A benchmark evaluation framework for measuring the faithfulness of LLM-generated clinical trial summaries across three stakeholder audiences is introduced and a knowledge-graph-augmented retrieval system was developed and evaluated, producing statistically significant improvements in NLI-based faithfulness scores.
Abstract
Large language models are increasingly used to summarize clinical trial results for healthcare providers, patients, and payers, but their tendency to hallucinate poses significant risks in this high-stakes context. This study introduces a benchmark evaluation framework for measuring the faithfulness of LLM-generated clinical trial summaries across three stakeholder audiences. The framework consists of 200 stratified trials drawn from the Aggregate Analysis of ClinicalTrials.gov database, evaluated using audience-specific prompt templates and a six-dimension faithfulness annotation schema. Baseline measurements were established for GPT-4o, Claude Sonnet 4.6, and Gemini 2.5 Flash across 1,800 generated summaries scored using a cross-encoder natural language inference (NLI) model. Unsupported Claims was identified as the dominant failure mode across all three models, with a mean annotation score of 1.55 out of three. A knowledge-graph-augmented retrieval system was developed and evaluated against the baseline, producing statistically significant improvements in NLI-based faithfulness scores (entailment +0.0125, faithfulness +0.0130, p<0.0001). Improvement pathways were model-dependent, with GPT-4o improving primarily through contradiction reduction while Claude Sonnet 4.6 and Gemini 2.5 Flash improved through increased entailment.
Background/Objectives: Large language models (LLMs) show promise for clinical decision support, yet their accuracy in interpreting specialized medical guidelines remains uncertain. Retrieval-augmented generation (RAG) may enhance performance by grounding responses in authoritative knowledge bases. This study aimed to compare the accuracy, comprehensiveness, and safety of RAG-enhanced versus standard LLMs for answering clinical questions derived from the German S3 guideline for oral cavity carcinoma. Methods: We conducted a prospective, single-blind benchmark study evaluating six LLMs: one RAG-enhanced model (Custom GPT with guideline access), one consensus-based model (ConsensusGPT), and four standard models (DeepSeek-V3.2, Mistral Small 3.2, Qwen3-Next-80B, GPT-OSS-120B). Fifty clinical questions covering 17 guideline domains were presented to each model three times, yielding 900 evaluations. Three expert reviewers assessed responses using 5-point Likert scales for accuracy, comprehensiveness, and clarity, under a single-blind procedure, the effectiveness of which was tested by a pre-specified manipulation check. We then ran a paired within-model experiment in which each base model was queried with and without guideline access through a transparent, openly released retrieval pipeline, and scored every response with a condition-blind automated judge alongside deterministic retrieval metrics computed from the logs. Secondary outcomes included hallucination rates and guideline citation behavior. Inter-rater reliability was assessed using intraclass correlation coefficients (ICCs). Results: In a paired within-model design that held each base model fixed, adding transparent guideline retrieval improved accuracy—significantly in the three weaker open-weight models (Mistral, Qwen3, and GPT-OSS) and directionally in the already-strong DeepSeek and GPT-5 bases. Because a pre-specified blinding check found that experts could still identify retrieval-augmented answers with 98.5% accuracy, we anchored causal interpretation on measures that do not depend on the human raters, ranked by their independence: deterministic, log-derived retrieval metrics first, and then an automated, condition-blind LLM judge, whose agreement with the experts (Spearman ρ = 0.81, 95.7% within-one agreement) establishes shared calibration rather than independence from their bias. Deterministically from the retrieval logs, citation groundedness rose from 0% to 51–89% and retrieval recall@5 was 92%. On the judge, content-level hallucination fell from 42% to 4% and accuracy rose by a pooled +0.64 points (95% CI 0.47–0.80); the accuracy gain persisted after adjustment for response length (+0.48, 95% CI 0.22–0.73), which retrieval shortened rather than lengthened. The accuracy gain was large for weaker base models and small or non-significant for already-strong ones, whereas the hallucination and auditability gains were consistent across all models. The human ratings reproduced the judge’s accuracy effect (+0.61, 95% CI 0.49–0.74), and GPT-5 run through the transparent pipeline showed no significant difference from the proprietary Custom GPT (judge accuracy 4.48 vs. 4.58). Conclusions: Guideline retrieval yields a reproducible, largely base-independent improvement in the safety and auditability of LLM answers to clinical guideline questions, with accuracy gains concentrated in weaker base models. Because retrieval-augmented answers are recognizable to experts, rigorous evaluation should rely on rater-independent measures, and residual hallucination continues to require human oversight.
Andreas Vollmer, Lara Schorn, Felix Schrader et al.· Diagnostics· 0 citations
By decoupling clinical information retrieval from generative chitchat, LENOHA enhances safety, preserves privacy, and markedly reduces energy use, offering a practical blueprint for sustainable and equitable medical AI deployment across diverse care settings.
Motoki Sato, Sou Nagata, Mizuho Ohnuma et al.· JMIR Medical Informatics· 0 citations
Bottom-up incremental scoring showed the closest alignment with human assessment in clinical AI evaluation, underscoring the need for standardised prompt architectures in clinical AI evaluation.
