Hua Xu

Objectives: Systematic evaluation of variability in large language model (LLM)-generated outputs is critical for assessing their reliability. We developed a general statistical framework and Python package (llmReproScore) for quantifying the repeatability and reproducibility of LLM outputs. Materials and Methods: Following US Food and Drug Administration terminology, our framework establishes four metrics for quantifying LLM repeatability and reproducibility along two dimensions: semantic and internal. Semantic metrics quantify variability in the meaning of LLM-generated outputs, whereas internal metrics quantify variability in token-generation certainty. We validated these metrics, and as an illustrative use case, applied them to diagnostic reasoning outputs generated across LLMs and prompting strategies using US Medical Licensing Examination questions and 90 real-world rare disease cases from the Undiagnosed Diseases Network. Results: Validation results showed that semantic scores aligned with human ratings (Spearman ρ = 0.88 for repeatability and 0.66 for reproducibility) and confirmed that internal metrics capture variability in token-generation certainty. In the illustrative use case, LLM repeatability was generally not associated with accuracy, suggesting that LLM output variability and accuracy capture distinct dimensions of model performance. Discussion: The proposed metrics can be used to characterize LLM output variability and systematically compare models, prompts, and generation settings. These comparisons can inform more rigorous selection of LLM configurations based on repeatability and reproducibility. Conclusion: We developed and validated generalizable metrics for quantifying LLM output variability. The proposed metrics can complement accuracy-based evaluation to inform experimental design and support more rigorous comparison and selection of models, prompts, and generation settings.

Phenotyping is fundamental to rare disease diagnosis, but manually curating structured phenotypes from clinical notes is labor intensive and difficult to scale. Existing artificial intelligence (AI) approaches optimize individual components of phenotyping but generally do not operationalize the full workflow of extracting features from clinical text, standardizing them to Human Phenotype Ontology (HPO) terms, and prioritizing rare disease-relevant HPO terms. We developed RARE-PHENIX, an end-to-end AI framework for rare disease phenotyping that integrates large language model-based phenotype extraction, ontology-grounded standardization to HPO terms, and supervised ranking of HPO terms. We trained RARE-PHENIX using data from 2671 patients across 11 Undiagnosed Diseases Network clinical sites, and externally validated it on 16,357 real-world clinical notes from 143 patients at Vanderbilt University Medical Center. Our findings suggest RARE-PHENIX generates structured, prioritized HPO terms highly concordant with clinician curation, demonstrates potential downstream disease-prioritization signal, and may support human-in-the-loop rare disease phenotyping in real-world settings.

Objectives: Accurate citation of relevant publications is essential for scientific integrity in biomedical research. Large language models (LLMs) excel at text generation but often hallucinate fabricated or inaccurate citations. Retrieval-augmented generation (RAG) can mitigate these errors, yet current approaches lack semantic precision in evidence retrieval. This study aims to develop a domain-specific RAG system for reliable, context-specific biomedical citation recommendations. Materials and Methods: We introduce CiteSure, a sentence-level citation recommendation tool designed to deliver reliable, evidence-based, and context-specific references using LLMs. CiteSure utilizes a 2-stage retrieval-augmented generation (RAG) framework, combining a domain-specific dense retriever (BioLLM2Vec) and reranker (BioRankLLaMA), adapted from LLaMA3-8B-Instruct using biomedical-specific training data. CiteSure leverages the complementary strengths of retrieval and generative LLM models, ensuring factual precision and contextual alignment. We evaluated CiteSure on a curated Alzheimer’s disease dataset, comparing it to standalone LLMs and traditional retrieval-based methods. Results: CiteSure achieved 100% factual accuracy and the highest relevance score of 77.50%, outperforming all baselines. BioLLM2Vec retrieved relevant articles with over 80% accuracy in the top 100 candidates. BioRankLLaMA consistently outperformed baseline rerankers across MAP, MRR, and Precision@5 metrics, confirming the benefit of domain-specific adaptation and contrastive fine-tuning. Discussion and Conclusion: Our results demonstrate that CiteSure, built on a 2-stage retrieval-augmented generation framework, effectively integrates domain-specific retrieval with LLM-based generation to achieve substantial improvements over baseline approaches. Our work underscores the importance of domain-specific adaptation in biomedical citation recommendation and provides publicly available datasets, models, and code for support future research.

Background: Distributed Research Networks (DRNs) offer significant opportunities for collaborative multi-site research and have significantly advanced healthcare research based on clinical observational data. However, generating high-quality real-world evidence using fit-for-use data from multi-site studies faces important challenges, including biases associated with various types of heterogeneity within and across sites and data sharing difficulties. Over the last 10 years, Privacy-Preserving Distributed Algorithms (PDA) have been developed and utilized in numerous national and international real-world studies spanning diverse domains, from comparative effectiveness research, target trial emulation, to healthcare delivery, policy evaluation, and system performance assessment. Despite these advances, there remains a lack of comprehensive and clear guiding principles for generating high-quality real-world evidence through collaborative studies leveraging the methods under PDA. Objective: The paper aims to establish 10 principles of best practice for conducting high-quality multi-site studies using PDA. These principles cover all phases of research, including study preparation, protocol development, analysis, and final reporting. Discussion: The 10 principles for conducting a PDA study outline a principled, efficient, and transparent framework for employing distributed learning algorithms within DRNs to generate reliable and reproducible real-world evidence.

