronutrients. However, performance for micronutrients requires further improvement and may affect overall dietary assessment.

This essay offers a re-examination of Tom Nairn's lifelong prediction of an impending Break-Up of Britain, against the backdrop of the Union's apparent resilience since the publication of his magnum opus in 1977. Specifically, it considers Nairn's conceptualization of the end of empire as the necessary precursor event that doomed the UK to inevitable destruction. What Perry Anderson terms the 'protracted delay' between Nairn's predictions and their fulfilment provides a useful vantage point to reconsider Nairn's thesis with the benefit of more than 50 years' hindsight. It is argued that Nairn's Gramscian model and his inherently structural view of empire tended to screen out the complex social interface between Britain and the wider imperial world. For all the crucial explanatory power he attributed to the empire's demise, he almost never exemplified his argument with reference to any specific part of it or referred to the corrosive effects of any given decolonizing moment. To gauge the empire's full significance, then, requires a larger frame of reference, comprising not only the unitary UK state but also the empire-wide affinities that were so crucial to popular investments in being British. To the extent that Nairn was right about the intricate nexus tying the Union to Britain's global coordinates, it was not always for the right reasons.

To evaluate the performance of secure cloud-based large language models (LLMs) in extracting glaucoma diagnosis, type, and severity from free-text clinical notes in the electronic health record (EHR). Retrospective chart review analysis. 1,250 subjects from the Bascom Palmer Ophthalmic Repository. Clinical notes of glaucoma-related encounters between 2014 and 2024 were extracted from the Bascom Palmer Ophthalmic Repository. Two fellowship-trained glaucoma specialists annotated clinical notes for glaucoma presence, type, and severity at the eye level. The dataset was split into development (10%), validation (10%), and test (80%) sets. Development and validation sets were used for prompt engineering and refinement, and the held-out test set was used for evaluation. Five LLMs (Claude Opus 4.6, DeepSeek-V3.2, GPT-5.2, Grok 4.1, and Qwen3.6-35B-A3B) were accessed via Azure AI Foundry within HIPAA-compliant containers. Model performance was assessed using standard metrics. Clinician-entered ICD-10 codes were also compared with adjudicated labels. Gwet AC1, accuracy, sensitivity, specificity, and F1-score. Inter-grader agreement was high for glaucoma detection (Gwet AC1= 0.930 (95% CI: 0.917-0.945), type classification (Gwet AC1= 0.917 (95% CI: 0.904-0.930), and severity staging (Gwet AC1= 0.901 (95% CI: 0.884-0.916). For glaucoma diagnosis, LLMs demonstrated high overall accuracy, with Claude achieving 97.5%, DeepSeek 96.0%, GPT 96.2%, Grok 94.4%, and Qwen 95.5%. F1 scores for glaucoma detection ranged from 95.4% to 98.9% across models. For glaucoma type classification, accuracies were 97.1%, 94.2%, 94.2%, 94.0%, and 94.4% for Claude, DeepSeek, GPT, Grok, and Qwen, respectively. F1 scores for the most prevalent type (POAG) ranged from 96.3% to 98.9%. For severity staging, accuracies were 95.0%, 94.8%, 94.5%, 94.0%, and 95.2%, respectively, with F1 scores ranging from 89.7% to 96.3% across severity categories and models. ICD-10 codes demonstrated substantially lower performance for type and severity staging, with overall accuracies of 89.2% and 58.5%, respectively. Secure cloud-based LLMs accurately extracted glaucoma diagnosis, type, and severity information from free-text ophthalmology notes, achieving performance approaching expert clinician adjudication while substantially outperforming ICD-based phenotyping approaches, particularly for disease severity classification. These findings demonstrate the potential of LLMs to transform unstructured clinical documentation into scalable, research-ready phenotypic data for large-scale glaucoma cohort development and EHR-based ophthalmic research.

