Abstract
Hospital artificial intelligence (AI) and robotics are spreading unevenly across the United States, yet national evidence on how these technologies are associated with hospital performance and population health remains limited. This observational study linked the 2024 American Hospital Association Annual Survey, capturing calendar-year 2023 adoption across 6,166 hospitals, to CMS, CDC, and County Health Rankings outcomes for 3,143 U.S. counties. Conditioning on 2019 baseline performance, hospitals using AI for staff scheduling had roughly 4% higher adherence to the SEP-1 sepsis bundle, and hospitals using AI for routine-task automation had approximately 5.1% lower 30-day pneumonia mortality. At the county level, the clearest timing-aligned outcome was contemporaneous hospital-setting mortality (2023): any access to workflow AI hospitals was associated with 25.5 fewer hospital deaths per 100,000 residents under cross-fitted doubly robust estimation (9.9% lower; non-cross-fitted estimate: 16.9 fewer per 100,000). A historical county benchmark (County Health Rankings premature mortality, 2020–2022) showed a concordant pattern, with the cross-fitted estimate corresponding to 975.8 fewer years of potential life lost per 100,000 before age 75 (p < 0.001), though this outcome predates the 2023 adoption measurement and should be interpreted as a lagged benchmark rather than a contemporaneous association. Workflow AI access was also associated with about 206 fewer preventable hospital stays per 100,000 residents (7.2% lower). Robotics showed narrower and more mixed associations. Yet access reflected a substantial digital divide: only 65.8% of Americans lived within 30 minutes of AI-enabled care, leaving about 114.6 million people outside that catchment (AI access Gini = 0.740), and despite a 56% increase in AI-enabled hospitals between 2022 and 2024, distributional inequality persisted. Overall, the evidence is consistent with adjusted associations between workflow AI access and better rescue outcomes where adoption occurs.
Significance Statement In this study of 6,166 hospitals and 3,143 U.S. counties, hospitals using AI for staff scheduling had roughly 4% higher adherence to the recommended sepsis care bundle, and cross-fitted doubly robust estimation associated workflow AI with 25.5 fewer contemporaneous hospital deaths per 100,000 residents (9.9% lower than comparable unexposed counties; non-cross-fitted baseline: 16.9 fewer per 100,000). A concordant historical benchmark (premature mortality, 2020–2022) and a 7.2% reduction in preventable hospital stays reinforce the pattern. Yet about 114.6 million Americans live beyond a 30-minute drive of any hospital using AI. Only 65.8% of Americans lived within 30 minutes of AI-enabled care (AI access Gini = 0.740), and despite a 56% increase in AI-enabled hospitals between 2022 and 2024, distributional inequality did not improve. The contribution is not a blanket claim that AI saves lives; it is a national estimate of health disparities created when hospital technologies diffuse unevenly, suggesting that extending AI-enabled care to underserved areas may represent a high-return intervention.
1 Introduction
Hospital artificial intelligence (AI) and robotics are being deployed unevenly across the United States. Large, digitally mature health systems tend to adopt these tools first, while many rural communities remain far from AI-enabled or robotics-enabled care 1,2. Yet national estimates of what these technologies are worth to care delivery—and what populations lose when adoption lags—remain limited.
This gap matters because the practical promise of hospital AI and robotics is operational, not merely symbolic. Hospitals face persistent staffing strain, burnout, and coordination failures in time-sensitive care 3,4. While visionaries have described a more efficient and humane future driven by AI 5,6, empirical evidence linking specific hospital AI and robotics capabilities to rescue-process adherence, mortality, and access remains scarce. Recent work describes the uptake of generative AI in electronic health records 7, but national evidence on measured healthcare value and uneven access is still emerging. This intervention-oriented framing aligns with recent work arguing that algorithmic systems in social settings should be evaluated by how they change organizational decisions and downstream outcomes rather than by predictive performance alone 8.
This gap also has an equity dimension. Prior work has shown that algorithmic systems in health care can reproduce or amplify disparities when design choices, implementation resources, or access are uneven 9,10. The present question is therefore not only whether hospitals that adopt AI perform differently, but whether uneven diffusion creates another layer of geographic inequality in who can plausibly reach AI-enabled care.
To the best of our knowledge, this is the first national analysis to estimate both the healthcare value and the access inequality of hospital AI and robotics while conditioning on historical baseline performance. The study links the 2024 American Hospital Association (AHA) Annual Survey, covering 6,166 hospitals, to Centers for Medicare & Medicaid Services (CMS), County Health Rankings, and CDC mortality outcomes for 3,143 U.S. counties. It distinguishes between cognitive workflow AI tools that automate information processing and robotics that support physical procedures. Here, “wealth-proxy bias” means that technology adoption can stand in for pre-existing hospital resources, market power, or quality rather than for technology value itself. Conditioning on 2019 baseline outcomes therefore asks whether 2023 adopters outperform otherwise similar non-adopters relative to their own historical starting points, rather than merely reflecting the fact that better-resourced hospitals adopt first.
At the county level, the empirical median of workflow AI exposure is zero, so the main estimand is most cautiously read as an extensive-margin contrast between any access and none, supplemented by continuous-treatment and spatial robustness checks. The primary estimates reported in the body of this paper therefore emphasize interpretable adjusted mean differences on the observed scale, while Double Machine Learning (DML), Targeted Maximum Likelihood Estimation (TMLE), overlap weighting, and spatial models are used as triangulating robustness checks rather than as replacements for that primary estimand.
Theoretical Framework
We interpret the results through the Resource-Based View (RBV) of the firm 11,12. In this setting, hospital AI workflow tools and robotics can be understood as organizational resources that may improve performance when hospitals can integrate them into care delivery. RBV is used here as an interpretive theoretical lens for understanding value realization after adoption rather than as a second empirical claim about diffusion.
Research Questions
RQ1 (Mechanism)
To what extent are the hospital AI workflow capabilities and robotics observed in the AHA survey associated with rescue and process outcomes (SEP-1, pneumonia mortality, and procedural safety) net of baseline performance?
RQ2 (Scaling)
Is access-weighted exposure to AI-enabled and robotics-enabled hospitals associated with lower county-level premature mortality?
RQ3 (Adoption/Equity)
How unevenly are AI and robotics adopted across counties, and what access penalty does that create for rural populations?
2 Results
Sample and Adoption Landscape
Our analytical sample included 6,166 U.S. hospitals in the 2024 AHA Annual Survey and all 3,143 U.S. counties after county-level aggregation. Adoption was uneven: approximately 31% of counties had at least one hospital using AI for any purpose, while robotics adoption was present in about 45% of counties. Throughout the county analyses, the focal workflow AI exposure is AI for automating routine tasks (AHA item MO14), and the focal robotics exposure is in-hospital robotics (AHA item MO21). These measures are distinct from the hospital-level staff-scheduling AI measure used in the SEP-1 mechanism analysis. Workflow AI and procedural robotics are therefore treated as related but non-interchangeable technologies (Table 1). Fig. 1 provides a national three-panel overview of continuous access, policy-threshold access, and distributional inequality.
Panel A: Continuous access map using Census block-group proxy drive time to the nearest AI-enabled hospital in the contiguous U.S., where AI-enabled status is defined using the workflow AI routine-task automation measure (AHA item MO14, AI for automating routine tasks) aligned with the primary county mortality models. Panel B: Binary policy-threshold map indicating block groups within a 30-minute drive of an AI-enabled hospital; in 2023, 65.8% of the U.S. population was within 30 minutes, and approximately 114.6 million remained outside this access threshold. Panel C: Lorenz curves for population-weighted travel burden to AI- and robotics-enabled hospitals, showing high but differential inequality (AI Gini = 0.740; Robotics Gini = 0.776). Together, the three panels visualize national technology access deserts versus oases and their distributional consequences.
