Hospital AI and dementia admissions: which tool, not whether

Design

Cross-sectional study of linked national claims, beneficiary and hospital IT survey data, multivariable regression

Population

340,509 Medicare fee-for-service beneficiaries aged 65+ with dementia and at least one hospitalisation in 2023

Primary outcome

Frequent hospitalisation, 30-day readmission, preventable admissions, Medicare payments and out-of-pocket spending

Effect

Greater AI adoption associated with lower frequent hospitalisation, readmission and preventable acute admission; treatment-recommendation tools associated with higher patient out-of-pocket spending

Hospitals are buying artificial intelligence tools faster than anyone is evaluating them, and older people with dementia are the population where the claims are loudest — high readmission rates, many potentially avoidable admissions, plenty of room to improve. This cross-sectional study linked 2023 Medicare inpatient claims for 340,509 beneficiaries with Alzheimer's disease and related dementias to a national survey of what their hospitals had actually installed.

The headline is favourable but the detail is the point. Greater adoption of patient-related AI tools was associated with lower odds of frequent hospitalisation, 30-day readmission and preventable acute admission. Broken down by function, risk-prediction tools for inpatients and tools identifying high-risk outpatients were consistently associated with lower utilisation; inpatient risk prediction was also associated with lower total Medicare spending, while high-risk outpatient identification was associated with higher spending — plausibly because finding more high-risk patients generates more care. Treatment-recommendation tools were associated with higher out-of-pocket spending for the patient.

So 'does hospital AI help' is the wrong question, and this study is most useful as a demonstration of why. Tools that flag who needs attention behaved differently from tools that suggest what to do, and the one group that shifted cost onto the patient was the latter. Nothing here is causal — hospitals that adopt AI differ systematically from those that do not, in staffing, size and case mix — and a cross-sectional design cannot rule out that the better-organised hospitals both bought the software and had fewer readmissions anyway.

• Ask what function a proposed AI tool performs before asking what it costs.
• Distinguish risk-flagging from treatment-recommending when evaluating a vendor's claims.
• Ask who bears the cost of the extra care a tool generates, including the patient.
• Treat association between adoption and outcomes as a reason to run a local evaluation, not to skip one.

Why it matters

It replaces a yes-or-no question about AI with the one that determines what a hospital should buy.

Don't overread it

Cross-sectional and associational — this cannot show that adopting a tool causes fewer readmissions.

The statistics, in plain English

Cross-sectional means everything was measured at once, so there is no way to know whether the software came before the better outcomes. The confounding worry is specific and large: hospitals that invest in AI are on average larger, better resourced and more likely to have care-coordination programmes, all of which independently reduce readmissions. Adjustment for hospital characteristics helps, but it can only correct for what was measured. The finding that different tool types moved outcomes in different directions is the most credible part, because confounding by hospital quality would be expected to push all of them the same way.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.