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.

OBJECTIVES: The automation of medical report generation using large language models (LLMs) could significantly reduce physicians' documentation burden while enhancing healthcare efficiency. However, the misuse of generative artificial intelligence in medical reporting can lead to important safety risks for patients. We addressed 2 questions: (1) What is the quality of medical reports generated by LLMs in English and French? and (2) Can we distinguish between human-written and LLM-generated medical reports? MATERIALS AND METHODS: We evaluated the quality of reports generated by several multilingual, open-weight LLMs using text similarity metrics on 4212 medical reports in English and French across multiple specialties. A bilingual expert panel of certified physicians (n = 4) and medical residents (n = 5) scored accuracy, fluency, and completeness of generated reports using a 1-5 Likert scale. Experts also completed a Turing-like test, blindly identifying reports as human or machine-generated. RESULTS: Phi-4 achieved the best overall performance (ROUGE-1: 0.70, BERTScore: 0.83). Expert evaluation confirmed high-quality reports in both languages (overall 4.6/5.0). Medical experts performed better than chance but struggled to differentiate human versus machine reports (accuracy: 0.60). Automatic classifiers showed strong performance (accuracy: 0.98). DISCUSSION: The high quality of LLM-generated reports supports their potential to enhance healthcare efficiency in multilingual settings. However, the discrepancy between human detection difficulty and automated detection success reveals inherent limitations in relying solely on human oversight for quality assurance and misuse prevention. CONCLUSIONS: Deployment of LLMs for medical reporting requires combining automated detection tools with human expertise to ensure patient safety. Dataset and code: https://github.com/ds4dh/medical_report_generation.

Patient stratification-identifying clinically meaningful sub-groups-is essential for advancing personalized medicine through improved diagnostics and treatment strategies. Electronic health records (EHRs), particularly those from intensive care units (ICUs), contain rich temporal clinical data that can be leveraged for this purpose. In this work, we introduce ICU-TSB (Temporal Stratification Benchmark), the first comprehensive benchmark for evaluating patient stratification based on temporal patient representation learning using three publicly available ICU EHR datasets. A key contribution of our benchmark is a novel hierarchical evaluation framework utilizing disease taxonomies to measure the alignment of discovered clusters with clinically validated disease groupings. In our experiments with ICU-TSB, we compared statistical methods and several recurrent neural networks, including LSTM and GRU, for their ability to generate effective patient representations for subsequent clustering of patient trajectories. Our results demonstrate that temporal representation learning can rediscover clinically meaningful patient cohorts; nevertheless, it remains a challenging task, with v-measuring varying from up to 0.46 at the top level of the taxonomy to up to 0.40 at the lowest level. To further enhance the practical utility of our findings, we also evaluate multiple strategies for assigning interpretable labels to the identified clusters. The experiments and benchmark are fully reproducible and available at https://github.com/ds4dh/CBMS2025stratification.

Effective representation of medical concepts is crucial for secondary analyses of electronic health records. Neural language models have shown promise in automatically deriving medical concept representations from clinical data. However, the comparative performance of different language models for creating these empirical representations, and the extent to which they encode medical semantics, has not been extensively studied. This study aims to address this gap by evaluating the effectiveness of three popular language models - word2vec, fastText, and GloVe - in creating medical concept embeddings that capture their semantic meaning. By using a large dataset of digital health records, we created patient trajectories and used them to train the language models. We then assessed the ability of the learned embeddings to encode semantics through an explicit comparison with biomedical terminologies, and implicitly by predicting patient outcomes and trajectories with different levels of available information. Our qualitative analysis shows that empirical clusters of embeddings learned by fastText exhibit the highest similarity with theoretical clustering patterns obtained from biomedical terminologies, with a similarity score between empirical and theoretical clusters of 0.88, 0.80, and 0.92 for diagnosis, procedure, and medication codes, respectively. Conversely, for outcome prediction, word2vec and GloVe tend to outperform fastText, with the former achieving AUROC as high as 0.78, 0.62, and 0.85 for length-of-stay, readmission, and mortality prediction, respectively. In predicting medical codes in patient trajectories, GloVe achieves the highest performance for diagnosis and medication codes (AUPRC of 0.45 and of 0.81, respectively) at the highest level of the semantic hierarchy, while fastText outperforms the other models for procedure codes (AUPRC of 0.66). Our study demonstrates that subword information is crucial for learning medical concept representations, but global embedding vectors are better suited for more high-level downstream tasks, such as trajectory prediction. Thus, these models can be harnessed to learn representations that convey clinical meaning, and our insights highlight the potential of using machine learning techniques to semantically encode medical data.

OBJECTIVES: Clinical trials (CTs) are essential for improving patient care by evaluating new treatments' safety and efficacy. A key component in CT protocols is the study population defined by the eligibility criteria. This study aims to evaluate the effectiveness of large language models (LLMs) in encoding eligibility criterion information to support CT-protocol design. MATERIALS AND METHODS: We extracted eligibility criterion sections, phases, conditions, and interventions from CT protocols available in the ClinicalTrials.gov registry. Eligibility sections were split into individual rules using a criterion tokenizer and embedded using LLMs. The obtained representations were clustered. The quality and relevance of the clusters for protocol design was evaluated through 3 experiments: intrinsic alignment with protocol information and human expert cluster coherence assessment, extrinsic evaluation through CT-level classification tasks, and eligibility section generation. RESULTS: Sentence embeddings fine-tuned using biomedical corpora produce clusters with the highest alignment to CT-level information. Human expert evaluation confirms that clusters are well structured and coherent. Despite the high information compression, clusters retain significant CT information, up to 97% of the classification performance obtained with raw embeddings. Finally, eligibility sections automatically generated using clusters achieve 95% of the ROUGE scores obtained with a generative LLM prompted with CT-protocol details, suggesting that clusters encapsulate information useful to CT-protocol design. DISCUSSION: Clusters derived from sentence-level LLM embeddings effectively summarize complex eligibility criterion data while retaining relevant CT-protocol details. Clustering-based approaches provide a scalable enhancement in CT design that balances information compression with accuracy. CONCLUSIONS: Clustering eligibility criteria using LLM embeddings provides a practical and efficient method to summarize critical protocol information. We provide an interactive visualization of the pipeline here.

Several machine learning approaches have been proposed to automatically derive clinical phenotypes from patient data. Nevertheless, methods leveraging similarity-based patient networks remain underexplored for temporal data. In this work, we propose a graph neural network (GNN) model that learns patient representation using different network configurations and feature modes. To explore the sequential nature of time series, features were extracted using a recurrent neural network (RNN) and embedded using information from the network structure via the GNN. Our method improves upon statistical and RNN baselines, with performance boosts up to 1% and 22% accuracy in the inductive and transductive settings, respectively. We also show that network configurations significantly impact performance in the transductive learning setting. Thus, automated phenotyping models based on GNNs could be used to support phenotype-based clinical research and ultimately for personalized clinical decision support.Data and Code Availability: This paper uses the MIMIC-III dataset [1], which is available on the PhysioNet repository [2]. The experiments are based on the public open source phenotyping benchmark of Harutyunyan et al. [3]. All our source code is publicly available at https://github.com/ds4dh/mimic3-benchmarks-GraDSCI23.