Special Topic: AI for the 25-bed hospital

Show Notes

A small rural hospital should buy the few AI tools that fill a gap in its staffing or clearly pay their way, and leave alone whatever it has no one to check.

Read the special topic and its sources: https://physicianintheloop.org/specials/ai-for-the-25-bed-hospital/

The voice on this episode is generated by AI. It reads a special topic from Physician in the Loop as written; every fact in it links to its source.

What is Physician in the Loop?

How technological innovation is changing medicine for patients and the people who practice on the front lines.

This is a Physician in the Loop special topic.

AI for the 25-bed hospital

A small rural hospital should buy the few AI tools that fill a gap in its staffing or clearly pay their way, and leave alone whatever it has no one to check.

CommonSpirit Health, which runs 136 hospitals in 24 states and 20 more through joint ventures, says artificial intelligence has brought it $200 million in financial value. Its chief financial officer told investors on Oct. 2 that a 36-member governance committee had reviewed more than 122 proposed uses and rejected 20, over bias, safety risks and vendor redundancy, and that its clinical ethics team holds "an absolute veto on any deployment, regardless of financial return." Becker's account gives no breakdown or definition of the $200 million, and the figure does not appear in the system's annual report.

A hospitalist in a 25-bed critical access hospital has no committee of 36. The hospitalist may be the only physician in the building overnight, admitting from an emergency department staffed by one doctor, managing the sickest patients without an intensivist and waiting hours to transfer the ones who need more. In a Society of Hospital Medicine survey published in 2018, half the rural hospitalists who served as their ICU's primary physicians said they felt obliged to practice beyond their scope, and nine in ten said they at least occasionally had too little support from intensivists. For hospitals like these the question is which two or three tools are worth having. The answer is to buy narrowly, where a tool fills a gap in staffing or the payment is clear, and to skip whatever the hospital has no one to check, however good the demonstration looks.

The money works differently in these hospitals. A critical access hospital has no more than 25 inpatient beds, keeps patients four days on average or less and, unless its state exempted it before 2006, must sit more than a 35-mile drive from the nearest hospital. Medicare pays it 101% of its reasonable costs, less a 2% budget cut, and its payments "are not based on the type or number of services provided." Software that codes charts more completely, the pitch behind much hospital AI, therefore does not raise what traditional Medicare pays such a hospital for its own inpatient and outpatient services; any gain there comes from other payers, and from professional fees, which Medicare still pays by its fee schedule. At the median critical access hospital, outpatient services bring in 85% of revenue, and about four in ten rural hospitals lose money on operations.

Small hospitals already lag. In 2024 half of critical access hospitals used predictive AI, against 80% of other hospitals, and independent hospitals trailed those in systems by 37% to 86%. Most of what they use arrives inside the record system: 80% of hospitals using predictive AI took models from their record vendor. Critical access, rural and other hospitals with fewer resources are buying these products "off the shelf," said the lead author of a study of hospitals' survey answers, which can mean they were designed for patients who look very different from their own. A model that misfires costs a small hospital more than a big one. When Michigan Medicine tested the first version of Epic's sepsis alert, in a study published in 2021, it fired on nearly one patient in five, and a clinician would have had to evaluate eight alerted patients to find one case of sepsis. Epic has since replaced that model and, STAT reported, recommends training it on a hospital's own data before clinical use, work a small hospital may have no one to do.

The tools with the best record in small hospitals mostly bring a distant clinician to the bedside, and AI is usually a small part of them. A 2017 study of a tele-emergency service run from Sioux Falls for 85 small rural hospitals found that it avoided about 1,200 transfers over 52 months, by its clinicians' estimate, saving patients money, although it did not save the hospitals any. Observational studies link remote intensivist coverage to lower mortality in the ICU and the hospital, although in a national study of Medicare patients the benefit was concentrated in large urban hospitals. A remote pharmacist reviewing orders after hours for six rural hospitals flagged potential medication errors in about one patient in five, and a tele-nocturnist program run by the Medical University of South Carolina for four rural hospitals reported, in a conference abstract, better quality at lower total cost. AI that rides on these services, such as software that flags a stroke on a scan for the radiologist reading it remotely, is worth more in a 25-bed hospital than in a tertiary center, because it stands in for a specialist who is not there. Evidence from rural hospitals themselves is thin, so a small hospital should ask for it rather than assume it.