2.1 RQ1: Hospital-Level Mechanism (Confirmatory)
To reduce the risk that adoption simply proxies for institutional wealth or the baseline performance of already high-performing hospitals, the hospital analysis used Augmented Inverse Probability Weighting (AIPW) estimators 13 (see SI Section S7 for full specification and triangulation) while controlling for bed size, ownership, teaching status, county-level social determinants, and baseline 2019 performance. AIPW is the primary estimator because adoption is strongly selected on observable characteristics, and its doubly robust structure remains consistent if either the propensity model or the outcome model is correctly specified. Throughout Tables 2–3, the average treatment effect (ATE) is the adjusted mean difference between adopters and otherwise comparable non-adopters in the native units of each outcome; positive ATEs indicate higher values on beneficial measures, whereas negative ATEs indicate lower rates on adverse outcomes.
Operational Rescue: Sepsis and Pneumonia
Outcomes dependent on timely intervention (“rescue”) showed the strongest robust associations (Table 2).
AI for Staff Scheduling: Associated with a 2.24-point increase in adherence to the SEP-1 sepsis bundle (p < 0.05).
AI for Routine Tasks: Associated with a 0.87-point reduction in the hospital-level pneumonia mortality measure (ATE ≈ − 0.87; p < 0.05).
These findings are consistent with the idea that AI tools may improve workforce allocation, ensuring staff are available to execute time-sensitive protocols. It is possible that, by reducing administrative friction, these tools allow clinical teams to focus cognitive resources on patient stabilization. We interpret the SEP-1 result cautiously because SEP-1 is documentation-intensive and debated as a stand-alone quality signal 14–16. The workflow AI pattern is not confined to SEP-1 itself: AI Workflows also showed a positive association with the sepsis 3-hour bundle component (+1.43 points; q < 0.05), AI Routine Tasks with lower pneumonia mortality, and MO14’s county association with CT6 was moderated by OP-18b rather than SEP-1.
Surgical Precision (Robotics)
Robotics adoption was associated with a significant 0.03-point improvement in the inverted Patient Safety Indicator 15 (Accidental Puncture or Laceration) measure—that is, fewer puncture/laceration events after direction was harmonized so positive coefficients indicate better safety (q < 0.05). However, associations with broader vascular outcomes like Deep Vein Thrombosis (PSI-12) were not statistically significant after adjustment. This highlights the specificity of the mechanism: robotics aids procedural precision, while workflow AI aids systemic process adherence.
Post-hoc baseline checks
To probe reverse causality (e.g., that high-performing hospitals simply adopt AI), we ran post-hoc baseline checks that regressed 2023 AI adoption (from the 2024 AHA release) on 2019 baseline outcomes. For our primary mechanisms (AI Staffing → SEP-1; AI Routine Tasks → Pneumonia), there was no significant association between 2023 adoption and 2019 performance (p > 0.10, see SI Table S2). This supports the interpretation that the benefits observed in 2023 are not merely reflections of historical quality but are consistent with newer operational capabilities. As an additional check, we tested hospital-level outcomes with no plausible rescue-process mechanism: elective hip/knee complication rates and colonoscopy follow-up. None showed significant associations with AI or robotics adoption (SI Table S15, Panel B), supporting causal specificity.
Bridging hospital mechanisms to population mortality
To probe whether the proposed operational-rescue pathway extends beyond SEP-1 documentation to measurable mortality outcomes, a two-stage bridge linked the focal workflow AI measure—AI for automating routine tasks (MO14)—and robotics (MO21) to hospital mechanism metrics and then to county mortality. In the first-stage models, MO14 was associated with lower 30-day pneumonia mortality (β = − 0.57 per percentage-point; q < 0.001), and MO21 was associated with faster ED throughput (β = +16.2 minutes; q < 0.001). In the second-stage models, pneumonia mortality independently predicted county YPLL (β = − 59.2; p = 0.002). When the county outcome was contemporaneous 2023 all-cause premature mortality (CT5; YPLL per 100,000 before age 75), adding pneumonia mortality attenuated the MO14 coefficient by 20.1%, while adding the full mechanism set (SEP-1, PSI-90, pneumonia mortality, and ED throughput) attenuated MO14 by 19.8% and MO21 by 26.4%. Future bridge specifications that additionally control for hospital payer mix (Medicare/Medicaid share), safety-net designation, and nurse-to-bed staffing ratios could further test the stability of these attenuation patterns, particularly for robotics where referral-selection concerns are strongest. These bridge results complement the CT6 analysis: in the primary cross-fitted county model, workflow AI was associated with lower hospital-setting mortality (cross-fitted AIPW ATE = − 25.5 deaths per 100,000; p < 0.001; about 9.9% relative reduction), while a two-part model indicated that the extensive margin drives the CT6 signal (ATE = − 15.7) and the intensive margin among adopters was approximately null (OLS slope = 0.047, p = 0.50; GPS slope = − 0.009, p = 0.91). Separately, higher SEP-1 compliance was independently associated with lower CT6 (ATE = − 9.3; p = 0.012), and poorer ED throughput independently predicted higher CT6 (ATE = +28.3; p = 0.004). Because the significant OP-18b × MO14 interaction appears on CT6 rather than on SEP-1 itself, the mechanism-consistent pattern is not reducible to SEP-1 documentation alone.
Contemporaneous Hospital-Setting Mortality Associations (CT6)
To test whether the proposed operational-rescue pathway is detectable in an outcome measured in the same year as technology exposure, we constructed CT6, an age-adjusted hospital death rate per 100,000 county residents (ages < 75) for 2023 from CDC WONDER mortality tabulated by county of residence, age band, and hospital place of death. Counties with suppressed CT6 cells after tabulation were excluded rather than imputed in the bridge analyses; of 3,143 counties, 247 (7.9%) were excluded due to CDC suppression (death counts 1–9), predominantly rural counties with populations below 20,000. Excluded counties had lower median populations (8,412 vs. 27,891) and higher baseline YPLL (9,847 vs. 8,562), suggesting that suppression-related selection, if anything, removes counties where health outcomes are worse, biasing remaining estimates toward the null. Small-area estimation or Bayesian partial-pooling approaches could recover estimates for these suppressed cells in future work; the present analysis takes the conservative approach of completecase exclusion rather than imputation under untested distributional assumptions. In the primary county sample, the untreated mean CT6 was 257.5 hospital deaths per 100,000 county residents. Under 5-fold cross-fitted doubly robust estimation, counties with any exposure to workflow AI for routine task automation (MO14) exhibited 25.5 fewer hospital deaths per 100,000 (95% CI [ − 39.7, − 11.3]; p < 0.001), a 9.9% reduction relative to the control mean (effective sample size = 409.1; 99th-percentile weight = 19.6). The non-cross-fitted AIPW baseline yielded a more conservative estimate of − 16.9 per 100,000 (95% CI [ − 29.9, − 6.2]; p = 0.008; 6.6% reduction). Robotics exposure (MO21) was directionally similar (cross-fitted: 25.8 per 100,000; non-cross-fitted: − 13.0; 95% CI [ − 27.9, 0.1]; p = 0.060), while naïve OLS slopes were weakly positive for robotics and near-zero for workflow AI (Table 4), underscoring why adjustment for baseline quality and selection matters in this setting. Counties above the median SEP-1 also showed significantly lower CT6 (ATE = − 9.3; 95% CI [ − 15.4, − 2.7]; q = 0.032), and poorer ED throughput independently predicted higher CT6 (ATE = +28.3; 95% CI [10.7, 46.9]; q = 0.032). Because SEP-1 is partly documentation-sensitive and debated as a stand-alone quality measure, we do not interpret that result in isolation; instead, we read it jointly with pneumonia mortality, the sepsis 3-hour bundle, preventable hospital stays, ED throughput, and CT614–16. The OP-18b × MO14 interaction is positive and FDR-significant (β = +0.0037; p = 0.004; q = 0.011). Because lower OP-18b indicates faster ED throughput, this implies that MO14’s protective association is strongest in counties with efficient baseline ED operations (low OP-18b) and weakens as throughput deteriorates (higher OP-18b). Because this moderation appears on CT6, a CDC-derived hospital-death outcome rather than on SEP-1 itself, the mechanism-consistent pattern is not reducible to SEP-1 documentation alone. Together, these results suggest that the clearest timing-aligned county signal is hospital-setting mortality, where the proposed operational pathway is most direct.
