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

Some capabilities required to effectively validate predictive AI provide a foundation for monitoring generative AI.

“It is likely that some of the capabilities required to effectively validate predictive AI provide a foundation for monitoring generative AI, but the science of ensuring that generative AI tools are fair, appropriate, valid, effective, and safe is still emerging.23,24”

• cites [23] Zack T, Lehman E, Suzgun M,. Assessing the potential of GPT-4 to perpetuate racial and gender biases in health care: a model evaluation study. Lancet Digit Health. 2024; 6(1): e12- e22. doi: 10.1016/S2589-7500(23)00225-X 38123252 sourceNot checked — no full text of the source
• cites [24] de Hond A, Leeuwenberg T, Bartels R,. From text to treatment: the crucial role of validation for generative large language models in health care. Lancet Digit Health. 2024; 6(7): e441- e443. doi: 10.1016/S2589-7500(24)00111-0 38906607 sourceCould not verify

  In the discussion of validation challenges (page unavailable), the paper contrasts standardized criteria for predictive AI with the more complex validation needs of generative AI. It calls for continuous validation of LLMs, but does not say that capabilities used to validate predictive AI provide a foundation for monitoring generative AI. I searched for that proposed transfer of capabilities and for monitoring. (LLM, conf=0.70)

  Experimental — accuracy not yet measured

There are widespread concerns that the performance of AI tools in early validation studies may not reflect their use in practice.

“There are widespread concerns that the performance of these tools in early validation studies may not reflect their use in practice.19,20”

• cites [19] Van Calster B, Steyerberg EW, Wynants L, van Smeden M. There is no such thing as a validated prediction model. BMC Med. 2023; 21(1): 70. doi: 10.1186/s12916-023-02779-w 36829188 PMC9951847 sourceSupported
• cites [20] Youssef A, Pencina M, Thakur A, Zhu T, Clifton D, Shah NH. External validation of AI models in health should be replaced with recurring local validation. Nat Med. 2023; 29(11): 2686- 2687. 37853136 sourceNot checked — no full text of the source

These trends parallel previously reported digital divides in hospital adoption of health IT.

“These trends parallel previously reported digital divides in hospital adoption of health IT.35,36”

• cites [35] Apathy NC, Holmgren AJ, Nong P, Adler-Milstein J, Everson J. Trending in the right direction: critical access hospitals increased adoption of advanced electronic health record functions from 2018 to 2023. J Am Med Inform Assoc. 2025; 32(1): 71- 78. doi: 10.1093/jamia/ocae267 39504475 PMC11648721 sourceSupported
• cites [36] Gabriel MH, Richwine C, Strawley C, Barker W, Everson J. Interoperable exchange of patient health information among US hospitals: 2023. ASTP Health IT Data Brief. 2024. Accessed November 6, 2025. https://www.ncbi.nlm.nih.gov/books/NBK606033/ 39413225Not checked yet

  Not checked in the free run.

Hospitals with high Medicaid discharges were more likely to have no implementation plans for generative AI.

“Independent hospitals and those with high Medicaid discharges were more likely to have no implementation plans, while those that reported robust local evaluation of predictive AI were more likely to plan to use generative AI in 1 year than to currently use it.”

72.6% of hospitals that did not use any predictive AI were delayed adopters of generative AI.

“52.8% (95% CI, 50.0%55.7%) of hospitals that used ML-based predictive AI were early generative AI adopters and 72.6% (95% CI, 65.5%-79.6%) of hospitals that did not use any predictive AI were delayed adopters.”

43.7% of nonfederal US hospitals were delayed adopters of generative AI (planning in 5 years, no plans, or unsure).

“872 hospitals (43.7%) were delayed adopters: 132 (6.7%) indicated plans in the next 5 years; 638 (32%) reported their hospital did not use generative AI and did not indicate plans to do so, and 102 (5%) did not know whether their hospital used generative AI.”

