Douglas Teodoro

Large language models (LLMs) in biomedicine face a fundamental conflict between static parameter knowledge and the dynamic nature of clinical evidence. Retrieval-augmented generation (RAG) addresses this by grounding generation in external data, yet it introduces new complexities in latency and architecture. This survey synthesizes the biomedical RAG landscape from 2020 to 2025 through a structured literature mapping of 178 screened records, yielding 94 representative studies classified into naive RAG (38 studies, 40.4%), advanced RAG (46 studies, 48.9%), and modular RAG (10 studies, 10.7%). Beyond a technological taxonomy, we formalize the biomedical RAG trilemma, identifying the inherent trade-offs between reasoning depth, inference latency, and data privacy that constrain current clinical deployment. We analyze how recent agentic workflows enhance diagnostic reasoning but risk prohibitive latency, and how privacy constraints dictate the choice between powerful cloud-based models and local deployment. Finally, we outline the alignment gap in multimodal RAG and propose future directions for self-correcting, verifiable clinical agents.

BACKGROUND Intensive care unit (ICU) electronic health records are sparse, irregularly sampled, and heterogeneous across hospitals, which limits how far patient representation-learning models transfer between centers. Benchmarks that evaluate such models under a common input schema and shared task definitions across multiple ICU databases remain scarce. OBJECTIVE Our objective is to determine whether Intensive Care Unit (ICU)-specific temporal representation-learning can generalize to support cross-cohort discovery and clinical risk prediction across hospitals. We use a temporal entity–attribute–value (t-EAV) input schema to train models under extreme sparsity, irregular sampling, and cross-center heterogeneity for supervised and unsupervised tasks. METHODS We harmonized 350,779 MIMIC-IV, eICU, SICdb, and HiRID stays, and evaluated cohort splits for 16 tasks extending ICU-TSB and YAIB benchmark task definitions. Using t-EAV inputs we evaluated statistical baselines and self-supervised models - recurrent (RNN) and graph neural networks trained from scratch, transformers fine-tuned from pretrained checkpoints and representations extracted zero-shot from general-purpose temporal foundation models. RESULTS No single method ranked highest across the 16 problems; the statistical heuristic (STAT) ranked highest on 6 supervised tasks on the mixed 24-hour cohort with AUROC 0.864 on mortality and 0.789 on sepsis prediction; tSW-GAT ranked highest on readmission (AUROC 0.675). Leiden graph community detection with the bidirectional LSTM reached the highest AMI (0.162) in mixed ICD-stratification. Under the tested zero-shot aggregation interface, generic-domain foundation models mostly trailed, with MOMENT the strongest of them in sepsis on the Fair COMMON test (AUROC 0.517, Appendix Table E3). Statistical heuristic and trained-from-scratch baselines ranked highest on most tasks. The held-out center was challenging for most supervised tasks. CONCLUSIONS General-purpose foundation models generally ranked below statistical heuristic and ICU-specific trained-from-scratch models. Several t-EAV baselines achieved higher multi-center performance than single-center cohorts, indicating task-dependent benefits.

Adverse drug events (ADEs) are a critical aspect of patient safety and pharmacovigilance, with significant implications for patient outcomes and public health monitoring. The increasing availability of electronic health records, social media, and patient forums provides valuable yet challenging unstructured data sources for ADE surveillance. To address these challenges, we introduce CONORM, a framework integrating named entity recognition (NER) and normalization (EN) for ADEs across diverse textual domains. CONORM comprises CONORM-NER and CONORM-EN, featuring a dual-encoder with dynamic context refining (DCR). DCR adaptively combines isolated entity embeddings with contextual representations. Our analyses demonstrate that CONORM adapts to text formality, enhances precision on out-of-distribution concepts, and reduces errors compared to context-agnostic baselines. CONORM was evaluated on tweets, forum posts, and structured product labels, covering 30,000 annotated ADE mentions across the SMM4H, CADEC, and TAC datasets. It achieved end-to-end F1-scores of 63.86%, 72.45%, and 84.99%, surpassing the strongest baseline by an average margin of 20%. These results highlight CONORM's robust adaptability across domains. CONORM offers a scalable solution for pharmacovigilance, with pre-computed target embeddings enhancing inference efficiency. Its generalization establishes it as a robust tool for ADE surveillance. The source code is available at https://github.com/ds4dh/CONORM .

