Journal of the American Medical Informatics Association最新文献

Background: The use of generative large language models (LLMs) with electronic health record (EHR) data is rapidly expanding to support clinical and research tasks. This systematic review characterizes the clinical fields and use cases that have been studied and evaluated to date.

Methods: We followed the Preferred Reporting Items for Systematic Review and Meta-Analyses guidelines to conduct a systematic review of articles from PubMed and Web of Science published between January 1, 2023, and November 9, 2024. Studies were included if they used generative LLMs to analyze real-world EHR data and reported quantitative performance evaluations. Through data extraction, we identified clinical specialties and tasks for each included article, and summarized evaluation methods.

Results: Of the 18 735 articles retrieved, 196 met our criteria. Most studies focused on radiology (26.0%), oncology (10.7%), and emergency medicine (6.6%). Regarding clinical tasks, clinical decision support made up the largest proportion of studies (62.2%), while summarizations and patient communications made up the smallest, at 5.6% and 5.1%, respectively. In addition, GPT-4 and GPT-3.5 were the most commonly used generative LLMs, appearing in 60.2% and 57.7% of studies, respectively. Across these studies, we identified 22 unique non-NLP metrics and 35 unique NLP metrics. While NLP metrics offer greater scalability, none demonstrated a strong correlation with gold-standard human evaluations.

Conclusion: Our findings highlight the need to evaluate generative LLMs on EHR data across a broader range of clinical specialties and tasks, as well as the urgent need for standardized, scalable, and clinically meaningful evaluation frameworks.

Objective: To characterize the nature and consequence(s) of interdependent physician electronic health record (EHR) work across inpatient shifts.

Materials and methods: Pooled cross-sectional analysis of EHR metadata associated with hospital medicine patients at an academic medical center, January-June 2022. Using patient-day observation data, we use a mixed effects regression model with daytime physician random effects to examine nightshift behavior (handoff time, total EHR time) as a function of behaviors by the preceding daytime team. We also assess whether nighttime patient deterioration is predicted by team coordination behaviors across shifts.

Results: We observed 19 671 patient days (N = 2708 encounters). Physicians used the handoff tool consistently, generally spending 8-12 minutes per shift editing patient information. When the day service team was more activated (highest tercile of handoff time, overall EHR time), nightshift experienced increased levels of EHR work and patient risk of overnight decline was elevated. (ie, Busy predicts busy). However, lower levels of dayshift activation were also associated with nightshift spillovers, including higher overnight EHR work and increased likelihood of patient clinical decline. Patient-days in the lowest and highest terciles of dayshift EHR time had a 1 percentage point increased relative risk of overnight decline (baseline prevalence of 4.4%) compared to the middle tercile (P = .04).

Discussion: We find evidence of spillovers in EHR work from dayshift to nightshift. Additionally, the lowest and highest levels of dayshift EHR activity are associated with increased risk of overnight patient decline. Results are associational and motivate further examination of additional confounding factors.

Conclusion: Analyses reveal opportunities to address task interdependence across shifts, using technology to flexibly shape and support collaborative teaming practices in complex clinical environments.

Objective: Accurate phenotyping is an essential task for researchers utilizing electronic health record (EHR)-linked biobank programs like the All of Us Research Program to study human genetics. However, little guidance is available on how to select an EHR-based phenotyping procedure that maximizes downstream statistical power. This study aims to estimate accuracy of three phenotype definitions of ovarian, female breast, and colorectal cancers in All of Us (v7 release) and determine which is most likely to optimize downstream statistical power for genetic association testing.

Materials and methods: We used empirical carrier frequencies of deleterious variants in known risk genes to estimate the accuracy of each phenotype definition and compute statistical power after accounting for the probability of outcome misclassification.

Results: We found that the choice of phenotype definition can have a substantial impact on statistical power for association testing and that no approach was optimal across all tested diseases. The impact on power was particularly acute for rarer diseases and target risk alleles of moderate penetrance or low frequency. Additionally, our results suggest that the accuracy of higher-complexity phenotyping algorithms is inconsistent across Black and non-Hispanic White participants in All of Us, highlighting the potential for case ascertainment biases to impact downstream association testing.

Discussion: EHR-based phenotyping presents a bottleneck for maximizing power to detect novel risk alleles in All of Us, as well as a potential source of differential outcome misclassification that researchers should be aware of. We discuss the implications of this as well as potential mitigation strategies.

Using 2023-2024 U.S. National Health Interview Survey data, we found that digital health literacy (dHL) mediated nearly half of the difference in telehealth use between Latino adults with non-English and English language preference. These findings identify dHL as a modifiable mechanism linking linguistic and digital access barriers, underscoring the need for multilingual, inclusive, and equitable telehealth design.

