About 71 percent of non-federal acute care hospitals in the United States reported using predictive AI built into their electronic health record (EHR) in 2024, up from 66 percent in 2023, according to a September 2025 federal data brief. The fastest growth was in billing and scheduling, not in treatment decisions, and small, rural, and independent hospitals lagged. Most hospitals say they check their models for accuracy and bias, but far fewer do so for every model.
Key takeaways
- The data come from a hospital survey run with the American Hospital Association, so they describe self-reported use, not measured benefit.
- Among hospitals that use predictive AI, use for billing rose from 36 percent to 61 percent between 2023 and 2024, and use for scheduling rose from 51 percent to 67 percent.
- System-affiliated hospitals reported 86 percent adoption in 2024, compared with 37 percent for independent hospitals.
- In 2024, 82 percent of hospitals evaluated models for accuracy and 74 percent for bias, though 21 percent answered "don't know" on bias.
- The survey definition of predictive AI is broad, so it does not show how many hospitals use newer generative tools.
What the 2023-2024 survey counted
The Assistant Secretary for Technology Policy (ASTP) in the U.S. Department of Health and Human Services published Data Brief No. 80 in September 2025. It draws on the 2023 and 2024 American Hospital Association Information Technology Supplement, a survey sent to hospital leaders and usually answered by the chief information officer. The 2024 survey ran from April to September 2024, and 51 percent of eligible hospitals responded. Answers were weighted to adjust for who replied.
The brief defines predictive AI as machine learning and other predictive models that classify patients or produce a risk score, such as readmission risk, early detection of a condition, or appointment no-shows. The question also covered non-machine-learning models, so the headline figure is wider than "machine learning" alone. It covers tools shown inside the EHR, and it says nothing about chatbots or note-drafting tools that sit outside that system.
Where the tools are used
Predicting health trajectories or risks for inpatients was the most common use, and it stayed steady over the year. The growth was elsewhere. Among hospitals using any predictive AI, reported use for simplifying or automating billing rose by 25 percentage points, scheduling by 16 points, and identifying high-risk outpatients for follow-up by nine points. Use for recommending treatments rose by only two percentage points. The authors suggest the low figure for treatment recommendations may reflect a higher risk of errors in that setting, which is their reading and not a measured result.
Hospitals also differ in where their models come from. In 2024, 80 percent used models from their EHR developer, 52 percent used third-party models, and 50 percent had built some themselves. Hospitals could pick more than one source.
Who has adopted and who has not
The table shows 2024 adoption of predictive AI by hospital group, taken from the brief's Table 1.
| Hospital group | Used predictive AI in 2024 |
|---|---|
| Small (fewer than 100 beds) | 59 percent |
| Large (400 or more beds) | 96 percent |
| Rural | 56 percent |
| Urban | 81 percent |
| Critical access | 50 percent |
| Independent | 37 percent |
| Part of a multi-hospital system | 86 percent |
The brief calls this an early sign of a persistent digital divide. Hospitals using the market-leading EHR vendor reported 90 percent adoption, against 50 percent for hospitals on other vendors, so access to AI follows the software contract as much as clinical need.
Checking and governing the models
In 2024, 82 percent of hospitals reported evaluating predictive AI for accuracy using their own data, 74 percent for bias, and 79 percent reported post-implementation monitoring. Overall rates were flat from 2023, but the share evaluating all or most models for bias rose from 44 percent to 57 percent. The share answering "don't know" was 15 percent for accuracy and 21 percent for bias, which suggests that some respondents could not say what happens to every model in use.
Governance was usually shared: 74 percent named more than one accountable group. A committee or task force dedicated to predictive AI was named by 66 percent, and information technology staff were named by 41 percent, the lowest of the listed groups. Evaluation rates were lower at independent hospitals, where 41 percent evaluated models for accuracy, than at system members, where 76 percent did.
What the survey cannot show
The survey records whether a hospital uses a tool, not whether patients did better. It relies on one respondent per hospital, and the response rate was about half. Because the definition includes older statistical models, a rise in adoption may partly reflect better awareness of tools that were already installed. Hospitals outside the United States are not covered, and Canadian figures would need a separate source.
- Ask which of a hospital's models were checked on local patient data, not only on the vendor's data.
- Ask who is accountable when a model's performance drifts.
- Ask whether the use is administrative or clinical, since the risks differ.
This article reports survey statistics for general information. It is not medical advice, and it does not rate any hospital or product. Questions about personal care belong with a clinician, and questions about a specific hospital's practices belong with that hospital.
FAQ
Does 71 percent mean most hospitals use AI to diagnose patients?
No. The figure covers any predictive model shown in the EHR, including billing, scheduling, and risk scores. Inpatient risk prediction was the most common use, and treatment recommendations were much less common.
Does the survey include chatbots and generative AI?
Not directly. The questions concern machine learning and predictive models displayed in or launched from the EHR. The authors only note that administrative use of generative AI may be influencing the growth in billing and scheduling tools.
Why do independent hospitals report lower adoption?
The brief documents the gap but does not test causes. It points to differences in size, system membership, and EHR vendor as characteristics that go with lower use.
Does evaluating a model for bias guarantee fair results?
No. The survey asks only whether a hospital evaluated models, not how well or against which standard. It also found that many hospitals evaluated only some of their models.
Related reading
- How computer vision is used in practice
- AI applications across other industries
- What artificial intelligence means
- How artificial neural networks work
More in AI in health.
Last verified: October 2026
Sources
- Chang W, Owusu-Mensah P, Everson J, Richwine C. Hospital Trends in the Use, Evaluation, and Governance of Predictive AI, 2023-2024. ASTP Data Brief No. 80, U.S. Department of Health and Human Services, September 2025
- American Hospital Association Information Technology Supplement, 2023 and 2024 (data source cited in the brief)