Journal of the American Medical Informatics Association最新文献

Objectives: To explore the complexities of eliminating race correction in clinical artificial intelligence (AI), the pitfalls of naive solutions, and to propose systematic strategies for equitable model development.

Background and significance: Race correction in clinical AI, as in traditional medicine, introduces biases with potentially harmful consequences. Simple removal of race from models is insufficient due to the lasting influence of historically biased data.

Approach: We analyze 4 standardized scenarios to demonstrate how race correction manifests in clinical AI: use of race-corrected variables, explicit inclusion of race, inference via proxy variables, and use of race-specific models.

Results: For each scenario, the intuitive solution to removing race correction fails to eliminate bias, often due to legacy effects embedded in the data. More thoughtful approaches are required.

Discussion: Ending race correction in clinical AI requires deliberate, context-sensitive interventions, inclusion of diverse stakeholders, and strategies to make model reasoning more transparent and auditable.

Objectives: Accurate triage in emergency departments (ED) is critical for appropriate resource allocation. While artificial intelligence (AI) has been explored for triage, prior models relied on summarized clinical scenarios. We aimed to develop and evaluate large language models (LLMs) trained on real-world clinical conversations to classify patient urgency.

Materials and methods: We used a nationally curated dataset of anonymized triage-level conversations from 3 tertiary Korean hospitals. Two BERT-based models were developed to classify urgency per the Korean Triage and Acuity Scale (KTAS) into urgent (KTAS 3) or non-urgent (KTAS 4-5). One model tokenized the entire conversation, while the other applied a hierarchical structure with sentence-level tokenization and speaker-role embeddings. Performance metrics included accuracy, precision, recall, and F1-score. We compared our models against ChatGPT GPT-4o and ClinicalBERT, and assessed explainability using SHapley Additive exPlanations (SHAP).

Results: A total of 5244 clinical conversations, 1057 triage-level dialogues were used, with 950 for training and 107 for testing. Our model with hierarchical structure achieved accuracies of 75.94%, significantly outperforming ChatGPT (56.68%) or fine-tuned ClinicalBERT (69.42%). For urgent cases, the best model achieved a recall of 0.9610, outperforming ChatGPT (0.5352). SHapley Additive exPlanations analysis confirmed that our model focused on clinically relevant cues aligned with KTAS criteria.

Conclusion: BERT-based LLMs trained on real-world ED conversations significantly outperform general-purpose models like ChatGPT in triage accuracy. This approach demonstrates the potential for enhancing clinical decision support with interpretable and efficient AI.

Objective: Artificial intelligence (AI) has impacted healthcare at urban and academic medical centers in the US. There are concerns, however, that the promise of AI may not be realized in rural communities. This scoping review aims to determine the extent of AI research in the rural US.

Materials and methods: We conducted a scoping review following the PRISMA guidelines. We included peer-reviewed, original research studies indexed in PubMed, Embase, and WebOfScience after January 1, 2010 and through April 29, 2025. Studies were required to discuss the development, implementation, or evaluation of AI tools in rural US healthcare, including frameworks that help facilitate AI development (eg, data warehouses).

Results: Our search strategy found 26 studies meeting inclusion criteria after full text screening with 14 papers discussing predictive AI models and 12 papers discussing data or research infrastructure. AI models most often targeted resource allocation and distribution. Few studies explored model deployment and impact. Half noted the lack of data and analytic resources as a limitation. None of the studies discussed examples of generative AI being trained, evaluated, or deployed in a rural setting.

Discussion: Practical limitations may be influencing and limiting the types of AI models evaluated in the rural US. Validation of tools in the rural US was underwhelming.

Conclusion: With few studies moving beyond AI model design and development stages, there are clear gaps in our understanding of how to reliably validate, deploy, and sustain AI models in rural settings to advance health in all communities.

Objective: Electronic health records (EHRs) lack a widely adopted standard for recording transgender and gender diverse (TGD) status, complicating research on TGD health. Computational models have been developed to identify TGD individuals in EHRs; however, gaps remain in understanding which components contribute to stronger phenotyping approaches. This scoping review evaluates EHR-based models for identifying TGD individuals, focusing on identifier types, performance, external validation, and ethical reporting to guide best practices.

Materials and methods: We searched PubMed, CINAHL, Web of Science, and Embase for peer-reviewed articles published before January 2024, following PRISMA-ScR guidelines. Included studies used EHR data to identify TGD individuals, verified TGD status, reported or allowed calculation of positive predictive value (PPV), and listed identifiers. Two authors screened and extracted data. We categorized models by data type and logic (structured, unstructured, and multimodal), summarized PPV distributions, and synthesized author-reported ethical considerations.

Results: Fourteen studies describing 50 models met inclusion criteria. Models using TGD-related diagnostic codes alone (n = 11) or requiring both structured and unstructured data (n = 6) showed the highest mean PPVs (85.3% and 97.1%). Models validated on larger confirmed TGD cohorts reported more stable performance, but external validation was rare. Most studies minimally addressed ethics; only 3 described protective measures or stakeholder engagement.

Discussion: Phenotyping of TGD individuals in EHR data remains heterogeneous in design and ethical transparency. Reported PPVs should be interpreted cautiously, as performance is influenced by study design, sample size, and verification methods.

Conclusions: Our recommendations emphasize the components that strengthen phenotyping approaches-identifier choice, multimodal intersection logic, validation practices, and ethical safeguards-rather than endorsing any single model.

Introduction: We created a probabilistic maternal-child electronic health record (EHR) linkage algorithm to promote clinical research in maternal-child health.

