Sohn, Sunghwan

Abstract Background Large language models (LLMs) are increasingly used in medication-related tasks despite limited evidence supporting their accuracy and safety. LLMs are vulnerable to adversarial attacks and may confabulate in response to fabricated information in the prompt. Given the substantial risk of harm from medication errors, the purpose of this study was to establish a benchmark task to evaluate LLM performance in distinguishing fabricated from real medications. Methods Two datasets (brand and generic), each consisting of 250 medication lists were developed, each including 4-6 medications and one fabricated medication (specifically a Pokémon character), with complete dosing information provided. Six LLMs were evaluated (GPT-5-Chat, GPT-4o-mini, DrugGPT, Gemma-3-27B-IT, Llama-3.3-70B-Instruct, and Qwen3-32B). For each medication list, models were queried for dosing information and disease indication across three experimental conditions: standard decoding, standard decoding with mitigation prompting, and deterministic decoding (temperature = 0). Each experiment was performed in triplicate. Outputs were deemed confabulations if the LLM failed to recognize and reject the fabricated medication. The primary outcome was the rate of confabulations. Exact paired-permutation tests were used to compare confabulation rates across LLMs and prompting approaches. Results Confabulation rates under standard conditions ranged from 2.7% to 99.6%. DrugGPT demonstrated the lowest baseline confabulation rates (2.7-6.4%), whereas other models demonstrated substantially higher rates (47.2-99.6%). Mitigation prompting significantly reduced confabulation rates compared to standard and deterministic decoding across all tasks and datasets (p < 0.001): GPT-5-Chat achieved confabulation rates of 0% with mitigation prompting from baseline rates of 66.4-88.5%. Deterministic decoding did not substantially impact confabulation rates. Conclusions LLMs performing medication-related tasks are susceptible to confabulation in response to invalid or fabricated inputs, raising important safety concerns for clinical use. While mitigation prompting reduced confabulation rates, this study underscores the importance of robust safeguards when deploying LLMs in medication-related decision-making.

Abstract Delirium is a state of acute confusion characterized by deficits in attention that occurs commonly in hospitalized older adults and is associated with adverse outcomes. Delirium may have significant implications for cognitive health, however the relationship of delirium to mild cognitive impairment (MCI) and dementia is not fully understood. To evaluate the impact of delirium on cognitive health trajectories, we leveraged the Mayo Clinic Study of Aging (MCSA), a population-based cohort study of cognitive health and aging in residents of Olmsted County, Minnesota, who are between the ages of 30 and 89 and without a previous diagnosis of dementia at enrollment. Emergency Department and hospital encounters were identified for all patients enrolled in the MCSA, and clinical notes following each encounter up to 30 days were analyzed using a previously validated natural language processing (NLP) algorithm for delirium detection. The association between delirium within 30 days of hospitalization and the development of MCI or dementia was assessed using Cox proportional hazards regression, adjusted for patient age, sex, and level of education. Initial MCSA assessment indicated normal cognition for all individuals in this cohort. We found that delirium was significantly associated with development of MCI and dementia, with 42 individuals experiencing delirium within 30 days of hospitalization; hazard ratio of 10.07 (95% CI 4.14, 24.49) for developing MCI and 19.10 (95% CI 8.40, 43.47) for dementia compared to those without delirium. Our results suggest that hospitalization with an episode of delirium can significantly impact cognitive trajectories in older adults.

Abstract Background: For large language models (LLMs) to reach their potential as information technology tools that make medication use safer, clinically relevant benchmarks capable of automated grading and designed specifically to measure the performance of LLMs for medication tasks are required. The purpose of this study was to design a suite of benchmarking tests reflective of Comprehensive Medication Management (CMM; the standard of care for medication optimization) and quantify the baseline performance of the latest LLMs. Methods: We established six benchmarks representing critical stages of the CMM process: drug formulation matching, drug order (sig) generation, drug route matching, drug-drug interaction identification, renal dose identification, and drug-indication matching. For each benchmark, we curated a clinician-annotated dataset comprising 250 standardized input-output pairs including both inpatient and outpatient medications. We evaluated the clinical knowledge retrieval capabilities of three LLMs: GPT-4o-mini, MedGemma-27B, and LLaMA3-70B. We employed a zero-shot prompting strategy, excluding in-context examples, to assess the models’ internal clinical knowledge rather than their few-shot learning potential. To check reliability, each model was run three times using a temperature of 0.7 (a mid-range value of an LLM setting controlling text generation randomness). Performance was assessed using task-specific evaluation metrics including precision (positive predictive value), recall (sensitivity), F1-score, accuracy, and correctness consistency across trials. Results: Across six benchmarks, LLaMA3-70B demonstrated the highest performance in four tasks: drug-formulation matching (F1, 54.0% [95 CI: 50.1-58]), drug-order generation (accuracy, 88.0%), drug-route identification (F1, 74.3% [95 CI: 71-78]), and drug-indication identification (accuracy, 97.6% [95 CI: 95.6-99.2]). In the drug–drug interaction task, GPT-4o-mini achieved the highest overall accuracy (70.4% [95 CI: 64.8-75.7]). For renal dose–adjustment identification, GPT-4o-mini demonstrated the highest F1 score (83.3% [95 CI: 77.6-88]). Correctness-consistency scores ranged from 8.0% to 97.6% across benchmarks, with no model exhibiting uniformly superior consistency. Conclusions: Model performance varied substantially across medication-related tasks. LLaMA3-70B demonstrated promising baseline performance in tasks involving formulation, ordering, route, and indication. GPT-4o-mini showed potential advantages in drug–drug interaction detection and renal dose adjustment. These findings underscore the need for task-specific evaluation when deploying models for medication-focused clinical decision support.