• Also stated in: § Results · plain

  “Weighted % | 100 | 31.5 | 24.7 | 43.7 |”

32% of nonfederal US hospitals reported no plans to use generative AI.

“872 hospitals (43.7%) were delayed adopters: 132 (6.7%) indicated plans in the next 5 years; 638 (32%) reported their hospital did not use generative AI and did not indicate plans to do so, and 102 (5%) did not know whether their hospital used generative AI.”

24.7% of nonfederal US hospitals planned to use generative AI integrated with the electronic health record in the next year.

“In this survey study of 2174 nonfederal US hospitals, 31.5% reported using generative AI in 2024 and 24.7% planned to do so in 1 year.”

• Also stated in: § Results · hedged

  “Another 540 hospitals (24.7%) were fast followers, indicating their hospital planned to use generative AI in the next year”

Absolute differences in generative AI adoption by hospital characteristics were generally smaller than for predictive AI.

“Absolute differences in generative AI adoption by hospital characteristics were generally smaller than for predictive AI (eTable 1 in Supplement 1).”

Independent hospitals were more likely to have no implementation plans for generative AI.

“Independent hospitals and those with high Medicaid discharges were more likely to have no implementation plans, while those that reported robust local evaluation of predictive AI were more likely to plan to use generative AI in 1 year than to currently use it.”

Major teaching hospitals had a higher proportion of early generative AI adopters (53.9%) than nonteaching hospitals (25.9%).

“53.9% (95% CI, 46.0%-61.9%) of major teaching hospitals were early adopters compared with only 25.9% (95% CI, 23.3%-28.6%) of nonteaching hospitals.”

In 2024, 31.5% of nonfederal US hospitals reported currently using generative AI integrated with the electronic health record.

“In this survey study of 2174 nonfederal US hospitals, 31.5% reported using generative AI in 2024 and 24.7% planned to do so in 1 year.”

• Also stated in: § Results · plain

  “A total of 762 hospitals (31.5%) were early adopters of generative AI, meaning they indicated that their hospital or health system currently used generative AI integrated with their EHR”

5% of nonfederal US hospitals did not know whether their hospital used generative AI.

“872 hospitals (43.7%) were delayed adopters: 132 (6.7%) indicated plans in the next 5 years; 638 (32%) reported their hospital did not use generative AI and did not indicate plans to do so, and 102 (5%) did not know whether their hospital used generative AI.”

Hospital experience with predictive AI was strongly associated with use of generative AI.

“Hospital experience with predictive AI was strongly associated with use of generative AI (Table 2).”

6.7% of nonfederal US hospitals planned to use generative AI in the next 5 years.

“872 hospitals (43.7%) were delayed adopters: 132 (6.7%) indicated plans in the next 5 years; 638 (32%) reported their hospital did not use generative AI and did not indicate plans to do so, and 102 (5%) did not know whether their hospital used generative AI.”

Fewer US hospitals currently used generative AI (31.5%) than used predictive AI (52.2%) or other predictive models (54.8%) integrated with the EHR.

“Overall, fewer respondents indicated their hospital currently used generative AI than used predictive AI (used by 1253 hospitals [52.2%]) or other predictive models (used by 1286 hospitals [54.8%]) integrated with the EHR.”

Hospitals that reported robust local evaluation of predictive AI were more likely to plan to use generative AI in 1 year than to currently use it.

“Independent hospitals and those with high Medicaid discharges were more likely to have no implementation plans, while those that reported robust local evaluation of predictive AI were more likely to plan to use generative AI in 1 year than to currently use it.”

Major teaching hospitals relative to nonteaching hospitals were 9.6 (95% CI, 1.2-18.0) percentage points more likely to be an early adopter or fast follower than a delayed adopter.

“as were major teaching hospitals relative to nonteaching hospitals (9.6 [95% CI, 1.2-18.0] percentage points)”

The 20% of hospitals with the highest proportion of Medicaid discharges were 4.9 (95% CI, 0.7-9.1) percentage points less likely to be early adopters or fast followers than hospitals with a lower proportion of Medicaid discharges.