Clinical practice guidelines require expert synthesis that large language models (LLMs) might partly automate, yet their ability to reproduce clinically actionable recommendations is poorly quantified. We evaluate an LLM (Claude Sonnet 4.6) against the 103 recommendations of the Spanish enhanced-recovery guideline Vía RICA 2026, grouped in 17 bundles. The model used the panel’s own closed corpus (617 documents) in a multilingual retrievalaugmented generation pipeline. Concordance was assessed twice: by optimal 1:1 bipartite matching (Hungarian) on cosine similarity, and by an LLM-as-a-judge clinical adjudicator (Claude Haiku 4.5) validated against a three-clinician panel (Fleiss’ κ = 0.538). The two schemes bracket a micro F1 of 0.61–0.69 and reveal four findings: (i) a systematic granularity bias, producing 1–8 recommendations per bundle regardless of ground-truth size; (ii) failure of cosine similarity to discriminate within narrow clinical domains; (iii) high reference-concordance precision (0.70–0.81) despite low exhaustiveness; and (iv) no transfer of the GRADE fields, evidence level agreeing no better than chance and strength systematically downgraded. An eight-fold larger retrieval budget left it intact. A corpus audit found 25 documents that formulate recommendations; excluding them lowers judged micro F1 to 0.602. The results delimit the current utility of generative AI for guideline development.
Andrea Moral, Antonio Arroyo, Juan Aparicio et al.· Machine Learning and Knowled...· 0 citations
The application of Large Language Models (LLMs) to diagnostic decision-making has garnered growing interest. However, existing benchmarks largely focus on textual reasoning or isolated visual question-answering (VQA) tasks, lacking holistic integration of clinical narratives and medical imaging, and thus failing to assess the multimodal diagnostic synthesis capability central to expert clinical judgment. To bridge this gap, we introduce MedReaMM, a benchmark specifically designed to evaluate models'ability to synthesize heterogeneous clinical evidence consisting of detailed patient histories alongside multiple medical images into accurate differential diagnoses under a complete-information paradigm. Constructed from case reports sourced from top-tier medical journals and curated clinical case databases, MedReaMM comprises 625 expert-validated cases with an average of 2.79 medical images per case and a total of 1,042 standardized diagnoses annotated with ICD-11 codes. These cases predominantly represent rare, atypical, or multi-system presentations that demand expert-level evidence integration beyond routine pattern recognition. We evaluate 23 Large Multimodal Models (LMMs) and find that most achieve diagnostic accuracy scores below 50%, underscoring a substantial gap in multimodal diagnostic synthesis capability. Further analysis reveals that medical knowledge proficiency, medical image understanding, and evidence integration are all highly correlated with diagnostic performance.
Lai Wei, Yuchao Chen, Zhenbiao Cao et al.· 0 citations
Patients with hallux valgus increasingly seek health information through consumer-facing artificial intelligence (AI)–driven patient education tools, particularly large language model–based conversational agents. Although these tools offer rapid and accessible responses, concerns remain regarding the reliability, usefulness, overall quality, and readability of AI-generated patient education materials. Evidence specifically evaluating AI-generated patient education for hallux valgus, a condition strongly influenced by patient expectations and treatment preferences, remains limited.
This cross-sectional comparative study evaluated the performance of three large language models—ChatGPT-4o, Gemini-2.5-Flash, and DeepSeek-V3—in responding to 20 patient-centered questions related to hallux valgus. Questions were developed using AI-assisted question generation and publicly available Google Trends search patterns and categorized into four clinical domains. AI-generated responses were anonymized and independently assessed by three orthopaedic surgeons for reliability, usefulness, and overall quality using 7-point Likert-based reliability and usefulness scales and the Global Quality Scale (GQS). Readability was analyzed using six standardized indices. Inter-rater agreement and between-model comparisons were statistically evaluated.
Gemini-2.5-Flash demonstrated modestly higher overall reliability, particularly in questions related to etiology and clinical presentation. DeepSeek-V3 achieved higher usefulness scores in the long-term outcomes and quality-of-life domain and produced significantly more readable content, as reflected by higher Flesch Reading Ease scores and lower grade-level indices. In contrast, Gemini-2.5-Flash generated linguistically more complex responses requiring higher educational levels for comprehension. Overall usefulness and global quality scores did not differ significantly among models. Qualitative review also identified occasional examples of oversimplified or potentially misleading information. Despite these differences, several readability metrics exceeded recommended patient health-literacy thresholds.
Contemporary AI-based conversational agents can provide patient-oriented information regarding hallux valgus with variable reliability and readability characteristics, although statistically significant differences were observed across models. A trade-off between factual accuracy and linguistic accessibility was observed. AI tools should therefore be regarded as adjuncts to, rather than replacements for, clinician-led patient education. Awareness of AI limitations and appropriate clinical guidance remain essential to ensure safe, accurate, and patient-centered information delivery.
Not applicable.
A. Koluman, Ebru Aloğlu Çiftçi, Mehmet Utku Çiftçi et al.· BMC Medical Informatics and...· 0 citations