We constructed CT6 alongside CT5 to distinguish a hospital-setting outcome from all-cause premature mortality. CT5 captures contemporaneous 2023 YPLL per 100,000 regardless of setting, whereas CT6 restricts to deaths occurring in hospitals, where care processes and “operational rescue” are most directly implicated. In the added sensitivity analyses, CT6 remained negative for both MO14 and MO21 under conventional cross-fitted clipping rules, whereas CT5 was directionally negative but substantially less stable even before overlap-focused reweighting. We therefore interpret CT6 as the clearest timing-aligned county outcome and CT5 as supportive but less stable. This contemporaneous county pattern is also complemented by a significant 7.2% reduction in Preventable Hospital Stays (ATE ≈ − 207 stays per 100,000; see SI Appendix, Table S15), suggesting that the operational signal is not confined to a single mortality outcome.
Stratification by rurality and market concentration
To test whether the extensive-margin county benefits are concentrated in particular settings, we stratified the primary MO14 → DV21 association by USDA rural–urban continuum codes and hospital market concentration (Herfindahl–Hirschman Index terciles). The negative association was present in both metropolitan and non-metropolitan counties, though the point estimate was larger in non-metropolitan settings (ATE ≈ − 1,180 YPLL, p = 0.02) than in metropolitan counties (ATE ≈ − 620 YPLL, p = 0.03), consistent with the extensive-margin interpretation that the first AI-enabled pathway into an otherwise unexposed catchment carries the strongest marginal association. In high-concentration markets (HHI top tercile), the association was directionally negative but imprecise, likely reflecting the smaller treated sample within single-system markets.
2.2 RQ2: Population Health Scaling (County Level)
We next asked whether county exposure to hospitals using workflow AI and robotics was associated with population health. We constructed an access-weighted measure: the share of the county population living within a 30-minute drive of an adopting hospital, weighted by adjusted patient days. This lets a county served by a large adopting hospital differ from one served by a small adopter.
Because the timing-aligned county evidence is strongest for CT6, we treat the historical DV21 analysis as a lagged population-health benchmark rather than the sole anchor of inference. In that benchmark model, cross-fitted doubly robust estimation associated any exposure to AI for automating routine tasks (MO14) with 975.8 fewer years of potential life lost (YPLL) per 100,000 population (95% CI [ − 1378.8, − 572.9]; p < 0.001). The non-cross-fitted AIPW baseline yielded a more conservative estimate of − 457.5 YPLL (95% CI [ − 882.7, − 111.5]; p = 0.004), corresponding to a 4.5% reduction relative to the control-group mean of 10,209.5 YPLL/100k. Fig. 2 displays the county-level ATEs across all major technology components and shows that routine-task automation (MO14) and in-hospital robotics (MO21) carry the most negative historical mortality estimates. Continuous-treatment estimators remained directionally negative. A formal two-part (hurdle) specification supports the extensive-margin interpretation: the participation effect for any MO14 exposure versus none was − 510.3 YPLL, while the intensive-margin slopes among adopters were near zero (OLS slope = +1.02, p = 0.47; GPS slope = − 1.75, p = 0.32).
This plot shows county-level AIPW Average Treatment Effects (ATEs) after adjustment for baseline (2019) mortality and socioeconomic factors. Points to the left of zero indicate lower premature mortality relative to comparable unexposed counties. Routine-task automation (MO14) carries the most negative point estimate (p ≈ 0.004), consistent with the hospital-level rescue-process mechanism emphasized in the text.
The focal MO14 → DV21 association is concordant across estimator families (Table 5): the primary cross-fitted AIPW estimate ( − 975.8 YPLL; p < 0.001) is supported by TMLE ( − 412.7; p = 0.002), overlap weighting ( 201.6; p = 0.004), state-level spatial block bootstrap ( − 498.4; 95% CI [ − 902.4, − 173.4]), and the non-cross-fitted AIPW baseline ( − 457.5; p = 0.004). This concordance across estimators with different assumptions provides strong evidence that the negative association does not depend on a single nuisance-fitting strategy or estimator family.
Threshold sensitivity analyses show that the DV21 association is not driven by a single cutoff: binarizing MO14 at 10, 25, and 50 percentage points preserved a negative sign in all cases (Table S12). Overlap-weighted (ATO) estimates, which target the subpopulation with the strongest covariate overlap, remained negative for both MO14 and MO21 on the historical DV21 outcome (Table S17; Section S7.6). Alternative access constructions reported in the SI—15/30/45/60-minute catchments and unweighted, bed-weighted, or adjusted-patient-day-weighted exposure definitions—also remained negative, suggesting that the county association is not peculiar to a single 30-minute implementation (SI Table S13).
Continuous-treatment, overlap, and spatial robustness
Because county workflow AI exposure is highly zero-inflated (median = 0; only 14.3% of counties exhibit non-zero MO14 access), the primary county estimand is an extensive-margin contrast between any exposure and none. Continuous-treatment estimators (DML, stabilized GPS-IPW) were directionally consistent but less precise (SI Tables S10–S13). Overlap-weighted ATO estimates remained negative for DV21 and attenuated toward non-significance for CT5 and CT6.
Treatment-misclassification sensitivity
A probabilistic misclassification analysis varying sensitivity and specificity of the binary treatment indicator produced negative median ATEs across all six tested scenarios for MO14 → DV21 (medians: − 83 to − 591 YPLL per 100,000), though only the Se= 0.95/Sp= 0.95 scenario retained a 95% interval excluding zero. The MO14 → DV21 result is therefore directionally robust to plausible nondifferential misclassification, but quantitatively attenuated under label noise. Contemporaneous county outcomes (CT5 and CT6; see SI Table S6 and Table S17) were more sensitive to assumed misclassification.
Spatial Autoregressive (SAR) and Spatial Error Model (SEM) specifications using a queen-contiguity weight matrix (SI Table S18) confirmed that the AI-access coefficient remained negative and significant after absorbing spatial dependence (SAR: − 422 YPLL, ρ = 0.341; SEM: − 533 YPLL, λ = 0.487; both p < 0.001).
Exploratory Risk Attenuation
We then explored whether AI access changed the relationship between unhealthy community behaviors and premature mortality. As illustrated in Fig. 3, the slope relating the Health Behaviors Index to YPLL was steeper in counties with low workflow AI access and flatter in counties with high workflow AI access; the negative interaction term is statistically significant (p < 0.05). This pattern is consistent with workflow AI acting as a risk attenuator: unhealthy behaviors still matter, but their association with mortality is weaker in counties that have greater access to AI-enabled hospitals.
Predicted county premature mortality (YPLL) is plotted against the Health Behaviors Index using the fully adjusted county model, with separate curves for High versus Low workflow AI access. High access is defined as counties above the median of the continuous access-weighted exposure score Ai (Ti = 𝕀 { Ai ≥ median(A) }; see §4.3 and Section S1.2). Shaded bands show 95% confidence intervals around model-implied predictions. The flatter high-access slope visualizes the negative Health Behaviors × AI Access interaction (p < 0.05), indicating attenuation of the behavior– mortality association in high-access counties.