Large language models (LLMs) increasingly inform mental health decisions by patients and clinicians. Inference-time activation steering can shift model behavior on a target dimension without altering weights or prompts and without disclosure to users, allowing treatment recommendations to be silently changed for commercial or ideological reasons. To determine whether directional activation steering can shift an open-weights LLM's depression treatment recommendations. This non-human subjects study applied directional activation steering to an open-weights LLM (DeepSeek V4 Flash) responding to 12 depression-advice scenarios (4 favoring medication, 4 favoring avoidance, 4 neutral), generated at 30 amplitudes from -1.5 to +1.5 in 0.1 increments plus an unsteered baseline. A single steering direction contrasting antidepressant medication with self-directed approaches (diet, exercise, meditation, dietary supplements), constructed from 16 paired training prompts and applied at the attention output of every transformer block; weights and system prompt were held constant. The extent to which medication and four self-care categories were addressed, scored 0 to 3 by a human-validated LLM rater (Claude Opus 4.7), the medication-versus-self-care balance, and clinician referral, estimated per unit of amplitude using mixed-effects models with a scenario random intercept. Across 372 generations, steering produced a graded, dose-dependent shift in the medication-versus-self-care balance, which declined by 0.32 per unit of amplitude (&#x3b2; = -0.32; 95% CI, -0.39 to -0.25; P < .001); medication extent fell and self-care extent rose. The shift was largest for scenarios with no stated treatment preference (&#x3b2; = -0.44; 95% CI, -0.54 to -0.34; P < .001). A clinician referral appeared in 322 of 372 responses (87%) and did not vary with steering amplitude (P = .63). In this open-weights LLM providing depression treatment information, inference-time activation steering shifted treatment recommendations without altering weights, prompt structure, or safety outputs, with the largest effect among users expressing no treatment preference. These findings suggest a need for LLM disclosure standards and independent auditing as such models inform clinical decisions.

Large language models (LLMs) are increasingly explored for qualitative analysis, but the effect of workflow design on thematic fidelity remains unclear. This study evaluated a structured human-AI collaboration framework using Claude Opus 4.6 to analyze 16 interview transcripts from patients with chronic obstructive pulmonary disease participating in a pulmonary telerehabilitation program. The workflow included code extraction, code combination, and theme generation, and was tested using hierarchical and direct strategies. AI-generated themes were compared with human-derived themes using sentence-t5-xxl embeddings and cosine similarity, with theme alignment performed using Hungarian and greedy matching. Output volume varied substantially across strategies, ranging from 53 to 357 codes and 11 to 17 themes. Direct grouping (average cosine similarity 0.891) and L3 grouping (0.890) achieved the highest similarity to human-generated themes. These findings suggest that grouping-based workflows can preserve key information, reduce redundancy, and improve thematic generation in LLM-assisted qualitative analysis.

The purpose of this study was to examine the relationships between demographic variables, ego-resilience and health-related quality of life (HRQoL) for individuals with upper limb amputations. As HRQoL continues to be an important measure of rehabilitative success, determining universal factors which correlate to and may predict HRQoL becomes more important in the clinical space. A sample of 90 previously administered outcomes from patients at a national upper limb prosthetic provider in the United States were gathered. The outcome measure, the Wellness Inventory, captured patient-reported data to screen for mental health status including ego-resilience, PTSD, depression, coping mechanisms. Scores from the Orthotics and Prosthetics Users' Survey (OPUS) HRQoL as well as the Ego-Resilience Scale (ER89) were utilized in this study, as well as pertinent demographic data. Comparative analyses were conducted on the data gathered. HRQoL and Ego-Resilience scores were analyzed alongside demographic factors: gender (77.8% male), age at time of amputation (mean age 38, SD&#x2009;=&#x2009;12.9), level of amputation, ethnicity and marital status. Correlational analysis showed positive relationship between ego-resilience and HRQoL (&#x3c1;&#x2009;=&#x2009;0.332, p&#x2009;=&#x2009;.002). Simple linear regression analysis found a significant relationship between ethnicity and HRQoL (&#x3b2;&#x2009;=&#x2009;4.237, p&#x2009;=&#x2009;.047), and ego-resilience and HRQoL (&#x3b2;&#x2009;=&#x2009;0.910, p=<.001). The multiple linear regression model identified ego-resilience and ethnicity as predictors for HRQoL (adjusted R&#xb2;&#x2009;=&#x2009;.129, F&#x2009;=&#x2009;7.576, p=<.001). Based on the findings in this study, ego-resilience has been identified as a significant predicting factor for HRQoL. Higher ego-resilience score and trait likely will result in higher HRQoL scores. Understanding of how demographic variables, such as ethnicity, may directly or indirectly impact HRQoL can also be beneficial in the recovery process.

