Citations

Abstract Heterogeneous treatment effects (HTE) are key to precision medicine, but most real-world studies lack the scale and diversity needed to detect them. While multi-site analyses offer a potential solution, data-sharing constraints often prevent access to patient-level information across institutions. We introduce POLARIS, a federated framework for time-to-event target trial emulation. POLARIS converts each site’s weighted Cox risk function into a compact tensor shared once with the coordinating center, enabling lossless reproduction of pooled estimates without sharing patient-level data. We applied POLARIS across five U.S. health systems to study risk of gastrointestinal outcomes after GLP-1 receptor agonist (GLP-1RAs) initiation versus sodium-glucose cotransporter 2 inhibitors (SGLT2is) and dipeptidyl peptidase 4 inhibitors (DPP4is). Results showed that GLP-1RAs were consistently associated with higher risks of nausea and vomiting, particularly among men, individuals with higher baseline HbA1c (≥ 8.5%), and lipid therapy. POLARIS provides a scalable solution for distributed target trial emulation and fine-grained assessment of HTE across diverse health systems.

Abstract Introduction While Large Language Models (LLMs) offer generative capabilities, generic models often lack domain-specific grounding, leading to hallucinations in high-stakes contexts. However, LLMs can serve as clinical decision support tools if augmented with specific domain data. This study evaluates the utility of Retrieval-Augmented Generation (RAG) applied to open-source LLMs, utilizing a curated knowledge base of authoritative sleep medicine textbooks, to quantify improvement in performance. Methods We evaluated four open-source models—Llama-70B, Llama-8B, Qwen-14B, and Qwen-235B, linked to a RAG system indexing five standard textbooks, including ICSD-3 TR. Models were tested on sleep board–style MCQs and clinical vignettes diagnostic accuracy. Performance was assessed across nine configurations derived from three factors: RAG pipeline complexity (Plain vs. Complicated), textbook preprocessing (Uncleaned vs. Cleaned/TOC-aligned), and retriever type (Dense-only vs. Hybrid). Metrics included MCQ Accuracy and the rates at which the correct diagnosis appeared as the top result or within the top 5 differential diagnosis lists for vignettes. Results RAG consistently outperformed the no-RAG baseline, yielding absolute performance gains of 5.6%–10.7% for MCQs and 8.1%–10.2% for the top diagnosis in cases across all models. For MCQs, Qwen-235B achieved the highest accuracy of 87.3% (vs. 81.7% baseline) using the Plain RAG configuration with Cleaned (TOC-aligned) Hybrid retrieval. Llama-70B followed with 83.6% in the same configuration. For case vignettes, Qwen-235B achieved a correct top diagnosis rate of 66.3% and a correct diagnosis rate within the top 5 differential diagnosis lists of 90.8% using Cleaned Hybrid retrieval configurations. Hybrid retrieval (Dense + Sparse) consistently surpassed Dense-only methods. Notably, while larger models benefited from Cleaned text, smaller models (e.g., Qwen-14B) achieved higher top diagnosis rates with Uncleaned text (57.1% vs. 51.0%), suggesting a dependency on redundant context in lower-parameter settings. Conclusion RAG significantly enhances knowledge performance of open-source LLMs in sleep medicine. Hybrid retrieval and curated, TOC-aligned knowledge bases yield optimal results for large models, whereas smaller models benefit from the redundancy of uncleaned text. These findings suggest that RAG systems can augment LLMs in domain specific knowledge for clinical decision support. Support (if any)

Study objectives To evaluate and compare the performance of two proprietary frontier large language models (LLMs), ChatGPT-5 and Grok-4, on diagnostic reasoning and foundational knowledge tasks within the specialty domain of sleep medicine. Methods The models were evaluated on two tasks: case-based reasoning using 79 clinical vignettes from the AASM Case Book of Sleep Medicine and knowledge assessment using 897 multiple-choice questions (MCQs) from board review materials. For vignettes, final diagnosis was scored by concept-level exact match, and differential diagnosis (DDx) was scored on a fixed top-5 output using concept-level matching with synonym normalization to compute precision, recall, and F 1-score. MCQ performance was the proportion correct. Inter-model performance was compared using the Mann–Whitney U test. Results Both models achieved high accuracy for final diagnosis (92.4% for both; 95% CI 86.4, 98.4) and MCQs (ChatGPT-5: 93.0%; Grok-4: 92.8%). However, performance on generating a comprehensive differential diagnosis was suboptimal, with modest F 1-scores for both ChatGPT-5 (0.55 ± 0.20) and Grok-4 (0.59 ± 0.20). There were no statistically significant differences in performance between the two models across any metric ( p > 0.05). Conclusions Frontier LLMs demonstrated high accuracy in sleep medicine tasks requiring knowledge recall and direct pattern recognition but showed more limited performance in complex clinical reasoning tasks such as generating a comprehensive differential diagnosis. These findings suggest that current general-purpose models may be more reliable for focused knowledge support than for broad hypothesis generation. Future studies should evaluate whether domain-adapted models or clinician-in-the-loop workflows can improve real-world diagnostic usefulness and safety.