Methods: We used EHR data from 1994 to 2024 to create an XGBoost model to predict maternal-child linkages. The model used standard EHR elements as predictor variables, including first name, last name, birthdate, address, phone number, email, and an EHR-embedded maternal-child indicator as the deterministic outcome.

Results: From 82 million unique records, 6.2 billion potential pairs met blocking criteria. Of the potential pairs, 33 364 674 contained the deterministic indicator and were used as cases, and an equal number of controls were randomly sampled. The final model obtained an accuracy of 92%, a precision of 98%, a recall of 87%, and an F1-score of 92%.

Conclusion: We derived and validated a probabilistic maternal-child linkage algorithm using routinely collected EHR data elements that could benefit future observational research in maternal-child health.

Objective: To support ambulatory care innovation, we created Observer, a multimodal dataset comprising videotaped outpatient visits, electronic health record (EHR) data, and structured surveys. This paper describes the data collection procedures and summarizes the clinical and contextual features of the dataset.

Materials and methods: A multistakeholder steering group shaped recruitment strategies, survey design, and privacy-preserving design. Consented patients and primary care providers (PCPs) were recorded using room-view and egocentric cameras. EHR data, metadata, and audit logs were also captured. A custom de-identification pipeline, combining transcript redaction, voice masking, and facial blurring, ensured video and EHR HIPAA compliance.

Results: We report on the first 100 visits in this continually growing dataset. Thirteen PCPs from 4 clinics participated. Recording the first 100 visits required approaching 210 patients, from which 129 consented (61%), with 29 patients missing their scheduled encounter after consenting. Visit lengths ranged from 5 to 100 minutes, covering preventive care to chronic disease management. Survey responses revealed high satisfaction: 4.24/5 (patients) and 3.94/5 (PCPs). Visit experience was unaffected by the presence of video recording technology.

Discussion: We demonstrate the feasibility of capturing rich, real-world primary care interactions using scalable, privacy-sensitive methods. Room layout and camera placement were key influences on recorded communication and are now added to the dataset. The Observer dataset enables future clinical AI research/development, communication studies, and informatics education among public and private user groups.

Conclusion: Observer is a new, shareable, real-world clinic encounter research and teaching resource with a representative sample of adult primary care data.

Objectives: To report on the feasibility of a simultaneous, enterprise-wide deployment of EHR-integrated ambient scribe technology across a large academic health system.

Materials and methods: On January 15, 2025, ambient scribing was made available to over 2400 ambulatory and emergency department clinicians. We tracked utilization rates, technical support needs, and user feedback.

Results: By March 31, 2025, 20.1% of visit notes incorporated ambient scribing, and 1223 clinicians had used ambient scribing. Among 209 respondents (22.1% of 947 surveyed), 90.9% would be disappointed if they lost access to ambient scribing, and 84.7% reported a positive training experience.

Discussion: Enterprise-wide simultaneous deployment combined with a low-barrier training model enabled immediate access for clinicians and reduced administrative burden by concentrating go-live efforts. Support needs were manageable.

Conclusion: Simultaneous enterprise-wide deployment of ambient scribing was feasible and provided immediate access for clinicians.

Objective: To develop a transfer-learning Bayesian sparse logistic regression model that transfers information learned from one dataset to another by using an informed prior to facilitate model fitting in small-sample clinical patient-level prediction problems that suffer from a lack of available information.

Methods: We propose a Bayesian framework for prediction using logistic regression that aims to conduct transfer-learning on regression coefficient information from a larger dataset model (order 105-106 patients by 105 features) into a small-sample model (order 103 patients). Our approach imposes an informed, hierarchical prior on each regression coefficient defined as a discrete mixture of the Bayesian Bridge shrinkage prior and an informed normal distribution. Performance of the informed model is compared against traditional methods, primarily measured by area under the curve, calibration, bias, and sparsity using both simulations and a real-world problem.

Results: Across all experiments, transfer-learning outperformed the traditional L1-regularized model across discrimination, calibration, bias, and sparsity. In fact, even using only a continuous shrinkage prior without the informed prior increased model performance when compared to L1-regularization.

Conclusion: Transfer-learning using informed priors can help fine-tune prediction models in small datasets suffering from a lack of information. One large benefit is in that the prior is not dependent on patient-level information, such that we can conduct transfer-learning without violating privacy. In future work, the model can be applied for learning between disparate databases, or similar lack-of-information cases such as rare outcome prediction.

Objective: Privacy preserving record linkage (PPRL) refers to techniques used to identify which records refer to the same person across disparate datasets while safeguarding their identities. PPRL is increasingly relied upon to facilitate biomedical research. A common strategy encodes personally identifying information for comparison without disclosing underlying identifiers. As the scale of research datasets expands, it becomes crucial to reassess the privacy risks associated with these encodings. This paper highlights the potential re-identification risks of some of these encodings, demonstrating an attack that exploits encoding repetition across patients.

Materials and methods: The attack leverages repeated PPRL encoding values combined with common demographics shared during PPRL in the clear (e.g., 3-digit ZIP code) to distinguish encodings from one another and ultimately link them to identities in a reference dataset. Using US Census statistics and voter registries, we empirically estimate encodings' re-identification risk against such an attack, while varying multiple factors that influence the risk.

Results: Re-identification risk for PPRL encodings increases with population size, number of distinct encodings per patient, and amount of demographic information available. Commonly used encodings typically grow from <1% re-identification rate for datasets under one million individuals to 10%-20% for 250 million individuals.

Discussion and conclusion: Re-identification risk often remains low in smaller populations, but increases significantly at the larger scales increasingly encountered today. These risks are common in many PPRL implementations, although, as our work shows, they are avoidable. Choosing better tokens or matching tokens through a third party without the underlying demographics effectively eliminates these risks.