Biomedical concept normalization (BCN)—mapping clinical terms to standardized concepts in biomedical terminologies—is a key bottleneck for semantic interoperability across health information systems, particularly in multilingual settings and when terms appear with minimal surrounding context, as is common in coded fields of real-world electronic health record databases. Large language models (LLMs) offer new opportunities to automate BCN, yet their behavior under these conditions remains poorly understood. This study introduces MedLexAlign, a multilingual benchmark comprising 52,011 unique terms mapped to 22,787 biomedical concepts across five European languages (English, French, German, Spanish, and Turkish). It proposes a knowledge-enhanced retrieve-then-rerank pipeline employing discriminative LLMs as dense retrievers and generative LLMs as rerankers, augmented with structured knowledge from the Unified Medical Language System (UMLS). We evaluate eight discriminative LLMs as dense retrievers and six generative LLMs (8B–32B parameters; across different reasoning modes) as rerankers, examining abbreviation handling, synonym sensitivity, and positional bias. The best-performing dense retriever, nemotron, achieved R@1 of 0.57 and R@10 of 0.80, outperforming kalm-gemma (R@1 = 0.56, R@10 = 0.78; p -value = 0.02). For reranking, UMLS-based knowledge enrichment yielded consistent gains: qwen3-32b-r achieved ΔR@1 of +0.10 with enrichment versus +0.05 without ( p -value < 0.001). However, LLMs exhibited sensitivity to lexical surface forms, handled synonymous representations inconsistently, and showed strong primacy bias toward earlier-ranked candidates. These findings demonstrate that structured knowledge enrichment is critical for effective LLM-based multilingual concept normalization, while surface-form sensitivity and positional biases remain important challenges for fully automated clinical pipelines.

This study conducts a clinical evaluation of a secure, locally deployed, quantized large language model (LLM) for automating modified Rankin Scale (mRS) score extraction from unstructured neurosurgical notes. We retrospectively selected 103 authentic clinical letters (2007–2025) from aneurysm patients at a tertiary neurosurgical centre. To comply with data privacy constraints, an open-source reasoning LLM (Qwen3-32B) with 4-bit quantization was deployed entirely on-premises. The LLM extracted mRS scores using a zero-shot approach with custom logits processors to enforce strict JSON formatting. Performance was compared to a reference standard (attending neurosurgeons’ consensus) and parallel scoring by medical residents and students. The LLM achieved excellent agreement with the attending consensus (QWK 0.95), matching the reliability of medical students (QWK 0.95) and residents (QWK 0.93). Exact agreement was 75%, and agreement within ±1 mRS point was 96%. Bayesian analysis strongly supported statistical equivalence between the model and human raters. The computationally optimized LLM demonstrated human-level classification reliability without task-specific fine-tuning. This approach successfully addresses key patient data privacy barriers and the formatting inconsistencies typical of open-ended generative models. Securely deploying a general-purpose, quantized LLM provides a scalable pathway for extracting functional outcomes and supports FAIR-aligned data systems.

BACKGROUND Generative artificial intelligence (GenAI) is increasingly used by medical students to summarize content While these tools may enhance efficiency, concerns have been raised that reliance on them could impede the development of summarization skills, which are essential for effective clinical communication and patient safety (de-skilling). OBJECTIVE To describe the development and evaluation of a GenAI-based online platform designed to train users in clinical case summarization and handoff communication through personalized feedback. METHODS The platform incorporates two expert-validated clinical vignettes, each with gold standard summaries for 4 target audiences (supervising physician, nurse, patient and night-shift physician). Personalized feedback is generated across three domains: content relevance, structure (based on the Situation, Background, Assessment, Recommendation [SBAR] framework), and language. Using a set of test summaries, multiple large language models (LLMs) were compared through a mixed-methods evaluation, yielding composite scores ranging from 7 (best performance) to 42 (worst performance). RESULTS Following two evaluation phases and iterative prompt optimization, Google Gemini 2.5 Pro emerged as the most reliable model, with a score of 11/42). The model consistently generated accurate, audience-appropriate feedback with few identified errors. CONCLUSIONS The GenAI based platform developed in this study, powered by Google Gemini 2.5 Pro, provides high-quality feedback on clinical summaries. These findings underscore its potential as an educational tool to develop clinical communication skills. Further research should assess its effectiveness in improving learners’ summarization performance in real‑world educational settings.

