JMIR AI最新文献

Unlabelled: With the rapid development of artificial intelligence (AI), particularly large language models, there is growing interest in adopting AI approaches within academic medical centers (AMCs). However, the vast amounts of data required for AI and the sensitive nature of medical information pose significant challenges to developing high-performing models at individual institutions. Furthermore, recent changes in government funding priorities may result in the decentralization of biomedical data repositories that risk creating significant barriers to effective data sharing and robust model development. This has generated significant interest in federated learning (FL), which enables collaborative model training without transferring data between institutions, thereby enhancing the protection of proprietary and sensitive information. While FL offers a crucial pathway to enable multi-institutional AI development while maintaining data privacy, it also exposes AMCs to novel governance, security, and operational risks that are not fully addressed by existing procedures. In response, this manuscript provides a perspective grounded in both leading international standards (NIST AI RMF [National Institute of Standards and Technology Artificial Intelligence Risk Management Framework], International Organization for Standardization (ISO) and International Electrotechnical Commission (IEC) 42001) and in the real-world governance experience of AMC leadership. We present a risk differentiation framework, an FL risk matrix, and a set of essential governance artifacts-each mapped to key institutional challenges and reviewed for alignment with core standards but offered as pragmatic, illustrative guides rather than prescriptive checklists. Together, these tools represent a novel resource to support AMC security, privacy, and governance leaders with standards-informed, context-sensitive tools for addressing the evolving risks of FL in biomedical research and clinical environments.

Background: The modified Rankin scale (mRS) is an important metric in stroke research, often used as a primary outcome in clinical trials and observational studies. The mRS can be assessed retrospectively from electronic health records (EHRs), but this process is labor-intensive and prone to interrater variability. Large language models (LLMs) have demonstrated potential in automating text classification.

Objective: We aimed to create a fine-tuned LLM that can analyze EHR text and classify mRS scores for clinical and research applications.

Methods: We performed a retrospective cohort study of patients admitted to a specialist stroke neurology service at a large academic hospital system between August 2020 and June 2023. Each patient's medical record was reviewed at two time points: (1) at hospital discharge and (2) approximately 90 days post discharge. Two independent researchers assigned an mRS score at each time point. Two separate models were trained on EHR passages with corresponding mRS scores as labeled outcomes: (1) a multiclass model to classify all seven mRS scores and (2) a binary model to classify functional independence (mRS scores 0-2) versus non-independence (mRS scores 3-6). Four-fold cross-validation was conducted using accuracy and the Cohen κ as model performance metrics.

Results: A total of 2290 EHR passages with corresponding mRS scores were included in model training. The multiclass model-considering all seven scores of the mRS-attained an accuracy of 77% and a weighted Cohen κ of 0.92. Class-specific accuracy was the highest for mRS score 4 (90%) and the lowest for mRS score 2 (28%). The binary model-considering only functional independence versus non-independence-attained an accuracy of 92% and a Cohen κ of 0.84.

Conclusions: Our findings demonstrate that LLMs can be successfully trained to determine mRS scores through EHR text analysis; however, improving discrimination between intermediate scores is required.

Background: Artificial intelligence (AI) is a topic of considerable hype, with many actors sensing its high potential for health care applications. Despite this, the adoption has been slow, with few applications being implemented in clinical practice.

Objective: The aim of our study was to investigate the challenges associated with using AI in health care, as well as provide suggestions for how further adoption of AI within health care organizations can be facilitated.

Methods: A qualitative case study with a mixed methods approach was conducted at one of Sweden's largest hospitals. Regulatory approved AI medical devices were analyzed, and primary qualitative data from 14 expert interviews were collected and cross-referenced with secondary quantitative data. The framework of technological innovation systems was used to analyze the system factors and their dynamics to identify blocking mechanisms and areas for improvement.

Results: The challenges related to knowledge development, diffusion, legitimation, and resource mobilization could trigger a cascade of positive activities, thereby significantly enhancing the overall performance of the innovation system. Creating dedicated testing environments to evaluate safety and efficacy would facilitate the routine clinical use and reinforce the use of AI innovations in health care organizations.

Conclusions: This analysis shows that the adoption of AI health care technology innovations can be accelerated through targeted strategies and supportive mechanisms triggering virtuous cycles that facilitate clinical validation and generate compelling use cases. The interconnection between guidance of search and entrepreneurial experimentation has been confirmed, providing the initial conditions for knowledge development, diffusion, and legitimation in the early stages of emerging technologies.

Background: Large language models (LLMs) are increasingly integrated into health care, where they contribute to patient care, administrative efficiency, and clinical decision-making. Despite their growing role, the ability of LLMs to handle imperfect inputs remains underexplored. These imperfections, which are common in clinical documentation and patient-generated data, may affect model reliability.

Objective: This study investigates the impact of input perturbations on LLM performance across three dimensions: (1) overall effectiveness in different health-related applications, (2) comparative effects of different types and levels of perturbations, and (3) differential impact of perturbations on health-related terms versus non-health-related terms.

Methods: We systematically evaluate 3 LLMs on 3 health-related tasks using a novel dataset containing 3 types of human-like variations (redaction, homophones, and typographical errors) at different perturbation levels.