2.3 RQ3: Adoption and Access Inequities
Despite the potential benefits, our geospatial analysis quantifies a severe national access gap. Fig. 4 illustrates that AI adoption is clustered in urban areas.
Local Moranś I cluster map of county AI adoption. HH (red) denotes high-adoption counties surrounded by highadoption neighbors, HL (orange) high-adoption counties surrounded by lower-adoption neighbors, LH (light blue) low-adoption counties adjacent to higher-adoption neighbors, LL (dark blue) low-adoption counties surrounded by low-adoption neighbors, and ns (gray) counties without statistically significant clustering. Clusters are most common in well-resourced urban centers, illustrating a stark digital divide.
The Access Penalty
Using 2020 U.S. Census block-group population (N = 334.7 million) and a 30-minute drive-time proxy (operationalized as a circuity-adjusted 15.4-mile radius, equivalent to 20 road-distance miles), we quantified the “Access Penalty” (Table 6). Here, P10 denotes the 10th percentile of the population-weighted travel-distance distribution (the best-served decile), and P90 denotes the 90th percentile (the entry to the worst-served decile).
Robotics: 266.0 million people (79.5% of the population) live within a 30-minute drive.
AI Automation: Only 220.1 million people (65.8% of the population) have similar access.
Inequality is highly concentrated. The P90 distance to an AI-enabled hospital is 61.3 miles. However, the most remote population decile accounts for approximately 62% of the total national travel burden; within this worst-served decile, the median travel distance is 101.5 miles (mean ≈ 200.6 miles). For robotics, the disparity is similarly acute: the worst-served decile faces a median travel distance of 46.4 miles (mean 171.1 miles), accounting for 68% of the total travel burden. In practical terms, those upper-tail distances mean longer travel, delayed access to advanced services, and heavier reliance on non-adopting local facilities in the most remote counties.
Additional spillover and interference diagnostics are reported in the Supplementary Information.
Distributional inequality metrics
To summarize the full distribution of travel burden beyond percentile ratios, we computed inequality indices over the population-weighted blockgroup distance to the nearest facility in each category (Table 7). Inequality is extreme even for baseline access to any hospital (Gini = 0.829; Atkinson(ε = 0.5) = 0.661; CV = 6.65), indicating that U.S. hospital infrastructure is highly spatially concentrated. Distance inequality remains high for advanced technologies (Robotics: Gini = 0.776; AI: Gini = 0.740) and is not driven by zero-distance observations; imposing a small minimum distance (0.25 miles) changes Gini by less than 10−4 in all cases.
Importantly, access inequality is not unique to advanced technologies: baseline proximity to any hospital is already highly unequal (Gini = 0.829; Atkinson(ε = 0.5) = 0.661), reflecting the entrenched urban–rural spatial structure of U.S. hospital infrastructure. AI and robotics access remains sharply unequal (Gini = 0.740 and 0.776), and the much larger upper-tail disparities in Table 6 (P90/P10 ratios) indicate a distinct technology “access penalty” layered on top of pre-existing hospital deserts rather than replacing them. Notably, P90/P10 highlights upper-tail extremes, whereas Gini summarizes dispersion across the full distribution; AI access is worse in the far tail (Table 6) even though its overall dispersion is slightly lower than robotics (Table 7).
Equity interpretation of the access penalty
The inequality indices reinforce that the “access penalty” should be interpreted as a distributional phenomenon, not merely a difference in average travel distance. Even baseline hospital access is highly unequal, implying that technology diffusion occurs within an already polarized geographic infrastructure. Against this backdrop, the large upper-tail disparities for AI and robotics (Table 6) highlight a potentially policy-relevant margin: small improvements for already well-served populations do little to change the national burden, whereas interventions that reduce extreme travel distances for the most remote decile have outsized equity returns.
Multi-year analysis using both the 2022 and 2024 AHA releases reveals notable temporal dynamics. AI-enabled hospitals increased 56% (1,119 to 1,743), and population coverage within 30 minutes rose from 66.2% (CY 2022, from the 2023 AHA release) to 75.2% (CY 2024, from the 2025 AHA release). The primary cross-sectional analysis in this paper uses the 2024 AHA release (CY 2023), which yields 65.8% coverage—the intermediate wave in this three-wave comparison. However, the Gini coefficient for AI access increased slightly (0.739 to 0.767), indicating that expansion did not proportionally reach the most underserved populations. The P90/P10 distance gap for AI access narrowed substantially (49.8 to 33.2 miles, a 33% reduction), suggesting the extreme upper tail benefited, even as overall distributional inequality persisted. Robotics access was essentially unchanged across both metrics. These trends reinforce the “access penalty” interpretation: rapid adoption can compress the worst-case access distances while leaving the broader distributional structure intact (SI Table S19).
3 Discussion
The paper asks two linked questions: what is the measured healthcare value of hospital AI and robotics, and who actually has access to them? The answer from this national sample is narrower and more useful than a generic claim that AI saves lives. Specific workflow AI capabilities were associated with better rescue and process outcomes, robotics showed more targeted procedural benefits, and both technologies were adopted unevenly across the country.
3.1 What the results say
Workflow AI
For RQ1, the hospital-level evidence is strongest for time-sensitive rescue pathways. AI for staff scheduling was associated with higher SEP-1 adherence, and AI for automating routine tasks was associated with lower pneumonia mortality. These are operational associations, not proof that all AI improves all outcomes. The stratified results in the Supplementary Information also show that the SEP-1 association is not confined to a single hospital archetype: it remains positive in acute-care hospitals (ATE = +2.30), in the highest-volume quartile (ATE = +3.11), and in rural hospitals (ATE = +4.02; SI Table S5). Falsification analyses using outcomes unlikely to be affected by workflow AI (elective hip/knee complications, colonoscopy follow-up rates) showed no significant associations (SI Table S15, Panel B), supporting specificity rather than a generic quality confounder. Just as importantly, the workflow AI mechanism story is not SEP-1-only: it is reinforced by lower pneumonia mortality, a positive association with the sepsis 3-hour bundle, lower preventable stays, and a significant OP-18b × MO14 interaction on CT6, a CDC-derived hospital-death outcome rather than on SEP-1 itself.
Robotics
Robotics results require more nuanced interpretation. At the hospital level, robotics adoption was associated with lower 30-day mortality for CABG ( − 0.24; p < 0.01) and stroke ( − 0.30; p < 0.01), but also with a positive MRSA SIR association (+0.08; p = 0.01). One plausible explanation is case-mix selection: hospitals adopting surgical robotics are often high-acuity referral centers that treat more complex and immunocompromised populations at elevated infection risk. That interpretation is consistent with prior work showing that robotic adoption reallocates patients toward adopting hospitals and can change market share and case mix 17,18.
At the county level, robotics access (MO21) remains negatively associated with DV21 under both cross-fitted AIPW and overlap weighting (ATO ATE = − 185.5; p = 0.004), while the contemporaneous CT6 association is − 25.8 under cross-fitting (p = 0.001) but attenuates under overlap weighting (ATO ATE = − 1.8; p = 0.527). We therefore interpret robotics as a narrower procedural-precision signal rather than a blanket quality effect, with the MRSA result likely reflecting residual case-mix and referral-intensity confounding that should be addressed in future service-line-specific surgical-cohort analyses. A separate capital-intensity sensitivity check points in the same direction: winsorized log CAPEX intensity strongly predicted robotics adoption (β = 0.46, p < 0.001) but only weakly predicted GenAI adoption (β = 0.058, p = 0.048), suggesting that wealth-proxy concerns are sharper for robotics than for workflow AI.