Objective: Information extraction (IE) from clinical texts has advanced rapidly with recent advances in natural language processing, particularly the advent of large language models (LLMs). However, inconsistent and incomplete reporting of methodologies limits reproducibility, comparability, and clinical translation. We aimed to develop a consensus-based reporting guideline tailored to clinical IE studies. Material and Methods: We developed the Clinical Information Extraction Reporting Guideline (CINEX) through a multi-phase process. The initiative was prospectively registered on the EQUATOR Network as a reporting guideline under development, and a detailed Delphi study protocol was published in advance. A scoping review informed an initial set of items, which was refined through a 3-round electronic Delphi study with 20 international experts, followed by a final consensus meeting. Items were iteratively refined based on predefined inclusion criteria and expert feedback. Results: The final CINEX guideline comprises 29 checklist items grouped into 5 domains: information model, architecture, data, annotation, and outcomes. The 3 eDelphi rounds included 20, 15, and 12 experts, respectively. Two items were added after round one. Consensus for inclusion was reached for a 21 and additional 7 items after rounds 2 and 3, respectively. Discussion: CINEX provides a structured framework to improve transparency, reproducibility, and interpretability in clinical IE research. By standardizing reporting of key methodological components, such as data provenance, annotation processes, and evaluation strategies, it facilitates meaningful comparison across studies and supports safer clinical implementation. Conclusion: CINEX complements existing AI reporting standards by addressing domain-specific challenges in clinical IE.

BACKGROUND Radiological reports are essential for clinical diagnosis. However, their preparation is laborious and time-consuming for radiologists. Artificial intelligence (AI) has the potential to help radiologists reduce this workload. OBJECTIVE This scoping review aims to map current research on AI-based radiological report generation, highlighting its applications, limitations, and future directions. METHODS We systematically retrieved 321 records from three major scientific databases (PubMed, Embase, Web of Science) and included 58 studies after screening. Benchmark datasets, radiological subspecialties, AI models, output formats, and evaluation methods were extracted and analyzed. RESULTS Nine public benchmark datasets and six radiological subspecialties were identified. Transformer-based architectures have become the dominant approach, with outputs primarily generated in natural language. Dataset diversity is limited, with most studies relying on chest X-ray images from two public datasets. Structured outputs for machine readability remain underexplored, and expert human evaluation is infrequently used as an assessment method. CONCLUSIONS Current AI research in radiological report generation is constrained by limited dataset diversity and evaluation practices. Future studies should expand imaging modalities and subspecialties, adopt unified vision-language Transformer models, generate more structured outputs for programmatic use, and incorporate expert human evaluation for quality assessment.

Background: Adverse drug events (ADEs) remain a critical safety issue in pharmaceutical research and development (Pharma R&D), necessitating robust methods for early detection and surveillance. Language models (LMs) are increasingly used in ADE analysis, addressing safety challenges during drug development and postmarket surveillance. Language modeling approaches, ranging from static embeddings to large language models (LLMs), capitalize on diverse data sources, such as clinical trial datasets, electronic health records, and social media posts, to predict ADEs, analyze real-world evidence, and improve drug screening and pharmacovigilance systems. Objective: This scoping review aims to map the application of LMs for the analysis of ADEs across the Pharma R&D lifecycle. Methods: Following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines, we searched PubMed, Web of Science, and Google Scholar for relevant papers published between January 2015 and October 2025. Results: This review identified 49 relevant papers. Overall, LM applications in Pharma R&D safety analysis are concentrated in 2 distinct phases: ADE prediction during the premarket phase (n=16) and ADE detection in postmarket surveillance (n=33). Conclusions: While some models demonstrate high predictive performance, persistent challenges, including data heterogeneity and limited external validation, hinder widespread adoption. Despite these barriers, discriminative and generative LMs have the potential to transform drug safety across the pre- and postapproval phases, especially when integrated with real-world pharmacovigilance frameworks.

Motivation: Biomedical named entity recognition (NER) presents unique challenges due to specialized vocabularies, the sheer volume of entities, and the continuous emergence of novel entities. Traditional NER models, constrained by fixed taxonomies and human annotations, struggle to generalize beyond predefined entity types. Results: To address these issues, we introduce GLiNER-BioMed, a domain-adapted suite of GLiNER models for biomedicine. Our approach first distills the annotation capabilities of large language models (LLMs) into a smaller, more efficient model, enabling the generation of high-coverage biomedical NER data. We subsequently train two GLiNER architectures, uni- and bi-encoder, at multiple scales to balance computational efficiency and performance. Experiments on eight biomedical datasets demonstrate that GLiNER-BioMed achieved state-of-the-art zero-shot performance (micro-F1 59.77%), exceeding the strongest baseline by 5.96 points (P < .001). In few-shot learning, the bi-encoder variant reached 70.39% (10-shot), consistently outperforming the strongest baseline across all settings (P < .05). Our findings show that the uni-encoder GLiNER-BioMed achieves the strongest zero-shot performance, while the bi-encoder offers superior few-shot gains and substantially higher inference throughput (+39%–568%), making it well-suited to annotation-limited, latency-sensitive, or large-label-space settings. Ablation studies further indicate that combining synthetic biomedical pre-training with general-domain post-training is essential for capturing domain-specific knowledge while maintaining precision-recall balance. Availability and implementation: The source code, datasets, and models are publicly available at https://github.com/ds4dh/GLiNER-biomed.