Introductory information and resources

Reviews, editorials, and other papers on what AI use looks like in the US healthcare system.

Why do probabilistic clinical models fail to transport between sites

Lasko TA, Strobl EV, Stead WW. Why do probabilistic clinical models fail to transport between sites. npj Digit Med. 2024;7:53. doi:10.1038/s41746-024-01037-4

This paper argues that failure of probabilistic clinical models to generalize across sites is an expected outcome. The authors highlight some common causes of clinical probabilistic models performing well in one site but often failing in another site. They categorize these common causes into experimental sources, which are controllable, and inherent factors, which occur due to the way the clinical data are generated. Experimental factors include model overfitting, information leakage during training, and asking models a question that they were not trained to answer. Inherent factors include the presence of site-specific processes and differences in causal prevalence. One solution they propose is to take advantage of the instability by training a Process Model that infers site-specific relationships and a Disease Model that uses those relationships to make inferences about a patient’s clinical state.

Artificial Intelligence Agents in Healthcare Research: A Scoping Review

Njei B, Al-Ajlouni YA, Kanmounye US, Boateng S, Nguefang GL, Njei N, et al. PLOS One. 2026.

This article reviews studies that examined the use of agentic AI systems in healthcare research and clinical applications. The authors categorized AI systems into conversational agents, workflow automation assistants, and multimodal decision support systems. Most studies focused on simulated or laboratory settings rather than real world clinical environments, with limited evidence regarding patient outcomes and safety. The article highlights the role of retrieval-augmented generation, multi-agent collaboration, and self-correction mechanisms in healthcare AI while emphasizing the need for stronger clinical validation, governance and safety oversight before widespread implementation.

Hospital Trends in the Use, Evaluation, and Governance of Predictive AI, 2023-2024

Chang W, Owusu-Mensah P, Everson J, Richwine C. Office of the Assistant Secretary for Technology Policy. 2025 Sep; Data brief No. 80.

The adoption of predictive AI in healthcare increased from 66% in 2023 to 71% in 2024. This study examines the use, evaluation, and governance of predictive AI in healthcare. The authors highlight the rapid growth of predictive AI in the administrative side to simplify billing and facilitate scheduling. Small, rural, independent, and critical access hospitals used predictive AI at lower rates compared to other hospital types.

Adoption of artificial intelligence in healthcare: survey of health system priorities, successes, and challenges

Poon EG, Harris Lemak C, Rojas JC, Guptill J, Classen D. J Am Med Inform Assoc. 2025;32(7):1093–1100.

In a survey of 43 non-profit US health systems, this study analyzes adoption of AI, successful use cases, and challenges faced related to AI implementation in healthcare. AI tools were most commonly used for clinical documentation, while barriers included concerns about relatability, financial constraints, and regulatory uncertainty.

How artificial intelligence tools influence health care delivery: categorization of clinical, operational, consumer, and hybrid uses

Angus DC, Khera R, Liu T, et al. JAMA. 2025;334(18):1650–1664.

This report comes out of the JAMA Summit on AI, which brought together stakeholders to review the state of AI in healthcare and discuss how it should be developed, evaluated, monitored, and regulated. The authors provide a framework for categorizing health and health care AI tools into four primary groups to frame evaluation efforts: clinical tools (e.g., sepsis alerts), business operation tools (e.g., staff scheduling), direct to consumer tools used by individuals (e.g., smartwatch biosensors for fall detection), and hybrid tools for clinical and administrative workflows (e.g., ambient AI scribes). Authors also recommend engaging diverse stakeholders in the AI lifecycle, implementing tools for evaluation and monitoring, and developing foundational data infrastructure to facilitate evaluation.