County mortality and attenuation
For RQ2, counties with access to workflow AI and robotics were associated with lower premature mortality. The moderation results add an important nuance: the relationship between unhealthy community behaviors and mortality was flatter in counties with higher workflow AI access. That pattern is consistent with AI helping hospitals translate limited staff and operational capacity into more reliable rescue performance, although the design remains observational.
Equity and diffusion
For RQ3, adoption itself is part of the story. AI was present in only about 31% of counties, while robotics was present in about 45%. The travel-time gap to AI-enabled care was much larger than the gap to any hospital, and materially larger than the comparable gap for robotics. In other words, the communities farthest from digitally mature hospitals also face the largest access penalty.
3.2 Interpreting extensive versus intensive margins
Our primary county-level estimand is an extensive-margin contrast because the empirical median of workflow AI exposure equals zero. The new two-part models make that choice more explicit. For DV21, the extensive-margin participation effect for any MO14 access versus none was − 510.3 YPLL per 100,000, whereas the intensive-margin slopes among adopters were near null (OLS = +1.02, p = 0.47; GPS = − 1.75, p = 0.32). For CT6, the corresponding extensive-margin effect was 15.7 hospital deaths per 100,000, while the intensive-margin slopes were likewise approximately zero (OLS = − 0.047, p = 0.50; GPS = − 0.009, p = 0.91). Together, these hurdle results indicate that the clearest county-level signal is the first AI-enabled pathway into an otherwise unexposed catchment, not a smooth linear dose-response among already exposed counties. A complementary distance-based analysis nonetheless suggests that geographically grounded proximity can exhibit a gradient even when institutional-intensity measures within adopters do not (SI Table S11).
3.3 Looking at who gets access
This paper makes a stronger case when framed around uneven adoption and measured value, not around a blanket claim that AI saves lives. The national contribution here is that where hospitals adopt workflow AI and robotics, some outcomes look better; where they do not, patients face a measurable access penalty. The multi-year panel (2022–2024) underscores this point: despite a 56% increase in AI-enabled hospitals and a 9-percentage-point gain in population coverage, distributional inequality (Gini) did not improve—reinforcing that where new capacity is added matters as much as how much is added.
3.4 Positioning relative to prior AI diffusion studies
Recent equity-focused critiques argue that medical AI can reproduce or amplify disparities when deployment resources and access are uneven 9,10. That concern is central here: our question is not only whether hospital AI and robotics are associated with better outcomes where present, but whether diffusion patterns leave some communities outside the current catchment of AI-enabled care. The access-inequality findings complement conceptual critiques by quantifying the scale of the technology access gap and its association with mortality outcomes at the county level.
Recent work by Hwang et al. used the AHA IT Supplement to map predictive-AI adoption across U.S. hospitals. Their Nature Health paper 19 characterizes the national landscape of AI implementation and the institutional factors that predict adoption, while a companion preprint 20 extends that mapping to associations with county-level elder mortality and CMS quality metrics. Both analyses document the same urban-concentrated clustering pattern we report. Our study differs in three ways: it distinguishes specific AI and robotics capabilities rather than a single predictive-AI indicator; it conditions on 2019 baseline performance to reduce wealth-proxy bias; and it constructs population-weighted access measures that estimate who can reach adopting hospitals, not just where adoption occurs. The two approaches are therefore complementary rather than competing.
3.5 Why block-group access rather than HSAs or ZIP-code crosswalks
Some prior geospatial AI studies aggregate community characteristics to Dartmouth Atlas Hospital Service Areas or ZIP-code crosswalks 19,20. We instead measure circuity-adjusted distance from roughly 240,000 Census block-group population centroids to current hospital coordinates and then aggregate access to the county level. This design better matches the substantive question in this paper—who can plausibly reach AI-enabled care—and avoids relying on administrative boundaries that predate many of the closures, consolidations, and referral changes relevant to present-day diffusion. It remains a simplified access proxy rather than a full routing or gravity model; gravity-based formulations such as the enhanced two-step floating catchment area (e2SFCA) 21,22 could refine the access gradient, particularly in dense urban markets. Likewise, formal interference-aware estimands that separate direct from spillover effects within referral networks remain future work.
3.6 Limitations
Observational design and temporal alignment
This remains an observational study, so residual confounding cannot be ruled out. Later post-adoption mortality outcomes were not yet available in the source data; the primary County Health Rankings premature mortality outcome therefore spans 2020–2022 even though the AHA survey measures technology adoption in 2023. We therefore treat DV21 as a lagged benchmark and CT6 as the clearest timing-aligned county outcome. Explicitly controlling for cumulative county-level COVID-19 death rates (2020–2022) changed the focal MO14 → DV21 estimate from − 411.0 to − 452.1 YPLL per 100,000 on the restricted complete-case sample, suggesting that differential pandemic burden is unlikely to drive the association.
Estimator robustness and residual imbalance
The primary county estimates use 5-fold cross-fitted AIPW, which is the mathematically appropriate baseline when machine-learning nuisance models are employed. The cross-fitted DV21 estimate (− 975.8 YPLL) and CT6 estimate (− 25.5 deaths per 100,000) are more negative than the non-cross-fitted baselines (− 457.5 and − 16.9, respectively), which we retain for comparability with earlier literature. The directional magnitude shift under crossfitting is consistent with the removal of own-observation over-fitting bias in the Random Forest nuisance models: when out-of-fold predictions replace in-sample fits, the outcome model’s residual-based correction is less attenuated, producing larger absolute ATEs. In preliminary checks, alternative nuisance learners (logistic/ridge propensity, gradient-boosted outcome) yielded intermediate cross-fitted estimates, suggesting that the shift reflects regularization dynamics rather than a single hyperparameter choice. The effective sample size under cross-fitted AIPW was 409.1 with a 99th-percentile weight of 19.6, indicating that while overlap is limited by the zero-inflated exposure distribution, the estimate is not produced by a tiny handful of extreme-weight counties. Non-cross-fitted weight-tail diagnostics and covariate balance before and after weighting are reported in SI Tables S21 and S20 and Fig. S2; the cross-fitted ESS and 99th-percentile weight are reported in the main text. The negative associations are concordant across TMLE ( − 412.7), overlap weighting ( − 201.6), and state-level spatial block bootstrap ( − 498.4). A pooled baseline placebo-family test across all eight exposure measures suggested some residual pre-treatment imbalance (Fisher combined p = 0.004), although the focal primary exposure family (MO14 and MO21) did not show strong pooled pre-association (p = 0.102; SI Table S4). Because the primary county outcome is continuous YPLL, the reported E-value should be read as a heuristic sensitivity translation rather than the last word on unmeasured confounding, and we place more weight on cross-fitted, overlap, falsification, and spatial checks than on coefficient-stability summaries (Oster) or E-value summaries (VanderWeele–Ding) alone 23,24.
Spatial dependence and interference
County-level spatial dependence is a valid concern. Moran’s I on OLS residuals was 0.251 (p = 0.001), confirming residual spatial autocorrelation. SAR and SEM specifications (SI Table S18) absorb spatial dependence through explicit lag and error structures; the AI coefficient remains negative and significant across all spatial specifications. Spatial spillovers are partially addressed through these specifications and a neighbor-exposure decomposition (SI Section S9), but a formal interference-aware estimand that separates direct from spillover effects within referral networks remains future work.
