Where Hospital AI Actually Lands — and Why That Is the Finding

Of 3,092 US hospitals that answered the 2023 American Hospital Association technology survey, about one in three — 35 percent — reported running AI-based predictive models. The interesting number is not that one. It is where those hospitals sit. Hwang and colleagues at Stanford put every facility on a map and asked a question accuracy benchmarks never reach: which patients live near a hospital that has these tools at all? The answer is a country of dense clusters and large blanks, and the blanks are not random.

For a physician this lands harder than any leaderboard. A model that helps only where help is already plentiful does not close a gap in care. It deepens it. So the right thing to study about a technology that is genuinely spreading is its distribution — and this preprint studies exactly that, with the appropriate humility about what a map can and cannot prove.

Design: a map, not a trial

This is an observational, cross-sectional analysis. The authors took the 2023 AHA Annual Survey and its information-technology supplement — which, for the first time, asked in detail whether a hospital uses predictive models — and graded each facility on a three-level scale: no predictive models, non-AI predictive models, or AI-based models. They joined this to community socioeconomic and shortage indicators and to 47 Centers for Medicare & Medicaid Services quality metrics measured quarterly from 2022 through mid-2025. Spatial statistics located the clusters; a random-forest model with SHAP values ranked what predicts adoption; and a longitudinal arm asked whether AI-adopting hospitals moved differently on quality over time. No patients were followed and nothing was randomised. Every result is an association.

What the data show

Adoption is sharply clustered — significant spatial autocorrelation on every implementation measure, with identifiable hotspots and coldspots rather than an even frontier. And the clustering runs the wrong way relative to need. Comparing implementation against need indicators, the authors report mismatch rates from 67.3 percent (social vulnerability) up to 72.4 percent (areas medically underserved for elderly and infant care). Read plainly: across most measures, more than two-thirds of deployment is misaligned with where the need is greatest, and the largest concentrations sit in the two worst quadrants — high need with low AI, and low need with high AI.

The predictor ranking is the part to keep. The single most influential factor was not bed count, ownership or community wealth. It was interoperability. The top SHAP feature was the Core Index, a measure of a hospital's ability to exchange health information; the inverse-signed Friction Index, capturing barriers to that exchange, was next, followed by delivery-system model, bed capacity and a minority-status indicator. Higher Core Index, more AI. And in the geographically weighted regression, Core Index was the one predictor that held a positive sign in every region while others flipped between positive and negative. Interoperability is where the relationship is stable; everything else is local. The quiet thesis is that AI settles where the data can already move.

What it cannot tell you

The quality arm is where the careful reader slows down. Of 20 metrics, 12 showed statistically significant trajectory differences for AI-adopting hospitals after correction. Some ran favourable — heart-failure 30-day mortality (slope difference -0.20) and pneumonia mortality (-0.32) drifted down, with fewer excess acute-care days after discharge. Others ran the wrong way: emergency-department median times lengthened (+1.16), more patients left the ED before being seen, and sepsis-care measures worsened (SEP-1 +0.91). That split is more believable than a clean win would be — and the authors are right to read it as association, not effect. Hospitals that adopt predictive AI differ from those that do not in staffing, case mix and baseline digitisation, any of which could move both adoption and outcomes. The study cannot separate the tool from the kind of hospital that buys it; its own random-forest model reached only an AUROC (area under the receiver-operating-characteristic curve) of 0.67, which the authors note reflects exactly such unmeasured confounding.

Two limits are structural and worth naming. The exposure is a survey checkbox: that a hospital "uses AI-based predictive models" says nothing about which model, for what task, on whom, or whether anyone acts on its output — the authors concede their categories are too coarse to tie to any specific outcome. And the unit is the hospital, not the patient, so inferring individual benefit from facility-level correlations would be an ecological overreach. None of this sinks the work. It fixes the ceiling on what the work can claim.

Why it travels

The figures are American and the survey instrument is American; neither maps onto a German or European system unedited. The mechanism does. If the ability to exchange data is the gate that decides which hospitals can use AI at all — and the most spatially consistent signal in the paper says it is — then the unglamorous work of standards, structured records and moving a discharge letter without a fax is not merely a precondition for AI. It is what determines whether AI ever reaches the places that need it most. The risk this preprint documents is not that the technology fails. It is that it succeeds unevenly and hardens a divide that existed before it arrived. For anyone allocating a digitisation budget, that is the case for funding the connections before the algorithms.

Source: Hwang Y-M, Ng MY, Pillai M, Sahai MP, Hernandez-Boussard T. AI Implementation in U.S. Hospitals: Regional Disparities and Health Equity Implications. medRxiv, posted 28 June 2025; doi:10.1101/2025.06.27.25330441; funded by The SCAN Foundation, no conflicts of interest reported. An unrefereed preprint with an observational, hospital-level design — its findings are associations, not causal effects; a peer-reviewed version has since appeared in Nature Health.