“The 20% of hospitals with the highest proportion of Medicaid discharges were 4.9 (95% CI, 0.7-9.1) percentage points less likely to be early adopters or fast followers than hospitals with a lower proportion of Medicaid discharges.”

For-profit hospitals were less likely to be early adopters than fast followers (−23.6 [95% CI, −38.4 to −8.8] percentage points).

“For-profit and government hospitals were less likely to be early adopters than fast followers (for profit: −23.6 [95% CI, −38.4 to −8.8] percentage points; government: −12.1 [95% CI, −21.0 to −3.3] percentage points), as were hospitals with a high share of Medicaid discharges (−8.9 [95% CI, −15.1 to −2.7] percentage points).”

In unadjusted analysis, hospitals in multihospital systems were more than twice as likely to be early adopters of generative AI or fast followers than independent hospitals.

“In unadjusted analysis, hospitals in multihospital systems were more than twice as likely to be early adopters of generative AI or fast followers than independent hospitals, and critical access and rural hospitals were substantially less likely to be early adopters or fast followers than their counterparts.”

Hospitals with high operating margins were 5.9 (95% CI, 0.4 to 11.4) percentage points more likely to be early adopters (vs fast followers).

“Hospitals with high operating margins were 5.9 (95% CI, 0.4 to 11.4) percentage points more likely to be early adopters (Table 3).”

Hospitals reporting no local evaluation practices were less likely to be early adopters than hospitals that reported 1 local evaluation practice.

“However, we found that both hospitals reporting no local evaluation practices and hospitals reporting all 3 evaluation practices were less likely to be early adopters than hospitals that reported 1 local evaluation practice.”

Hospitals that reported 1 local evaluation practice for predictive AI were 23.3 (95% CI, 14.7-32.0) percentage points more likely to be early adopters or fast followers than hospitals that reported no evaluation practices.

“Hospitals that reported 1 local evaluation practice for predictive AI were 23.3 (95% CI, 14.7-32.0) percentage points more likely to be early adopters or fast followers than hospitals that reported no evaluation practices.”

• Also stated in: § Discussion · hedged

  “Hospitals that reported at least 1 of the 3 evaluation steps (evaluation for accuracy or bias or postimplementation evaluation) were less likely to be delayed adopters.”

In unadjusted analysis, critical access and rural hospitals were substantially less likely to be early adopters or fast followers than their counterparts.

“In unadjusted analysis, hospitals in multihospital systems were more than twice as likely to be early adopters of generative AI or fast followers than independent hospitals, and critical access and rural hospitals were substantially less likely to be early adopters or fast followers than their counterparts.”

In multivariate analysis, hospitals that used ML-based predictive AI integrated with the EHR were 26.2 (95% CI, 16.8-35.6) percentage points more likely to be early adopters or fast followers than hospitals that did not use predictive AI.

“In multivariate analysis, hospitals that used ML-based predictive AI integrated with the EHR were 26.2 (95% CI, 16.8-35.6) percentage points more likely to be early adopters or fast followers than hospitals that did not use predictive AI (Figure 2A and Table 3).”

• Also stated in: § Discussion · plain

  “First, hospitals that used predictive AI integrated with the EHR were substantially less likely to be delayed adopters of generative AI.”

• Also stated in: § Conclusions · plain

  “Hospitals with limited experience using predictive AI were less likely to adopt generative AI, while fast followers reported more AI evaluation practices than early adopters.”

This digital divide is harder to ascertain in adjusted analyses.

“However, this digital divide is harder to ascertain in adjusted analyses.”

Hospitals that reported using predictive AI from their EHR developer were 22.4 (95% CI, 11.0-33.9) percentage points more likely to be early adopters than hospitals that reported only using models from other sources.

“Hospitals that reported using predictive AI from their EHR developer were 22.4 (95% CI, 11.0-33.9) percentage points more likely to be early adopters than hospitals that reported only using models from other sources.”

