The increasing volume of patient messages via electronic health record (EHR) portals has contributed significantly to clinician workload. Automating responses to these messages can help alleviate this burden, but it is essential to ensure that the generated responses are grounded in accurate clinical evidence. As part of the ArchEHR-QA 2025 BioNLP ACL shared task, we explore unsupervised methods for generating patient question responses that are both contextually accurate and evidence-backed. We investigate three novel approaches: zero-shot prompting, clustering-based evidence selection, and attention-based evidence attribution, along with a hybrid model that combines clustering and attention. Our methods do not require model fine-tuning and leverage the inherent structure of the input data to identify the most relevant supporting evidence from clinical notes. Our best-performing approach, which integrates clustering and attention, demonstrates a substantial improvement in factuality over baseline zero-shot methods, highlighting the potential of unsupervised strategies for enhancing the clinical utility of large language models in EHR contexts.

Large language models perform surprisingly well on many zero-shot classification tasks, but are difficult to fairly compare to supervised classifiers due to the lack of a modifiable decision boundary. In this work, we propose and evaluate a method that transforms binary classification tasks into pairwise comparisons between instances within a dataset, using LLMs to produce relative rankings of those instances. Repeated pairwise comparisons can be used to score instances using the Elo rating system (used in chess and other competitions), inducing a confidence ordering over instances in a dataset. We evaluate scheduling algorithms for their ability to minimize comparisons, and show that our proposed algorithm leads to improved classification performance, while also providing more information than traditional zero-shot classification.

The U.S. Food and Drug Administration (FDA) collects real-world adverse events, including device-associated deaths, injuries, and malfunctions, through passive reporting to the agency’s Manufacturer and User Facility Device Experience (MAUDE) database. However, this system’s full potential remains untapped given the extensive use of unstructured text in medical device adverse event reports and lack of FDA resources and expertise to properly analyze all available data. In this work, we focus on addressing this limitation through the development of an annotated benchmark corpus to support the design and development of state-of-the-art NLP approaches towards automatic extraction of device-related adverse event information from FDA Medical Device Adverse Event Reports. We develop a dataset of labeled medical device reports from a diverse set of high-risk device types, that can be used for supervised machine learning. We develop annotation guidelines and manually annotate for nine entity types. The resulting dataset contains 935 annotated adverse event reports, containing 12252 annotated spans across the nine entity types. The dataset developed in this work will be made publicly available upon publication.

The introduction of Large Language Models (LLMs) has advanced data representation and analysis, bringing significant progress in their use for medical questions and answering. Despite these advancements, integrating tabular data, especially numerical data pivotal in clinical contexts, into LLM paradigms has not been thoroughly explored. In this study, we examine the effectiveness of vector representations from last hidden states of LLMs for medical diagnostics and prognostics using electronic health record (EHR) data. We compare the performance of these embeddings with that of raw numerical EHR data when used as feature inputs to traditional machine learning (ML) algorithms that excel at tabular data learning, such as eXtreme Gradient Boosting. We focus on instruction-tuned LLMs in a zero-shot setting to represent abnormal physiological data and evaluating their utilities as feature extractors to enhance ML classifiers for predicting diagnoses, length of stay, and mortality. Furthermore, we examine prompt engineering techniques on zero-shot and few-shot LLM embeddings to measure their impact comprehensively. Although findings suggest the raw data features still prevail in medical ML tasks, zero-shot LLM embeddings demonstrate competitive results, suggesting a promising avenue for future research in medical applications.