JMIR Publications

ABSTRACT

Background:

As machine learning (ML) technologies shift the focus from development to real-world deployment in the recent decade, U.S. healthcare providers and hospital administrators have increasingly embraced ML technologies — particularly through integration with electronic health record (EHR) systems. This evolving landscape underscores the need for empirical evidence on ML adoption and its determinants; however, the relationship between hospital characteristics and ML adoption within EHRs is not thoroughly explored.

Objective:

This study examines the extent of ML adoption within electronic health record (EHR) systems across U.S. hospitals and identifies hospital-level factors associated with ML adoption within EHR.

Methods:

We used data from the 2022 American Hospital Association (AHA) Annual Survey and the 2023 AHA Information Technology (IT) Supplement Survey, including 2,360 general and acute care hospitals. Descriptive statistics were used to assess ML adoption rates and functions, while multivariate regression models estimated the associations between hospital characteristics and ML adoption within the EHR. Subsequent analyses further explored adoption patterns across clinical, operational and individual ML functions.

Results:

Overall, 69.5% of hospitals had adopted at least one ML function into their EHR systems by 2023. The most commonly adopted ML functions were for predictions of inpatient risks (65.3%) and outpatient follow-up (57.6%). Hospitals primarily relied on EHR vendors for EHR development. Model evaluation practices were limited: only 61.0% assessed model accuracy, and 42.9% evaluated model bias. The probability of ML adoption was significantly higher in non-for-profit (0.083; 95% CI [0.035, 0.13]; P<.001), large (0.17, 95% CI [0.1, 0.24]; P<.001), metropolitan (0.089; 95% CI [0.051, 0.13]; P<.001), and health-system–affiliated hospitals (0.31; 95% CI [0.28, 0.33]; P<.001). Greater numbers of ML functions were associated with large (0.41; 95% CI [0.14, 0.67]; P=.002), metropolitan (0.52; 95% CI [0.30, 0.74]; P<.001) and health-system-affiliated hospitals (2.31; 95% CI [1.96, 2.67]; P<.001). Regarding the adoption of clinical ML functions, hospital size, ownership, health system affiliation and metropolitan status play a significant positive role; however, for-profit hospitals showed particular interest, especially in ML scheduling functions.

Conclusions:

ML adoption is influenced by organizational resources and strategic goals, with potential risks for digital inequity. Meanwhile, limited quality control and evaluation underscore the need for regulatory oversight and increased support for under-resourced hospitals. While the adoption of ML into EHR systems is expanding, resource-based disparities in adoption and oversight raise critical concerns. Targeted policies are necessary to ensure equitable, safe, and effective use of ML in healthcare settings.