Large language models are increasingly used to obtain health information, but their quality in pediatric anesthesia remains insufficiently evaluated. This study aimed to assess the reliability and readability of four widely used AI chatbots in this context. This cross-sectional observational study developed 18 pediatric anesthesia-related questions using Medical Subject Headings terms, online search trend analysis, and commonly queried topics reflecting parental information needs. Each question was submitted under standardized conditions to four generative AI-driven chatbots: OpenAI's GPT-5.1 Thinking, Google's Gemini 3 Pro, Anthropic's Claude Opus 4.5 Extended Thinking, and DeepSeek-V3.2-Speciale. Models were accessed in their vendor-deployed configurations without task-specific fine-tuning. The generated responses were evaluated for information reliability using the Ensuring Quality Information for Patients (EQIP) instrument, DISCERN tool, Global Quality Score (GQS), and Journal of the American Medical Association (JAMA) benchmark criteria. Readability was assessed using seven validated indices including Flesch Reading Ease Score, Flesch-Kincaid Grade Level, Gunning Fog Index, Simple Measure of Gobbledygook, Coleman-Liau Index, Automated Readability Index, and Linsear Write Formula. A total of 72 chatbot-generated responses were included for analysis. Significant between-model differences were observed in DISCERN, EQIP, and GQS, while JAMA benchmark scores were consistently low across all models. DeepSeek and Gemini showed higher median reliability scores across several instruments, although significant pairwise differences mainly involved ChatGPT. None of the evaluated models achieved the recommended sixth-grade readability level across any index. Correlations between reliability and readability were non-significant, suggesting that these represent independent dimensions of information quality. Current LLM-based chatbots provided pediatric anesthesia information with variable reliability and consistently suboptimal readability. Although certain models demonstrated relatively higher information quality, limited transparency and excessive reading complexity may restrict their suitability for public-facing educational use. These findings highlight the need for improved quality control, enhanced transparency, and readability-focused optimization in pediatric perioperative education.

The National Institutes of Health Stroke Scale (NIHSS) is critical to acute stroke care but is often documented in unstructured notes. Large language models (LLMs) can enable automated extraction, though smaller models often underperform relative to frontier systems. Chain-of-Verification (CoVe) prompting introduces a structured self-verification step that may improve performance. We evaluated eight LLMs on 312 discharge summaries. Small models included LLaMA 3.2 3B, Ministral 3B, Gemma 3 4B, and Qwen 3 4B. Frontier models included GPT-5.2, Gemini 3 Pro, Claude Opus 4.5, and Grok 4. Each model was tested under a baseline and CoVe prompt. Outcomes were subscore exact-match accuracy, subscore mean absolute error (MAE), total score exact-match accuracy, and total score MAE. At baseline, small models achieved 53.2&#xa0;&#xb1;&#xa0;10.0% subscore accuracy and subscore MAE 0.84&#xa0;&#xb1;&#xa0;0.22, compared with 88.5&#xa0;&#xb1;&#xa0;10.1% and 0.15&#xa0;&#xb1;&#xa0;0.16 in frontier models (both p&#xa0;<&#xa0;0.001). Total exact accuracy was low in both groups (7.7&#xa0;&#xb1;&#xa0;12.9% vs 35.9&#xa0;&#xb1;&#xa0;32.4%). CoVe significantly improved small-model performance (subscore accuracy 65.0&#xa0;&#xb1;&#xa0;10.9%; subscore MAE 0.55&#xa0;&#xb1;&#xa0;0.21; total MAE 4.84&#xa0;&#xb1;&#xa0;2.30 vs 7.19&#xa0;&#xb1;&#xa0;3.54 at baseline; all p&#xa0;<&#xa0;0.001), although total exact accuracy remained modest (9.6&#xa0;&#xb1;&#xa0;15.7%). Frontier models showed no significant group-level change with CoVe. CoVe prompting substantially improves NIHSS extraction in small LLMs while producing negligible effects in frontier models. Although smaller model performance remains insufficient for standalone clinical deployment, CoVe prompting offers a promising avenue for further exploration.