Objective: Healthcare decisions are increasingly made with the assistance of machine learning (ML). ML has been known to have unfairness-inconsistent outcomes across subpopulations. Clinicians interacting with these systems can perpetuate such unfairness by overreliance. Recent work exploring ML suppression-silencing predictions based on auditing the ML-shows promise in mitigating performance issues originating from overreliance. This study aims to evaluate the impact of suppression on collaboration fairness and evaluate ML uncertainty as desiderata to audit the ML.

Materials and methods: We used data from the Vanderbilt University Medical Center electronic health record (n = 58 817) and the MIMIC-IV-ED dataset (n = 363 145) to predict likelihood of death or intensive care unit transfer and likelihood of 30-day readmission using gradient-boosted trees and an artificially high-performing oracle model. We derived clinician decisions directly from the dataset and simulated clinician acceptance of ML predictions based on previous empirical work on acceptance of clinical decision support alerts. We measured performance as area under the receiver operating characteristic curve and algorithmic fairness using absolute averaged odds difference.

Results: When the ML outperforms humans, suppression outperforms the human alone (P < 8.2 × 10-6) and at least does not degrade fairness. When the human outperforms the ML, the human is either fairer than suppression (P < 8.2 × 10-4) or there is no statistically significant difference in fairness. Incorporating uncertainty quantification into suppression approaches can improve performance.

Conclusion: Suppression of poor-quality ML predictions through an auditor model shows promise in improving collaborative human-AI performance and fairness.

Background: Despite rapid integration into clinical decision-making, clinical large language models (LLMs) face substantial translational barriers due to insufficient structural characterization and limited external validation.

Objective: We systematically map the clinical LLM research landscape to identify key structural patterns influencing their readiness for real-world clinical deployment.

Methods: We identified 73 clinical LLM studies published between January 2020 and March 2025 using a structured evidence-mapping approach. To ensure transparency and reproducibility in study selection, we followed key principles from the PRISMA 2020 framework. Each study was categorized by clinical task, base architecture, alignment strategy, data type, language, study design, validation methods, and evaluation metrics.

Results: Studies often addressed multiple early stage clinical tasks-question answering (56.2%), knowledge structuring (31.5%), and disease prediction (43.8%)-primarily using text data (52.1%) and English-language resources (80.8%). GPT models favored retrieval-augmented generation (43.8%), and LLaMA models consistently adopted multistage pretraining and fine-tuning strategies. Only 6.9% of studies included external validation, and prospective designs were observed in just 4.1% of cases, reflecting significant gaps in translational reliability. Evaluations were predominantly quantitative only (79.5%), though qualitative and mixed-method approaches are increasingly recognized for assessing clinical usability and trustworthiness.

Conclusion: Clinical LLM research remains exploratory, marked by limited generalizability across languages, data types, and clinical environments. To bridge this gap, future studies must prioritize multilingual and multimodal training, prospective study designs with rigorous external validation, and hybrid evaluation frameworks combining quantitative performance with qualitative clinical usability metrics.

Objectives: To assess the performance, generalizability, and computational efficiency of instruction-tuned Large Language Model Meta AI (LLaMA)-2 and LLaMA-3 models compared to bidirectional encoder representations from transformers (BERT) for clinical information extraction (IE) tasks, specifically named entity recognition (NER) and relation extraction (RE).

Materials and methods: We developed a comprehensive annotated corpus of 1588 clinical notes from 4 data sources-UT Physicians (UTP) (1342 notes), Transcribed Medical Transcription Sample Reports and Examples (MTSamples) (146), Medical Information Mart for Intensive Care (MIMIC)-III (50), and Informatics for Integrating Biology and the Bedside (i2b2) (50), capturing 4 clinical entities (problems, tests, medications, other treatments) and 16 modifiers (eg, negation, certainty). Large Language Model Meta AI-2 and LLaMA-3 were instruction-tuned for clinical NER and RE, and their performance was benchmarked against BERT.

Results: Large Language Model Meta AI models consistently outperformed BERT across datasets. In data-rich settings (eg, UTP), LLaMA achieved marginal gains (approximately 1% improvement for NER and 1.5%-3.7% for RE). Under limited data conditions (eg, MTSamples, MIMIC-III) and on the unseen i2b2 dataset, LLaMA-3-70B improved F1 scores by over 7% for NER and 4% for RE. However, performance gains came with increased computational costs, with LLaMA models requiring more memory and Graphics Processing Unit (GPU) hours and running up to 28 times slower than BERT.

Discussion: While LLaMA models offer enhanced performance, their higher computational demands and slower throughput highlight the need to balance performance with practical resource constraints. Application-specific considerations are essential when choosing between LLMs and BERT for clinical IE.

Conclusion: Instruction-tuned LLaMA models show promise for clinical NER and RE tasks. However, the tradeoff between improved performance and increased computational cost must be carefully evaluated. We release our Kiwi package (https://kiwi.clinicalnlp.org/) to facilitate the application of both LLaMA and BERT models in clinical IE applications.