• Also stated in: § Discussion · plain

  “Hospitals that used predictive AI from their EHR developer were substantially more likely to be early adopters of generative AI than fast followers.”

Hospitals that reported all 3 local evaluation practices for most or all models were 12.1 (95% CI, 4.5-19.6) percentage points less likely to be early adopters than hospitals reporting 1 evaluation practice.

“Hospitals that reported all 3 local evaluation practices for most or all models were 12.1 (95% CI, 4.5-19.6) percentage points less likely to be early adopters than hospitals reporting 1 evaluation practice.”

• Also stated in: § Discussion · plain

  “However, we found that both hospitals reporting no local evaluation practices and hospitals reporting all 3 evaluation practices were less likely to be early adopters than hospitals that reported 1 local evaluation practice.”

Government hospitals were less likely to be early adopters than fast followers (−12.1 [95% CI, −21.0 to −3.3] percentage points).

“For-profit and government hospitals were less likely to be early adopters than fast followers (for profit: −23.6 [95% CI, −38.4 to −8.8] percentage points; government: −12.1 [95% CI, −21.0 to −3.3] percentage points), as were hospitals with a high share of Medicaid discharges (−8.9 [95% CI, −15.1 to −2.7] percentage points).”

Hospitals that used Epic were 21.9 (95% CI, 16.3-27.4) percentage points more likely than Oracle users to be an early adopter or fast follower than a delayed adopter.

“Hospitals that used Epic were 21.9 (95% CI, 16.3-27.4) percentage points more likely than Oracle users (reference category) to be an early adopter or fast follower than a delayed adopter (Table 3), as were major teaching hospitals relative to nonteaching hospitals (9.6 [95% CI, 1.2-18.0] percentage points) and hospitals with high operating margins (5.2 [95% CI, 0.6-9.9] percentage points).”

Hospitals with high operating margins were 5.2 (95% CI, 0.6-9.9) percentage points more likely to be early adopters or fast followers.

“and hospitals with high operating margins (5.2 [95% CI, 0.6-9.9] percentage points).”

Hospitals with a high share of Medicaid discharges were less likely to be early adopters than fast followers (−8.9 [95% CI, −15.1 to −2.7] percentage points).

“For-profit and government hospitals were less likely to be early adopters than fast followers (for profit: −23.6 [95% CI, −38.4 to −8.8] percentage points; government: −12.1 [95% CI, −21.0 to −3.3] percentage points), as were hospitals with a high share of Medicaid discharges (−8.9 [95% CI, −15.1 to −2.7] percentage points).”

Fast followers reported more AI evaluation practices than early adopters.

“Hospitals with limited experience using predictive AI were less likely to adopt generative AI, while fast followers reported more AI evaluation practices than early adopters.”

In 2024, more than half of US hospitals reported either currently using (31.5%) generative AI integrated with their EHR or planning to do so in the next year (24.7%).

“In 2024, more than half of US hospitals reported either currently using (31.5%) generative AI integrated with their EHR or planning to do so in the next year (24.7%), representing rapid adoption in health care settings following the widespread emergence of generative AI in late 2022.”

• Also stated in: § Conclusions · plain

  “In this national survey study of US hospitals, nearly one-third of US hospitals had adopted generative AI integrated with their EHR in 2024, with another quarter planning to implement it in the next year.”

Hospitals in more alternative payment models were more likely to be early adopters or fast followers (2.6 [95% CI, 1.2-4.0] percentage points per additional alternative payment model).

“Hospitals in more alternative payment models were also more likely to be early adopters or fast followers (2.6 [95% CI, 1.2-4.0] percentage points per additional alternative payment model).”

There is evidence of a digital divide, whereby some hospitals adopt generative AI tools without complete safeguards.

“The findings of this study suggest that by the end of 2025, half of nonfederal acute care hospitals in the United States will use generative AI; there is evidence of a digital divide, whereby some hospitals adopt these tools without complete safeguards.”