Abstract Background The purpose of this study was to develop a series of benchmarks specifically for large language model (LLM) performance in the identification of drug-drug interactions (DDIs). Methods We evaluated six LLMs using a clinician-designed dataset of 750 DDI scenarios. Tasks included: (1) a pointwise two-drug classification task (2) a pairwise three-drug discrimination task and (3) a listwise 4–6 drug selection task. Zero-shot prompting was applied and performance was assessed using precision, recall, F1 score, accuracy, and self-consistency. Results Model performance varied by task structure and interaction severity. GPT-5-Chat demonstrated the highest recall and F1 score in the pointwise task (80.7% and 80.2%), whereas DrugGPT achieved superior accuracy in the pairwise (93.7%) and listwise task (92.6%). All models had improved performance with more severe compared to less severe DDIs (accuracy 72.0-99.2% vs. 58.3-86.7%). All models tended to have higher self-consistency in the pairwise and listwise experiments compared to the pointwise experiment (68.8-93.6% vs. 31.0-74.2%). No model exhibited uniformly high reliability across all tasks. Conclusions Current LLMs show inconsistent capabilities in identifying DDIs. Performance degrades as reasoning space expands with limited stability. These findings emphasize the need for multi-format evaluation frameworks and reliability-aware assessment when considering LLMs for medication-safety applications.

Abstract Background Delirium is often underdiagnosed in clinical practice and is not routinely coded for billing. While manual chart review can identify delirium, it is labor-intensive and impractical for large-scale studies. Natural language processing (NLP) can analyze unstructured text in electronic health records (EHRs) to extract meaningful clinical information. Methods To support national integration of NLP for EHR-based delirium identification across different institutions, we launched the Delirium Interest Group within the national Evolve to Next-Gen Accrual to Clinical Trials (ENACT) NLP Working Group. This paper outlines our initial efforts to standardize, evaluate, and translate an NLP-based delirium detection model into the i2b2/ENACT platform. Results Multisite contextual inquiry identified several key challenges, including variations in local screening practices (eg, tools used, documentation frequency, and quality control), the need for harmonized definitions in the context of EHRs, and the complexity of modeling temporal logic. Multisite NLP evaluation revealed variable performance degradation driven by differences in delirium screening practices, clinical documentation patterns and semantics, and note syntactic structures. Conclusion Our work represents an important first step toward enabling scalable and standardized NLP-based delirium detection across institutions. By engaging diverse institutions through the ENACT NLP Working Group, we identified shared challenges and site-specific variations that impact model implementation and performance. Our collaborative approach enabled the development of a more robust framework for delirium identification across heterogeneous EHR systems. Future efforts will build on this foundation to enhance the validity, usability, and translational impact of delirium detection.

Abstract Objective: Electronic Health Record (EHR) data are critical for advancing translational research and AI technologies. The ENACT network offers access to structured EHR data across 57 CTSA hubs. However, substantial information is contained in clinical narratives, requiring natural language processing (NLP) for research. The ENACT NLP Working Group was formed to make NLP-derived clinical information accessible and queryable across the network. Methods: We established the ENACT NLP Working Group with 13 sites selected based on criteria including clinical notes access, IT infrastructure, NLP expertise, and institutional support. We divided sites into five focus groups targeting clinical tasks within disease contexts. Each focus group consisted of two development sites and two validation sites. We extended the ENACT ontology to standardize NLP-derived data and conducted multisite evaluations using the Open Health Natural Language Processing (OHNLP) Toolkit. Results: The working group achieved 100% site retention and deployed NLP infrastructure across all sites. We developed and validated NLP algorithms for rare disease phenotyping, social determinants of health, opioid use disorder, sleep phenotyping, and delirium phenotyping. Performance varied across sites (F1 scores 0.53–0.96), highlighting data heterogeneity impacts. We extended the ENACT common data model and ontology to incorporate NLP-derived data while maintaining Shared Health Research Informatics NEtwork (SHRINE) compatibility. Conclusion: This demonstrates feasibility of deploying NLP infrastructure across large, federated networks. The focus group approach proved more practical than general-purpose approaches. Key lessons include the challenge of data heterogeneity and importance of collaborative governance. This work also provides a foundation that other networks can build on to implement NLP capabilities for translational research.