Access measurement
The access construction is deliberately simple and interpretable: nearest-facility proximity within a 30-minute, circuity-adjusted catchment. Validation against Open-StreetMap routing for 500 randomly sampled block groups (stratified by rurality and terrain) showed a median network-to-Euclidean circuity ratio of 1.32 (IQR: 1.21–1.45), with mean absolute error of 3.8 minutes for the 30-minute threshold classification, RMSE of 5.2 minutes, and concordance (binary agreement on within/outside 30 minutes) of 91.4%. Error was modestly higher in mountainous census regions (median ratio 1.41) and lower in flat metropolitan areas (median ratio 1.27). The measure does not directly encode supply-demand competition, provider capacity, or hospital-specific AI depth. Future work should therefore compare this benchmark with enhanced twostep floating catchment and related gravity-style access formulations, which can incorporate distance decay and provider capacity more directly 21,22, as well as with full network routing in especially rural or mountainous regions. The AHA survey is also self-reported, so implementation depth almost certainly varies across hospitals. We do not provide external validation for a subsample. However, a probabilistic misclassification sensitivity analysis (SI Section S7) confirmed directional robustness of the MO14 → DV21 association across tested scenarios, though statistical significance was preserved only under the most favorable assumptions, consistent with classical attenuation from non-differential measurement error.
Robotics confounding and mechanism scope
For robotics in particular, positive naïve OLS slopes alongside negative adjusted AIPW estimates (Table 4) suggest that residual referral and case-mix selection into tertiary centers remains plausible; robotics is therefore interpreted as a narrower procedural association rather than a blanket mortality claim. Service-line-specific or acuity-stratified follow-up—for example, restricting to hospitals with high CABG or stroke volumes, or using instrumental-variable and selection-adjustment strategies based on differential robotics adoption timing across service lines—would be especially valuable in future work. Large-scale interventions can also shift patient composition and selective survival, so future causal designs should probe referral and compositional dynamics more directly 18. Finally, the paper measures hospital capabilities and geographic access, not the internal decision process by which hospital leaders choose to adopt AI or robotics. For that reason, the most accurate narrative is not “who personally decides to adopt AI,” but rather “where adoption occurs and who gets access to the resulting care.” With additional AHA waves, eventstudy style analyses could also better separate adoption timing from contemporaneous shocks.
3.7 Conclusion
The central contribution of this study is to quantify both the measured value of hospital AI and robotics where they are adopted and the access penalty imposed where they are not. This is not a blanket claim that AI saves lives. It is instead a national estimate showing that specific workflow AI capabilities are associated with better rescue-consistent outcomes where adoption occurs— staff-scheduling AI with 2.24-point higher SEP-1 adherence, workflow AI access with 25.5 fewer hospital deaths per 100,000 under cross-fitted estimation (9.9% reduction), and 975.8 fewer YPLL per 100,000 on the historical benchmark (p < 0.001)— while robotics results are narrower and more mixed, consistent with a procedural-precision rather than a blanket-quality signal. But those gains are not geographically neutral: AI was present in only about 31% of counties, only 65.8% of Americans lived within 30 minutes of AI-enabled care, and roughly 114.6 million people remained outside that threshold. Despite a 56% increase in AI-enabled hospitals between 2022 and 2024, distributional inequality did not materially improve. Because later post-adoption outcomes were not yet available, these estimates should be read as the best available 2023 benchmark rather than as a final causal statement. Even so, the pattern suggests that the largest public-health returns may come not only from improving hospital AI capability where it already exists, but from extending that capability into the communities that remain outside the current catchment.
4 Materials and Methods
4.1 Design and Assumptions
This is a lagged-outcome adjusted observational study. Identification relies on (i) conditional exchangeability: after controlling for baseline outcomes and covariates, technology adoption is as-if independent; and (ii) temporal precedence of stable hospital capabilities. Table 8 summarizes measurement timing and release timing, which motivates the use of 2019 baseline controls and the distinction between historical DV21 and contemporaneous CT5/CT6 mortality outcomes. The hospital-level analysis is prioritized (2023 adoption → 2023 outcomes conditional on 2019 baseline) because it offers the closest contemporaneous alignment available in the source data.
4.2 Data Sources
Timing clarification
We use the 2024 release of the AHA Annual Survey, which captures hospitals’ self-reported technology adoption during calendar year 2023. County Health Rankings 2024 compiles county covariates measured in 2023 and earlier, and our CDC WONDER and CMS/NBER outcome measures are also drawn from calendar year 2023. Accordingly, our key county exposures (MO14, MO21) and contemporaneous CDC outcomes (CT5, CT6) are measured in the same calendar year (2023), although reported in 2024 data products.
We integrated data from:
2024 AHA Annual Survey: Technology adoption for calendar year 2023 (N = 6, 166).
CMS Hospital Compare (released 2024): Outcomes for calendar year 2023 (SEP-1, Mortality, PSI).
County Health Rankings (2024): Contextual covariates (SES, Access) for 2023 and earlier.
4.3 Variables and Pre-specification
Outcomes
Two confirmatory hospital endpoints were pre-specified based on the operational-rescue hypothesis: SEP-1 bundle compliance and pneumonia mortality. All other hospital outcomes (e.g., PSI-15, ED wait times) are treated as exploratory; for readability, PSI-15 is reported in inverted direction so that positive coefficients denote fewer accidental punctures/lacerations. At the county level, DV21 denotes County Health Rankings YPLL per 100,000 (2020–2022), CT5 denotes CDC-derived 2023 all-cause YPLL per 100,000, and CT6 denotes 2023 hospital-setting deaths per 100,000 county residents younger than 75. Exposure: Hospital adoption was aggregated to the county level using an access-weighted approach: population share within 30 minutes of an AI-enabled hospital, weighted by hospital Adjusted Patient Days (adjpd) to reflect utilization volume. Two workflow AI measures anchor the main claims: staff-scheduling AI for the hospital SEP-1 mechanism analysis, and routine-task automation (MO14) for the county scaling analyses.
Robotics is captured by in-hospital robotics (MO21). For main county models, the continuous exposure was dichotomized to fit the primary doubly robust estimand. Because county MO14 access is highly zero-inflated (median = 0; only a small minority of counties exhibit non-zero exposure), the primary county estimand is the extensive margin of any exposure versus none; sensitivity analyses using continuous measures, alternative thresholds (>10/>25/>50 percentage points), and alternate 15/30/45/60-minute catchments are reported in the SI. Prespecification versus exploration: The core design of this study was prospectively specified in the first author’s doctoral dissertation proposal at Drexel University and formally reviewed by the dissertation committee before completion of the final linked analyses. That prospectively specified core defined the focal hospital AI/robotics exposures, the confirmatory hospital endpoints (SEP-1 and pneumonia mortality), the primary county extensive-margin estimands for MO14 and MO21 on DV21 and CT6, and the baseline-adjusted covariate strategy. The present manuscript therefore distinguishes those prospectively specified analyses from later reviewer-responsive robustness checks and manuscript-revision extensions. In practice, SEP-1 and pneumonia mortality are the confirmatory hospital endpoints; the focal county estimands are the extensive-margin MO14 and MO21 contrasts on DV21 and CT6. Additional hospital outcomes, moderation terms, change-score analyses, cross-fitted reestimation, estimator triangulation, and other sensitivity checks are reported as exploratory or robustness analyses unless explicitly noted otherwise.
Protocol provenance and ethics
This study used only archival secondary data: licensed hospital-survey data from the AHA and public-use hospital- and county-level outcome/covariate files from CMS, CDC, and County Health Rankings. The analyses involved no contact or intervention with human participants, and the analytic files contained no individually identifiable private information. Accordingly, the project did not involve human-subjects research requiring IRB review.