By the end of 2025, half of nonfederal acute care hospitals in the United States will use generative AI.

“The findings of this study suggest that by the end of 2025, half of nonfederal acute care hospitals in the United States will use generative AI; there is evidence of a digital divide, whereby some hospitals adopt these tools without complete safeguards.”

• AssumptionRests on an assumption: Hospitals that reported plans to use generative AI in the next year will actually use it by the end of 2025.

Estimates were weighted to reflect the population of hospitals represented in the AHA Annual Survey using inverse probability weights based on hospital characteristics.

“Estimates were weighted to reflect the population of hospitals represented in the AHA Annual Survey using data available for all hospitals in the annual survey and estimating the probability that a hospital responded to the AHA IT Supplement based on those characteristics.”

Representatives from 2174 hospitals responded to survey questions on AI.

“Representatives from 2174 hospitals responded to survey questions on AI.”

Hospitals that used Epic were far more likely to be early adopters or fast followers than hospitals using other EHR developers.

“Hospitals that used Epic were far more likely to be early adopters or fast followers than hospitals using other EHR developers.”

• Also stated in: § Conclusions · hedged

  “Hospitals that reported Epic as their primary EHR were far more likely to be early adopters or fast followers in generative AI, indicating that hospitals may be relying on their EHR systems to facilitate AI use.”

Hospitals may be relying on their EHR systems to facilitate AI use.

“Hospitals that reported Epic as their primary EHR were far more likely to be early adopters or fast followers in generative AI, indicating that hospitals may be relying on their EHR systems to facilitate AI use.”

The strength of these associations, combined with the substantial share of hospitals that use the 3 largest vendors in our data, may indicate that a handful of EHR developers have substantial ability to influence competition among health AI firms.

“The strength of these associations, combined with the substantial share of hospitals that use the 3 largest vendors in our data, may indicate that a handful of EHR developers have substantial ability to influence competition among health AI firms.”

This likely indicates that established competencies ensuring the effectiveness of AI are important to adopting new technology.

“This likely indicates that established competencies ensuring the effectiveness of AI are important to adopting new technology.”

We observed the continuation of a digital divide across many hospital characteristics in adoption of generative AI.

“We observed the continuation of a digital divide across many hospital characteristics in adoption of generative AI.22”

• cites [22] Nong P, Adler-Milstein J, Apathy NC, Holmgren AJ, Everson J. Current use and evaluation of artificial intelligence and predictive models in US hospitals. Health Aff (Millwood). 2025; 44(1): 90- 98. doi: 10.1377/hlthaff.2024.00842 39761454 sourceNot checked — no full text of the source

The associations between hospital characteristics and generative AI use is attenuated by correlations with 2 important factors: predictive AI experience and the hospital's primary EHR developer.

“This is likely because the associations between hospital characteristics and generative AI use is attenuated by correlations with 2 important factors: predictive AI experience and the hospital’s primary EHR developer.”

Differences in generative AI adoption across hospital characteristics, primary EHR vendor, and experience evaluating predictive AI indicate that the digital divide observed with predictive AI may be replicated in the use of generative AI.

“Differences in generative AI adoption across hospital characteristics, primary EHR vendor, and experience evaluating predictive AI indicate that the digital divide observed with predictive AI may be replicated in the use of generative AI.22”

• cites [22] Nong P, Adler-Milstein J, Apathy NC, Holmgren AJ, Everson J. Current use and evaluation of artificial intelligence and predictive models in US hospitals. Health Aff (Millwood). 2025; 44(1): 90- 98. doi: 10.1377/hlthaff.2024.00842 39761454 sourceNot checked — no full text of the source

EHR developers appear central to the adoption of generative AI.

“Third, EHR developers appear central to the adoption of generative AI.”

Three factors differentiated early adopter, fast follower, and delayed adopter hospitals in multivariable analysis.

“Three factors differentiated early adopter, fast follower, and delayed adopter hospitals in multivariable analysis.”