Abstract Background: Accurate staging is essential for treatment selection, prognosis assessment, and trial eligibility. While prior studies attempted automating pathological (p) staging, few address clinical (c) staging and no studies have compared LLM-derived staging with clinician staging during clinical visit and retrospective cancer registry staging. Methods: We developed a multi-agent framework using Gemini-2.0-Flash-001 and GPT-5 to (1) identify reports (2) extract key data, and (3) apply AJCC 8th edition staging criteria. LLM staging was compared with clinician documentation and registry staging in breast cancer patients. Clinician-registry agreement served as the reference standard. Two independent breast oncologists reviewed discordant cases. Results: We analyzed 122 randomly selected breast cancer patients across all three Mayo Clinic sites from 2018 - 2023. LLM performance matched or exceeded human inter-rater agreement for pathological staging (95-99.2% vs. 95-98.4%). Clinical staging was more challenging, with LLM concordance of 73-77.9% for cT and 87-89.3% for cN versus 87.8% and 91.0% clinician-registry agreement. Among 89 patients with clinician-registry concordance, LLM concordance was 98.9% for pT/pN, 79.8% for cT, and 91.0% for cN. Experts favored LLM staging in 27.8% (5/18) of cT and 25.0% (2/8) of cN discordances. Error analysis revealed LLM challenges in handling non-mass enhancements (6 cases), selecting radiologic measurements (2 cases), identifying discrete masses (2 cases), and miscellaneous (3 cases). Conclusion: LLMs achieved human-level performance for pathological staging (98%), supporting human-in-the-loop deployment. For clinical staging (79.8% cT, 91.0% cN), future work must enhance multimodal reasoning and integrate physical examination data. Prospective validation is needed to assess real-world impact. With targeted refinements and oversight, automated staging can transform registry workflows while augmenting clinical decision-making. Citation Format: Arshad Mohammed, Umair Ayub, Pooja Advani, Shakeela W. Bahadur, Amye J. Tevaarwerk, Tufia C. Haddad, Elisabeth I. Heath, Brenda J. Ernst, Ben Zhou, Cui Tao, Sara J. Holton, Karthik V. Giridhar, Irbaz B. Riaz. Automating clinical and pathological staging for breast cancer patients [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 2744.

Abstract Objective Evaluate how RAG architecture, including corpus structure, retrieval strategy, and pipeline complexity, affects LLM-based medical problem solving and knowledge retrieval in sleep medicine. Materials and Methods We benchmarked four open-source LLMs (Llama-3-8B, Llama -3 -70B, Qwen 2.5-14B, and Qwen 2.5-235B) using a knowledge base of five sleep medicine textbooks. We compared performance across three dimensions: corpus structure (raw text vs table-of-contents aligned.), retrieval strategy (dense embedding vs hybrid sparse-dense), and pipeline complexity (baseline vs augmented). Evaluation metrics included board-style multiple choice question (MCQ) accuracy and clinical case vignette diagnostic ranking. Results RAG improved MCQ accuracy for all models. Llama-8B saw the largest gain of 10.6% (61.8% to 72.4%), while Qwen-235B reached 87.3%. In clinical cases, Llama-8B accuracy dropped by 7.1% when using raw text and dense retrieval due to context noise. This was corrected by using structured hybrid configurations. Hybrid retrieval consistently outperformed dense-only methods. Overall, structured corpora improved primary diagnosis accuracy by 6.1% on average, with Qwen-235B reaching a peak 10.2% increase. Discussion RAG effectiveness depends on the balance between model size and data structure. Large models handle uncurated text well, but smaller models are easily distracted by irrelevant data. Hybrid retrieval is necessary to maintain precision with specialized medical terms. A structured corpus paired with a baseline hybrid pipeline offers the best stability and speed for clinical use. Conclusion Rigorous data curation and hybrid retrieval are as essential as model scale for deploying safe, guideline-compliant AI in sleep medicine.

ABSTRACT Behavioral and social factors (BSFs) substantially influence the risk, onset, and progression of Alzheimer’s disease and related dementias (ADRD), yet BSF-related knowledge remains scattered across heterogeneous sources, posing critical challenges for data harmonization and evidence synthesis. To address this gap, we present the B ehavioral S ocial Data and Knowledge O ntology for AD RD (BSO-AD), a FAIR-compliant semantic resource for representing and harmonizing BSFs and ADRD-related knowledge. BSO-AD was developed following established ontology design principles with reuse of existing ontologies and controlled terminologies, including the Social Determinants of Health Ontology, Drug Repurposing-Oriented Alzheimer’s Disease Ontology, AD-Onto, ICD-9-CM, and ICD-10-CM. Relationships between BSFs and ADRD were derived through literature mining. BSO-AD contains 1,690 classes, 152 object properties, 49 data properties, and 40 annotation properties. Ontology evaluation through Hootation-based domain expert review and a scalable LLM-assisted ontology assessment framework demonstrated high domain coverage and strong semantic coherence. BSO-AD provides a semantic foundation for data harmonization and knowledge integration in BSF-related ADRD research, while the LLM-assisted evaluation framework supports scalable and automated ontology assessment.

Abstract Models that estimate the probability of an adverse clinical outcome require an operational cutoff to translate continuous estimated probabilities into discrete labels that can trigger clinical action. Although statistical methods identify ‘optimal’ cut-offs, threshold selection ultimately reflects value judgments regarding harm tolerance, resource allocation, and workflow feasibility. We describe a governance-informed approach to selecting a deployment threshold for an asthma exacerbation (AE) prediction model integrated into clinical workflows. Using prevalence-adjusted performance metrics and real-world provider capacity modeling, we evaluated multiple candidate thresholds and quantified downstream workload and missed-event trade-offs. We demonstrate that statistically optimal thresholds may produce operationally infeasible alert volumes or unacceptable miss rates. We propose a structured threshold governance framework integrating statistical performance, clinical utility, stakeholder input, and human oversight safeguards. This case illustrates how threshold decisions should be treated as organizational governance processes rather than purely technical optimizations.