Formally, let Ai ∈ [0, 100] denote the continuous, access-weighted county AI exposure score (Section S1.2). We estimate a binary average treatment effect (ATE) for the primary extensive-margin estimand as Ti = 𝕀 {Ai > 0} (any exposure vs. none). For median-split sensitivity analyses, we define High-Access counties as Ti = 𝕀 {Ai > median(A) }. This preserves the underlying continuous exposure while keeping the main-text estimand aligned to the observed data-generating process. The county propensity and outcome models use the full pre-specified covariate set: 2019 YPLL, County Health Rankings indices for social/economic factors, physical environment, and health behaviors, Medicaid expansion status, log population, census-division indicators, and state-level leave-one-out spatial lags for the major AI/robotics exposure measures (MO11– MO15, MO21, MO1, and MO2). Because the model relies on composite County Health Rankings domains rather than every underlying subcomponent separately, Table 9 reports a compact main-text subset of the most substantively important confounders and their pre/post-weighting balance.
4.4 Statistical Analysis
The main text reports Augmented Inverse Probability Weighting (AIPW) estimates 25. AIPW is the primary estimator because adoption is observably non-random and strongly correlated with hospital and county resources; the estimator explicitly combines propensity-score reweighting with outcome regression while preserving an interpretable adjusted mean difference on the native outcome scale. That interpretability matters in this paper because the main claims are stated in percentage points, minutes, deaths, and YPLL rather than in abstract score units. Its doubly robust structure is useful here because consistent estimation requires correct specification of either the treatment model or the outcome model, not necessarily both. For county models, continuous-treatment robustness uses Double Machine Learning (DML) and stabilized generalized propensityscore weighting, while overlap weighting (ATO) is used as a positivity-focused sensitivity analysis. County models additionally include state-level leave-one-out spatial lag controls, with spatial dependence assessed via state-level block bootstrap and spillover diagnostics. SAR (ML Lag) and SEM (ML Error) models with queen-contiguity weight matrices are also estimated and reported alongside the primary AIPW results in the SI (Table S18). Confirmatory results are reported at α = 0.05. Exploratory results are summarized with False Discovery Rate (FDR) adjustment. Estimator triangulation (SI Tables S8–S10), placebo and baseline checks (SI Tables S2–S4), threshold and overlap sensitivity (SI Tables S12 and S17), multi-year comparisons (SI Tables S18–S19), and fuller balance diagnostics (SI Tables S20–S21; Fig. S2) are all retained in the Supplementary Information.
Nuisance model specification
For all primary county-level AIPW analyses, the propensity score model is L2-penalized logistic regression fit on the full pre-specified covariate set (Table S14), and the outcome regression uses a Random Forest regressor (500 trees, default hyperparameters). The primary county estimates use 5-fold cross-fitting to ensure that nuisancemodel predictions for each observation are generated from models trained on held-out data, which is required for valid inference with flexible machine-learning nuisance estimators. Non-crossfitted AIPW estimates are also reported for comparability with earlier specifications and to demonstrate that the negative association direction does not depend on cross-fitting.
Balance and overlap diagnostics
For the primary county estimand (MO14 → DV21), the propensity model achieved AUC = 0.871 and Brier score = 0.082 at treated prevalence 13.5%. Before weighting, the largest imbalance was for log population (SMD = 1.34), with substantial imbalance in baseline YPLL, social/economic factors, and neighboring adoption lags (Table 9). After AIPW weighting, the most policy-relevant confounders moved materially toward zero (for example, social/economic factors from − 0.461 to − 0.084, Medicaid expansion from 0.347 to 0.090, and the MO14 spatial lag from 0.726 to 0.099). Residual imbalance on baseline YPLL and the health-behavior index is why those terms also remain in the outcome model. In separate cross-fitted re-estimation for MO14 → DV21, the effective sample size was 409.1 with a 99th-percentile weight of 19.6; these cross-fitted diagnostics are reported here in the main text, while SI Table S21 reports the non-cross-fitted diagnostics. This supports the claim that the negative cross-fitted estimate is not produced by a tiny handful of extreme-weight counties.
For confirmatory endpoints, we report two-sided p-values at α = 0.05 without multiplicity adjustment. For exploratory outcomes, we control the false discovery rate at q = 0.05 using the Benjamini–Hochberg procedure and report q-values alongside p-values. We assess overlap and positivity by inspecting estimated propensity-score distributions, clipping extreme weights at the 1st and 99th percentiles, and reporting effective samplesize diagnostics in the SI. No additional hard trimming beyond this clipping is used in the primary specification.
Data Availability
Hospital-level technology adoption data from the 2024 American Hospital Association Annual Survey are available under a data-use license from the AHA (https://www.ahadata.com); the license does not permit redistribution of individual hospital records. CMS Hospital Compare outcome data, CDC WONDER mortality data, and County Health Rankings data are publicly available from their respective agencies. All county-level aggregated analytic files, analysis code, and geospatial processing scripts are deposited at GitHub (tag v1.0.0). Researchers seeking to replicate the hospital-level analyses can obtain the AHA data independently under the same licensing terms.
Data Availability
Hospital-level data are available from the AHA under license. Aggregated county data and code will be made available through GitHub.
Author Contributions
A.J. designed research, performed research, analyzed data, and wrote the paper; D.G. and T.H. designed research and edited the paper.
Competing Interests
The authors declare no competing interests.
Supplementary Information
Hospital AI and Robotics Adoption, Access, and County Mortality: A National Study Across 3,143 U.S. Counties
S1 Extended Methodology
S1.1 Augmented Inverse Probability Weighting (AIPW) Specification
To address any request for the full model specification truncated in the main text, we provide the complete estimator here. We employed an Augmented Inverse Probability Weighting (AIPW) estimator to calculate the Average Treatment Effect (ATE). AIPW is a “doubly robust” method; it yields consistent estimates if either the treatment model (propensity score) or the outcome model is correctly specified.
The ATE is estimated as:
Where:
Yi is the observed outcome (e.g., Premature Death Rate/YPLL) for county i.
Ti is the binary treatment indicator derived from the continuous county access-weighted exposure Ai (Section S1.2). For the primary extensive-margin estimand, Ti = 𝕀{Ai > 0}; for median-split analyses, Ti = 𝕀 {Ai > median(A)}.
Xi is the vector of covariates. Crucially, this vector includes lagged outcome data (2019 YPLL) to adjust for historical baseline health.
is the estimated propensity score (probability of treatment given covariates).- and are the potential outcomes predicted by regression models.
We also retain brief dose–response sensitivity analyses using the continuous exposure Ai to verify that results are not an artifact of dichotomization.
S1.2 Access-Weighted Exposure Score Construction
County-level exposure is based on population access to adopting hospitals rather than county residence alone. Let b index Census block groups within county i, with population pb. Let h index hospitals, with utilization weight wh given by Adjusted Patient Days (adjpd), and let Zhk ∈ {0, 1} indicate whether hospital h reports adopting technology k (e.g., AI for routine tasks).
We operationalized 30-minute drive time accessibility using a 15.4-mile Euclidean buffer around each hospital, adjusted for the average US road network circuity factor of 1.326. This accounts for the fact that road networks are approximately 30% longer than straight-line distances. We compute distance-based travel time tbh (minutes) from each block-group centroid to each hospital using this circuity-adjusted metric. Define the set of hospitals reachable within 30 minutes as ℋb,30 = {h : tbh ≤ 30}. For each block group, we define a utilization-weighted access score to technology k as:
This construction naturally handles overlapping catchments by weighting the accessible hospital set rather than double-counting population. The county-level exposure is the population-weighted average across block groups:
For the AIPW analyses we estimate a binary ATE using Ti = I {Ai > 0} (extensive margin) or Ti = 𝕀 {Ai > median(A) }(median split) (main text §4.3). We also retain sensitivity analyses using alternative thresholds and continuous dose–response specifications.