Results: Contrary to expectations, LLMs demonstrate notable robustness to common variations, and in more than half of the cases (151/270, 55.92%), the performance was stable or improved. In some cases (38/270, 14.07%), variations resulted in an increased performance, especially when dealing with lower perturbation levels. Redactions, often stemming from privacy concerns or cognitive lapses, are more detrimental than other variations.

Conclusions: Our findings highlight the need for health care applications powered by LLMs to be designed with input variability in mind. Robustness to noisy or imperfect inputs is essential for maintaining reliability in real-world clinical settings, where data quality can vary widely. By identifying specific vulnerabilities and strengths, this study provides actionable insights for improving model resilience and guiding the development of safer, more effective artificial intelligence tools in health care. The accompanying dataset offers a valuable resource for further research into LLM performance under diverse conditions.

Background: Images created with generative artificial intelligence (AI) tools are increasingly used for health communication due to their ease of use, speed, accessibility, and low cost. However, AI-generated images may bring practical and ethical risks to health practitioners and the public, including through the perpetuation of stigma against vulnerable and historically marginalized groups.

Objective: To understand the potential value of AI-generated images for health care and public health communication, we sought to analyze images of substance use disorder and recovery generated with ChatGPT. Specifically, we sought to investigate: (1) the default visual outputs produced in response to a range of prompts about substance use disorder and recovery, and (2) the extent to which prompt modification and guideline-informed prompting could mitigate potentially stigmatizing imagery.

Methods: We performed a mixed-methods case study examining depictions of substance use and recovery in images generated by ChatGPT 4.o. We generated images (n=84) using (1) prompts with colloquial and stigmatizing language, (2) prompts that follow best practices for person-first language, (3) image prompts written by ChatGPT, and (4) a custom GPT informed by guidelines for images of SUD. We then used a mixed-methods approach to analyze images for demographics and stigmatizing elements.

Results: Images produced in the default ChatGPT model featured primarily White men (81%, n=34). Further, images tended to be stigmatizing, featuring injection drug use, dark colors, and symbolic elements such as chains. These trends persisted even when person-first language prompts were used. Images produced by the guideline-informed custom GPT were markedly less stigmatizing; however, they featured almost only Black women (74%, n=31).

Conclusions: Our findings confirm prior research about stigma and biases in AI-generated images and extend this literature to substance use. However, our findings also suggest that (1) images can be improved when clear guidelines are provided and (2) even with guidelines, iteration is needed to create an image that fully concords with best practices.

Background: Large language models (LLMs) have fundamentally transformed approaches to natural language processing tasks across diverse domains. In health care, accurate and cost-efficient text classification is crucial-whether for clinical note analysis, diagnosis coding, or other related tasks-and LLMs present promising potential. Text classification has long faced multiple challenges, including the need for manual annotation during training, the handling of imbalanced data, and the development of scalable approaches. In health care, additional challenges arise, particularly the critical need to preserve patient data privacy and the complexity of medical terminology. Numerous studies have leveraged LLMs for automated health care text classification and compared their performance with traditional machine learning-based methods, which typically require embedding, annotation, and training. However, existing systematic reviews of LLMs either do not specialize in text classification or do not focus specifically on the health care domain.

Objective: This research synthesizes and critically evaluates the current evidence in the literature on the use of LLMs for text classification in health care settings.

Methods: Major databases (eg, Google Scholar, Scopus, PubMed, ScienceDirect) and other resources were queried for papers published between 2018 and 2024, following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines, resulting in 65 eligible research articles. These studies were categorized by text classification type (eg, binary classification, multilabel classification), application (eg, clinical decision support, public health and opinion analysis), methodology, type of health care text, and the metrics used for evaluation and validation.

Results: The systematic review includes 65 research articles published between 2020 and Q3 2024, showing a significant increase in publications over time, with 28 papers published in Q1-Q3 2024 alone. Fine-tuning was the most common LLM-based approach (35 papers), followed by prompt engineering (17 papers). BERT (Bidirectional Encoder Representations from Transformers) variants were predominantly used for multilabel classification (50%), whereas closed-source LLMs were most commonly applied to binary (44.0%) and multiclass (30.6%) classification tasks. Clinical decision support was the most frequent application (29 papers). Over 80% of studies used English-language datasets, with clinical notes being the most common text type. All studies employed accuracy-related metrics for evaluation, and the findings consistently showed that LLMs outperformed traditional machine learning approaches in health care text classification tasks.

Conclusions: This review identifies existing gaps in the literature and highlights future research directions for further investigation.

Large language models (LLMs) are increasingly used by patients and families to interpret complex medical documentation, yet most evaluations focus only on clinician-judged accuracy. In this study, 50 pediatric cardiac intensive care unit notes were summarized using GPT-4o mini and reviewed by both physicians and parents, who rated readability, clinical fidelity, and helpfulness. There were important discrepancies between parents and clinicians in the realm of helpfulness, along with important insights by clinicians assessing clinical accuracy and parents assessing readability. This study highlights the need for dual-perspective frameworks that balance clinical precision with patient understanding.