Ambient scribes, the software that listens to a visit and drafts the note, are spreading quickly. A study of more than 1,800 clinicians who used them at five academic medical centers, compared with 6,770 who did not, linked the scribes to 16 fewer minutes a day of documentation and half a visit more a week. The added revenue, $167 a month per clinician, was statistically significant but "nominal," in Mass General Brigham's words; the authors wrote that the estimate could be read as a conservative lower bound. In a randomized trial by researchers at UCLA, one product cut time spent in notes by 9.5% and the other had no significant effect, according to the maker of the first. A research firm for private-market investors puts prices at $99 to $600 a month per clinician. None of this evidence comes from rural hospitals. In a 25-bed hospital the case for a scribe rests on keeping clinicians, not on billing. That can be a good case where losing one physician can leave the nights uncovered, but it should be tested as one, by whether clinicians stay.

Other tools are luxuries for now. Prediction scores for deterioration, readmission or sepsis need someone to validate them on local patients, watch them and act on them; in a large study of a sepsis alert at five hospitals, patients did better when a clinician confirmed the alert within three hours. Guidance from the Joint Commission and the Coalition for Health AI asks hospitals for a formal governance structure, local validation and ongoing monitoring scaled to the setting. A 25-bed hospital can name one person to govern three tools. It cannot govern thirty, and features switched on and left unwatched are not free.

The objection is that this counsels timidity and will widen the gap between rural and urban care. Rural patients deserve the same tools, the need is greatest where expertise is thinnest, and grant money is arriving: the federal Rural Health Transformation Program is distributing $50 billion over five years, and its permitted uses include training and technical assistance to adopt artificial intelligence and remote monitoring, and software and hardware for information technology. That objection mistakes narrow for timid. A tool that stands in for the missing intensivist at three in the morning is the most ambitious purchase a small hospital can make. And the program's money ends in 2030, while a subscription does not; Chartis, a consultancy, says the program is "unlikely to offset the impact of Medicaid cuts" made by the 2025 budget law. Big systems' governance can be borrowed rather than built. A North Carolina network launched last month with a $4.4 million grant to help rural and resource-limited hospitals choose and safely implement AI, because, as one of its leaders put it, large systems have teams to evaluate and govern AI tools and "most rural hospitals and small practices do not."

From the hospitalist's side of the bed, the rule is to start from the gap rather than the product. The hours and jobs a small hospital cannot staff, such as nights, ICU coverage, pharmacy review and imaging reads, are where software earns its keep, and the best buys keep a clinician behind every alert: remote intensivist, pharmacist and radiology services that use AI to sort their work. Claims about payment deserve suspicion. Better coding does not raise what traditional Medicare pays a critical access hospital for its own services, so any gain must come from other payers or from professional fees, and the hospital's cost-report accountant should say how the expense itself will be treated before the contract is signed.

Just as important, a small hospital should switch on only what someone will watch. Every predictive feature in the record system should have a named person who reviews its alerts and outcomes each month, a feature no one will watch should be switched off, and the record vendor should be asked how its models perform at hospitals of the same size. And one-time money should go on one-time things. Grants are better spent on connections, hardware and training, which last, than on subscriptions the hospital cannot carry once the grant runs out.

CommonSpirit can afford a committee of 36 to sort more than 122 ideas. A 25-bed hospital should pick the two or three that fill its gaps and make sure someone is watching them.

This special topic was read by an AI voice. Its sources are linked at physicianintheloop.org.