Abstract With global aging, assessing functional status is vital for precision medicine. Electronic Health Records (EHRs), particularly unstructured data, hold abundant information on patient mobility. This study explores using Large Language Models (LLMs) to extract and standardize mobility status from unstructured EHR data (i.e., clinical notes). We annotated 600 clinical notes from three health care institutions located in southeastern Minnesota and west-central Wisconsin, focusing on expressions of mobility and associated impairment. Leveraging the open-source Llama 3 model, we tested various prompting strategies—including zero-shot, few-shot, and task decomposition—and evaluated their performance. Error analysis showed that while the model sometimes inferred impairments without explicit evidence, most errors were clinically reasonable, often reflecting borderline or ambiguous cases. While considering reasonable inference as correct, at the patient-level, Mobility Extraction achieves a micro-average accuracy of 0.952 with an F1-score of 0.962, and Impairment Classification produces a micro-average accuracy of 0.912 and an F1-score of 0.948. A local, deterministic setup improved trustworthiness by ensuring consistent outputs, safeguarding privacy, and demonstrating cross-institution generalizability. These findings highlight the feasibility of LLM-based solutions for extracting mobility functional status from unstructured EHR data, supporting both clinical applications and research.

Abstract Objectives Heart failure (HF) impacts millions of patients worldwide, yet the variability in treatment responses remains a major challenge for healthcare professionals. The current treatment strategies, largely derived from population based evidence, often fail to consider the unique characteristics of individual patients, resulting in suboptimal outcomes. This study aims to develop computational models that are patient-specific in predicting treatment outcomes, by utilizing a large Electronic Health Records (EHR) database. The goal is to improve drug response predictions by identifying specific HF patient subgroups that are likely to benefit from existing HF medications. Materials and Methods A novel, graph-based model capable of predicting treatment responses, combining Graph Neural Network and Transformer was developed. This method differs from conventional approaches by transforming a patient's EHR data into a graph structure. By defining patient subgroups based on this representation via K-Means Clustering, we were able to enhance the performance of drug response predictions. Results Leveraging EHR data from 11 627 Mayo Clinic HF patients, our model significantly outperformed traditional models in predicting drug response using NT-proBNP as a HF biomarker across five medication categories (best RMSE of 0.0043). Four distinct patient subgroups were identified with differential characteristics and outcomes, demonstrating superior predictive capabilities over existing HF subtypes (best mean RMSE of 0.0032). Discussion These results highlight the power of graph-based modeling of EHR in improving HF treatment strategies. The stratification of patients sheds light on particular patient segments that could benefit more significantly from tailored response predictions. Conclusions Longitudinal EHR data have the potential to enhance personalized prognostic predictions through the application of graph-based AI techniques.

Abstract Background Error analysis plays a crucial role in clinical concept extraction, a fundamental subtask within clinical natural language processing (NLP). The process typically involves a manual review of error types, such as contextual and linguistic factors contributing to their occurrence, and the identification of underlying causes to refine the NLP model and improve its performance. Conducting error analysis can be complex, requiring a combination of NLP expertise and domain-specific knowledge. Due to the high heterogeneity of electronic health record (EHR) settings across different institutions, challenges may arise when attempting to standardize and reproduce the error analysis process. Objectives This study aims to facilitate a collaborative effort to establish common definitions and taxonomies for capturing diverse error types, fostering community consensus on error analysis for clinical concept extraction tasks. Materials and Methods We iteratively developed and evaluated an error taxonomy based on existing literature, standards, real-world data, multisite case evaluations, and community feedback. The finalized taxonomy was released in both .dtd and .owl formats at the Open Health Natural Language Processing Consortium. The taxonomy is compatible with several different open-source annotation tools, including MAE, Brat, and MedTator. Results The resulting error taxonomy comprises 43 distinct error classes, organized into 6 error dimensions and 4 properties, including model type (symbolic and statistical machine learning), evaluation subject (model and human), evaluation level (patient, document, sentence, and concept), and annotation examples. Internal and external evaluations revealed strong variations in error types across methodological approaches, tasks, and EHR settings. Key points emerged from community feedback, including the need to enhancing clarity, generalizability, and usability of the taxonomy, along with dissemination strategies. Conclusion The proposed taxonomy can facilitate the acceleration and standardization of the error analysis process in multi-site settings, thus improving the provenance, interpretability, and portability of NLP models. Future researchers could explore the potential direction of developing automated or semi-automated methods to assist in the classification and standardization of error analysis.