Background: Social isolation and social support are clinically important but inconsistently represented in structured electronic health record data. Their identification from clinical narratives is complicated by nuanced, context-dependent language and by variation in how social connectedness is documented across clinical settings, patient populations, and geographic locations. Models intended for health care use must therefore distinguish social isolation from documented support, negation, ambiguous social circumstances, and non-social clinical uses of similar terminology while remaining robust to differences in documentation practices. Objective: This study aimed to develop and evaluate locally deployable, fine-tuned language models for context-aware classification of social isolation, social support, and contextually similar nonrelevant clinical text across multiple health systems. Materials and Methods: We conducted a retrospective study using electronic health record data from 326,847 adults aged 50 years or older who received care at 3 Tennessee health systems from 2020 to 2023. Candidate clinical text was identified using social-environment ICD-10-CM Z codes and an established social isolation lexicon. Human annotators classified 9,748 spans from 9,578 clinical notes contributed by 6,881 patients as social isolation, social support, or no social isolation. The annotated corpus included social isolation and support concepts including negations, infection-control and microbiology terminology, institutional living arrangements, and other contextually ambiguous examples. Uncertainty-based active learning was used to prioritize informative spans for annotation. FLAN-T5-Large, BERT, RoBERTa, and Gemma-2-2B were fine-tuned for 3-class span classification. Full fine-tuning and parameter-efficient fine-tuning using low-rank adaptation approaches were evaluated. Performance was assessed using site-based internal-external cross-validation, with macro-F1 as the primary metric. Results: The annotated corpus included 2,318 social isolation spans, 638 social support spans, and 6,792 spans classified as no social isolation. Fully fine-tuned FLAN-T5-Large achieved the highest and most consistent performance, with a mean macro-F1 of 0.92 (SD 0.04). Class-specific F1 scores were 0.91 (SD 0.03) for social isolation, 0.90 (SD 0.04) for social support, and 0.94 (SD 0.05) for no social isolation. Gemma-2-2B achieved a macro-F1 of 0.89 (SD 0.10), compared with 0.80 (SD 0.21) for RoBERTa and 0.77 (SD 0.17) for BERT. Full fine-tuning also outperformed parameter-efficient fine-tuning in the corresponding held-out evaluation. Remaining errors primarily involved patients who lived alone but received assistance, the presence of family members without meaningful support, grief-related language, and non-social clinical uses of isolation terminology. Conclusions: Fine-tuned language models can distinguish social isolation, social support, and contextually similar but unrelated language across heterogeneous clinical settings. FLAN-T5-Large provided the strongest and most stable performance, including for the less frequent social support class. The 3-class annotation framework, inclusion of difficult negative examples, and multisite evaluation offer a practical approach for converting nuanced social-connectedness documentation into structured data. Further work should evaluate end-to-end identification in unselected clinical notes and prospective integration into clinical and population health workflows.

Large Language Models (LLMs) have demonstrated significant potential in medicine, with many studies adapting them through continued pretraining or fine-tuning on medical data. However, a key question remains: to what extent do LLMs memorize medical training data—that is, recall or regenerate content seen during continued pretraining or fine-tuning. In this work, we investigate memorization of LLMs in medicine, assessing its prevalence (frequency), characteristics (what is memorized), volume (how much), and potential downstream impacts. We systematically analyze common adaptation scenarios: (1) continued pretraining on medical corpora, (2) fine-tuning on standard medical benchmarks, and (3) fine-tuning on real-world clinical data, including over 13,000 unique inpatient records from Yale New Haven Health System. The results demonstrate that memorization is prevalent and significantly higher than that in the general domain. Memorization has distinct characteristics during continued pretraining and fine-tuning, and it is persistent: up to 87% of content memorized during continued pretraining remains after fine-tuning. Memorization can be categorized into three types: beneficial (e.g., accurate recall of clinical guidelines), uninformative (e.g., templated language), and harmful (e.g., sensitive clinical content). We offer practical recommendations to facilitate beneficial memorization, minimize uninformative memorization, and mitigate harmful memorization to protect patient privacy and improve medical utility.

The BRAIN Initiative Cell Atlas Network (BICAN) is generating large-scale multimodal datasets to profile cell types in the human, non-human primate, and mouse brain. The diversity of single-cell and spatial transcriptomic and epigenomic assays, combined with varied experimental contexts, multiple data-generating laboratories and distributed infrastructure, poses substantial challenges for data integration and reuse in BICAN. To address this, we implemented a standards framework that enables layered integration of these data into knowledge-ready products for interoperable brain cell atlases. This framework organizes data based on three progressively structured layers. First, we introduced an assay-agnostic modeling layer that unifies the representation of single-cell and spatial omics data using a common set of biological entities and processes assessed by diverse experimental techniques. Second, we implemented harmonized metadata standards that capture key experimental features linked to biospecimen provenance across heterogeneous tissue sources, species, and preparations, supporting integration and validation while minimizing burden on data contributors. Third, we present an extensible representation for data-driven cell type taxonomies that integrates molecular data with annotations, ontology mappings, and evidence. Together, these contributions represent an end-to-end framework that transforms heterogeneous datasets into structured, interoperable resources that support broad community reuse via mapping algorithms, annotation systems, and visualization platforms. This approach links biospecimen provenance with cell-level outputs and embeds these in a standardized taxonomy format, enabling downstream applications such as cross-dataset integration, reference mapping, and knowledge-driven analysis. More broadly, our work demonstrates a generalizable strategy for enabling an efficient data-to-knowledge pipeline in a large-scale consortium setting.