Uptake of Generative AI Integrated With Electronic Health Records in US Hospitals.

Importance: There is widespread enthusiasm about generative artificial intelligence (AI), but no systematic evidence on its implementation across health care organizations. Objective: To describe adoption of generative AI integrated with the electronic health record (EHR) by nonfederal acute care hospitals, how adoption relates to experience using and evaluating predictive AI, and hospital characteristics. Design, Setting, and Participants: This survey study of nonfederal acute care US hospitals used the 2024 American Hospital Association (AHA) Information Technology (IT) Supplement survey. The survey was completed by individuals most knowledgeable about health IT at the participating hospitals. Exposures: Experience with predictive AI, source of predictive AI, local evaluation practices (evaluation for accuracy and bias as well as postdeployment evaluation), and EHR developer were collected from the 2024 AHA IT Supplement. Hospital characteristics, including critical access hospital status, multihospital system membership, and teaching status, were collected from the 2024 AHA Annual Survey. Hospital operating margins, uncompensated care burden, and percentage of discharges from Medicaid were collected from the 2022 Medicare Cost Report. Main Outcomes and Measures: Whether the hospital was an early adopter of generative AI integrated with their EHR (currently used generative AI), fast follower (planned to use in the next year), or delayed adopter (planned to use in 5 years, no plans, or do not know). Results: A total of 2174 hospitals (1003 [weighted percentage, 50.4%] small; 1382 [weighted percentage, 60.8%] urban core-based; 1668 [weighted percentage, 68.8%] part of a multihospital system) responded to questions about their use of AI (51.5% response rate). Overall, 762 hospitals (weighted percentage, 31.5%) were early adopters of generative AI in 2024, 540 (weighted percentage, 24.7%) were fast followers, and 872 (weighted percentage, 43.7%) were delayed adopters. In unadjusted analyses, independent hospitals and critical access hospitals were less likely to be either early adopters or fast followers than delayed adopters. In adjusted analyses, hospitals that used predictive AI were more likely to be early adopters or fast followers than delayed adopters (difference, 26.2 [95% CI, 16.8-35.6] percentage points). Users of Epic were more likely to be early adopters and fast followers than users of other EHRs (eg, likelihood of being an early adopter or fast follower, Epic vs Oracle users: 21.9 [95% CI, 16.3-27.4] percentage points). Hospitals that reported conducting all local evaluation practices (accuracy, bias, postdeployment) were slower to adopt than hospitals that reported only 1 evaluation practice (all local evaluation processes: 12.1 [95% CI, 4.5-19.6] percentage points less likely to be early adopters than fast followers). Conclusions and Relevance: In this survey study of US hospitals, more than half of US hospitals reported that they would likely implement generative AI by the end of 2025. Results indicate the value of providing support to ensure hospitals can adopt beneficial generative AI and the need for developing and disseminating best practices for generative AI evaluation across organizations.