Background: Chatbots are increasingly used to deliver health education, patient engagement, and access to healthcare services. GARDE-Chat is an open-source platform designed to facilitate the development, deployment, and dissemination of chatbot-based digital health interventions across different domains and settings.

Materials and methods: GARDE-Chat was developed through an iterative process informed by real-world use cases to guide prioritization of key features. The tool was developed as an open-source platform to promote collaboration, broad dissemination, and impact across research and clinical domains.

Results: GARDE-Chat's main features include (1) a visual authoring interface that allows non-programmers to design chatbots; (2) support for scripted, large language model (LLM)-based and hybrid chatbots; (3) capacity to share chatbots with researchers and institutions; (4) integration with external applications and data sources such as electronic health records and REDCap; (5) delivery via web browsers or text messaging; and (6) detailed audit log supporting analyses of chatbot user interactions. Since its first release in July 2022, GARDE-Chat has supported the development of chatbot-based interventions tested in multiple studies, including large pragmatic clinical trials addressing topics such as genetic testing, COVID-19 testing, tobacco cessation, and cancer screening.

Discussion: Ongoing challenges include the effort required for developing chatbot scripts, ensuring safe use of LLMs, and integrating with clinical systems.

Conclusion: GARDE-Chat is a generalizable platform for creating, implementing, and disseminating scalable chatbot-based population health interventions. It has been validated in several studies, and it is available to researchers and healthcare systems through an open-source mechanism.

Objective: Extracting social determinants of health (SDoHs) from medical notes depends heavily on labor-intensive annotations, which are typically task-specific, hampering reusability and limiting sharing. Here, we introduce SDoH-GPT, a novel framework leveraging few-shot learning large language models (LLMs) to automate the extraction of SDoH from unstructured text, aiming to improve both efficiency and generalizability.

Materials and methods: SDoH-GPT is a framework including the few-shot learning LLM methods to extract the SDoH from medical notes and the XGBoost classifiers which continue to classify SDoH using the annotations generated by the few-shot learning LLM methods as training datasets. The unique combination of the few-shot learning LLM methods with XGBoost utilizes the strength of LLMs as great few shot learners and the efficiency of XGBoost when the training dataset is sufficient. Therefore, SDoH-GPT can extract SDoH without relying on extensive medical annotations or costly human intervention.

Results: Our approach achieved tenfold and twentyfold reductions in time and cost, respectively, and superior consistency with human annotators measured by Cohen's kappa of up to 0.92. The innovative combination of LLM and XGBoost can ensure high accuracy and computational efficiency while consistently maintaining 0.90+ AUROC scores.

Discussion: This study has verified SDoH-GPT on three datasets and highlights the potential of leveraging LLM and XGBoost to revolutionize medical note classification, demonstrating its capability to achieve highly accurate classifications with significantly reduced time and cost.

Conclusion: The key contribution of this study is the integration of LLM with XGBoost, which enables cost-effective and high quality annotations of SDoH. This research sets the stage for SDoH can be more accessible, scalable, and impactful in driving future healthcare solutions.

Objective: Frequent premature ventricular complexes (PVCs) can lead to adverse health conditions such as cardiomyopathy. The linear correlation between PVC frequency and heart rate (as positive, negative, or neutral) on a 24-hour Holter recording has been proposed as a way to classify patients and guide treatment with beta-blockers. Our objective was to evaluate the robustness of this classification to measurement methodology, different 24-hour periods, and nonlinear dependencies of PVCs on heart rate.

Materials and methods: We analyzed 82 multi-day Holter recordings (1-7 days) collected from 48 patients with frequent PVCs (burden 1%-44%). For each record, linear correlation between PVC frequency and heart rate was computed for different 24-hour periods and using different length intervals to determine PVC frequency.

Results: Using a 1-hour interval, the correlation between PVC frequency and heart rate was consistently positive, negative, or neutral on different days in only 36.6% of patients. Using shorter time intervals, the correlation was consistent in 56.1% of patients. Shorter time intervals revealed nonlinear and piecewise linear relationships between PVC frequency and heart rate in many patients.

Discussion: The variability of the correlation between PVC frequency and heart rate across different 24-hour periods and interval durations suggests that the relationship is neither strictly linear nor stationary. A better understanding of the mechanism driving the PVCs, combined with computational and biological models that represent these mechanisms, may provide insight into the observed nonlinear behavior and guide more robust classification strategies.

Conclusion: Linear correlation as a tool to classify patients with frequent PVCs should be used with caution. It is sensitive to the specific 24-hour period analyzed and the methodology used to segment the data. More sophisticated classification approaches that can capture nonlinear and time-varying dependencies should be developed and considered in clinical practice.