To evaluate the efficacy of large language models (LLMs) in extracting medication-related information from glaucoma clinical notes in the electronic health record (EHR). Cross-sectional. 1,250 subjects in the Bascom Palmer Ophthalmic Repository. Extracted clinical notes from glaucoma-related encounters between 2014 and 2024 were labeled by two glaucoma specialists with a third serving as an adjudicator. Graders were asked to label current topical medications (CTM), proposed changes to topical medications (&#x394;TM), current oral medications (COM), and proposed changes to oral medications (&#x394;OM) in a structured fashion. The dataset was split into development (10%), validation (10%), and test (80%) sets stratified by clinician. Development and validation sets were used to engineer and refine prompts, and the held-out test set was used for model assessment. Five LLMs (Claude Opus 4.6, DeepSeek-V3.2, GPT 5.2, Grok 4.1, and Qwen3.6-35B-A3B) were accessed via Microsoft Azure AI Foundry within a HIPAA-compliant environment. Inter-grader agreement was assessed with Gwet AC1. LLM performance was initially assessed in a binary fashion with F1 scores, and the degree of text match among positive cases was evaluated using exact match accuracy and Jaccard Index (JI). F1 score, exact match accuracy, JI. Gwet AC1 for intergrader agreement was 0.799, 0.888, 0.985, and 0.988 for CTM, &#x394;TM, COM, and &#x394;OM, respectively. F1 scores for CTM were 0.985, 0.971, 0.978, 0.968, and 0.970 for Claude, Deepseek, GPT, Grok, and Qwen, respectively; for &#x394;TM: 0.905, 0.826, 0.897, 0.842, 0.855, respectively; for COM: 0.923, 0.887, 0.899, 0.906, 0.894, respectively; for &#x394;OM: 0.958, 0.815, 0.937, 0.835, 0.940, respectively. Among positive cases, range of exact match accuracies for CTM (N=1354) was 0.730-0.882 and range of JIs was 0.809-0.918. For &#x394;TM (N=404), exact match accuracy range was 0.619-0.780 and JI range was 0.668-0.827. For COM (N=47), exact match accuracy range was 0.766-0.872 and JI range was 0.765-0.870. For &#x394;OM (N=25), exact match accuracy range was 0.583-0.920 and JI range was 0.583-0.922. The GLLaucoMed pipeline demonstrated high performance in extracting and standardizing medication data from unstructured clinical notes, including both current medications and proposed changes. Claude and GPT exhibited the strongest performance.

Specialized clinical artificial intelligence (AI) tools are entering medical practice despite scarce independent evaluation. We quantitatively evaluate two clinical AI tools, OpenEvidence and UpToDate Expert AI, built on large language models (LLMs) against three frontier LLMs: GPT-5.2, Gemini 3.1 Pro and Claude Opus 4.6. Our evaluation has three stages: (1) 500 MedQA questions testing medical knowledge, (2) 500 HealthBench items measuring alignment with clinicians and (3) the real clinical queries (RCQ) benchmark, built from 100 de-identified queries from physicians to a general-purpose language model in a live clinical environment. For the RCQ benchmark, 12 US clinicians performed randomized, blinded review of model outputs, producing 1,800 model-question annotations. Frontier LLMs outperformed clinical AI tools in all three evaluations. Clinical AI tools performed comparably to auto-enabled Google Search AI Overview on the RCQ. These findings highlight the need for independent, real-world evaluation of AI tools before they enter clinical settings.

Personalized perioperative fluid therapy is important for reducing postoperative complications and adverse outcomes. Although large language models (LLMs) show promise in healthcare, their application in fluid therapy remains challenged by hallucinations, limited domain-specific knowledge, and insufficient personalization. To address these limitations, Retrieval-Augmented Generation (RAG) is an effective method, while Knowledge Graphs (KGs) provide more accurate and reliable information. In this paper, we constructed a Personalized Fluid Therapy Knowledge Graph (PFTKG) comprising 6,490 entities and 15,687 relationships, and adapted GraphRAG, a graph-based RAG strategy that employs community detection and recursive summarization to support finding-level retrieval of clinically relevant information. We compared GraphRAG with document-based retrieval-augmented generation (DocRAG) and mainstream prompting strategies, including Vanilla, Chain-of-Thought (CoT), and Reflection-of-Thoughts (RoT), across three LLMs: GPT-4o, Claude Opus 4, and Gemini 2.5 Pro. Performance was evaluated using a 300-question knowledge-based question set and a 262-question retrospective case-based question set derived from 206 abdominal surgery patients. Evaluation included accuracy, honesty, error composition, response length, and response time. On the knowledge-based question set, GraphRAG achieved the highest average accuracy: 96.89% for multiple-choice questions and 66.44% for open-ended questions. On the retrospective case-based question set, GraphRAG also showed the strongest overall performance, with an average accuracy of 71.12%, compared with 62.47% for DocRAG, 54.20% for CoT, 52.67% for Vanilla, and 52.54% for RoT. Adding a "Don't know" option increased explicit acknowledgment of uncertainty, and GraphRAG reduced context-irrelevant errors compared with DocRAG. These results support GraphRAG as a domain-adapted retrieval strategy for personalized perioperative fluid therapy question answering. By integrating a clinical knowledge graph with hierarchical summarization and finding-level retrieval, it improved answer accuracy and promoted more conservative responses under uncertainty in both knowledge-based and retrospective case-based evaluations, supporting its use in future clinically integrated studies.