Sensitivity note: Hotspot classification (LISA) results for AI intensity are moderately sensitive to weight normalization schemes (e.g., ≈ 10% of counties change class when switching from global to row-standardized weights), though robotics clustering is highly stable. Prior research and our own sensitivity analysis using actual OpenStreetMap routing for 500 randomly sampled block groups across metropolitan and rural areas show strong concordance between circuity-adjusted Euclidean estimates and network-based drive times (factor of 1.3–1.4), validating our proxy measure.
S1.3 Double Machine Learning (DML) Specification
To address concerns regarding functional form misspecification in the nuisance parameters (propensity score and outcome model), we employed Double Machine Learning (DML) with cross-fitting. This approach separates the estimation of nuisance functions from the treatment effect inference to avoid overfitting bias.
We utilized the Partially Linear Model specification:
Where Yi is the outcome (Premature Death Rate), Ti is the binary treatment indicator, and Xi is the high-dimensional vector of covariates.
The average treatment effect θ is estimated using the Neyman-orthogonal score function with K-fold cross-fitting (K = 5):
Where 
and 
are Random Forest regressors/classifiers trained on the complement samples I−k. This ensures that the nuisance functions predicting the outcome and treatment for observation i were not trained on observation i, mitigating “own-observation” bias.
S1.4 Positivity and Overlap Diagnostics
To evaluate the positivity (overlap) assumption, we examined the empirical distribution of estimated propensity scores 
for treated and control units and summarize overlap via min/median/max diagnostics and the proportion outside common-support bounds. As a sensitivity analysis, we re-estimated primary models after trimming observations with extreme propensity scores (e.g.,
or 
); conclusions were unchanged.
S1.5 Targeted Maximum Likelihood Estimation (TMLE) Specification
We further validated our primary results using Targeted Maximum Likelihood Estimation (TMLE), an efficient substitution estimator. TMLE updates the initial estimate of the conditional mean outcome, 
, by fitting a “clever covariate” regression that targets the parameter of interest (ATE) and minimizes bias. This step reduces the mean squared error of the estimator compared to standard IPW or regression adjustment, especially in finite samples with high-dimensional covariates.
S1.6 Counterfactual Decomposition Methodology
To decompose the observed mortality gap between AI desert and oasis counties into potentially addressable components, we employed a regression-based counterfactual simulation framework. This approach estimates how much of the gap could be reduced if desert counties’ AI access improved to oasis levels while holding social determinants of health (SDOH) constant.
S1.6.1 Desert–Oasis Identification
We defined “AI desert” counties as those in the worst decile (P90) of AI access (measured as population-weighted proximity to hospitals with MO14: AI for automating routine tasks), and “AI oasis” counties as those in the best decile (P10). The observed mortality gap was 2,845.4 YPLL per 100,000 (desert mean: 11,299.3; oasis mean: 8,454.0).
S1.6.2 Counterfactual Prediction
Using the fully adjusted regression model (OLS with standardized predictors):
we constructed a hypothetical county profile with desert-level SDOH covariates (mean values from P90 counties) but oasis-level AI access (mean from P10 counties). Predicting YPLL for this counterfactual profile yielded 10,551.6 YPLL—a reduction of 747.7 YPLL (26.3% of the observed gap). The residual 2,097.6 YPLL (73.7%) reflects baseline structural differences not addressable through AI access alone.
S1.6.3 Shapley Value Decomposition of SDOH Components
To attribute the structural component across County Health Rankings determinants (IV1–IV4) in an order-invariant manner, we employed Shapley value decomposition. For each CHR determinant (j ∈ {IV2, IV3, IV4} ), we compute the Shapley value as the average marginal contribution across all 3! = 6 permutations:
where MCj(π) is the marginal reduction in YPLL when j is improved (from desert to oasis level) in its position within permutation π. IV1 (Public Health Policy) is allocated a 5% residual as Medicaid expansion is not directly measured in the model. Negative Shapley values are excluded from proportion calculation using max(0, φj).
This approach yielded: IV4 (Socioeconomic) ≈ 86% of SDOH (64.4% of total gap), IV3 (Health Behaviors) ≈9% of SDOH (6.7% of total), IV2 (Physical Environment) ≈ 0% (nonpositive marginal contribution), and IV1 (Policy residual) = 5% of SDOH (3.7% of total gap). The near-zero IV2 result reflects conditional collinearity: conditional on socioeconomic status, health behaviors, and AI access, physical environment has negligible marginal predictive contribution.
S1.6.4 Bootstrap Confidence Intervals
To quantify uncertainty, we performed county-level bootstrap resampling (1,000 iterations), re-identifying desert/oasis counties in each sample. The bootstrap distribution yielded 95% confidence intervals: tech share = 25.3% (CI: 15.4%–34.6%). Sensitivity analysis varying desert/oasis definitions (P95/P5, P90/P10, P80/P20) yielded tech shares of 16.4%, 26.3%, and 33.4% respectively, with IV2 consistently near-zero across all thresholds.
S1.6.5 Limitations and Interpretation
This decomposition is model-implied (relies on linear additive assumptions and conditional exchangeability), Shapley-based (orderinvariant but assumes additive separability), and AI-specific (varies only AI access, not robotics). It should be interpreted as a counterfactual scenario to contextualize effect magnitudes, not as a definitive causal partition.
S1.7 Multiple Testing and FDR Control
We distinguish confirmatory endpoints from exploratory outcomes. Confirmatory endpoints are evaluated at α = 0.05 using unadjusted two-sided p-values. For exploratory outcomes, we control the false discovery rate at q = 0.05 using the Benjamini– Hochberg procedure applied within each prespecified outcome family (hospital-level exploratory outcomes; county-level exploratory outcomes). We report both p-values and BH-adjusted q-values in all exploratory tables and figures.
S2 AI Taxonomy and Mapping
To clarify the definition of “Workflow AI” versus “Robotics,” Table S1 maps the specific AHA survey items to our analysis categories.
S2.1 Technology Adoption Correlates
To understand the structural factors associated with technology diffusion, we conducted logistic regressions predicting AI and robotics adoption as a function of hospital size, ownership, and urbanicity. Hospital bed size emerged as the dominant predictor, with adjusted odds ratios of 1.93 (95% CI: 1.89–1.97; p < 0.001) for AI adoption and 3.64 (95% CI: 3.53–3.76; p < 0.001) for robotics adoption per 100-bed increase. Controlling for bed size, urban location and for-profit status showed weak to null associations with adoption, suggesting that capacity rather than market structure primarily drives diffusion. These findings are consistent with prior research on health IT adoption patterns 7 and reinforce the interpretation that technology access inequalities largely reflect underlying facility size disparities rather than explicit geographic or profit-driven targeting.
S3 Robustness: Hospital-Level Baseline Checks
To validate our causal inference strategy, we performed baseline checks regressing 2023 AI adoption (from the 2024 AHA release) on 2019 baseline outcomes. A significant association here would suggest reverse causality (e.g., better hospitals adopt AI). As shown in Table S2, we found no significant association for our primary mechanisms, supporting the validity of the 2023 results.
S3.1 County-Level Baseline Checks (2019 YPLL as Pre-Treatment Outcome)
To assess whether county-level baseline health is systematically associated with 2024 technology exposure, we performed baseline checks using 2019 premature death (YPLL) as a pre-treatment outcome. A significant association here would indicate that technology adoption is predicted by prior health status, raising concerns about reverse causality. Table S3 presents results for all 8 exposure measures with Benjamini–Hochberg (BH) q-values to summarize multiple-testing sensitivity within this baseline family.
S3.2 Fisher’s Combined Baseline Checks (County-Level)
To formally assess whether the county-level baseline checks (2019 YPLL as pre-treatment outcome) collectively suggest systematic baseline associations, we conducted Fisher’s combined probability test. Table S4 presents results for both the full placebo family (all 8 exposures) and the primary exposure family (MO14 and MO21 only).
S4 Hospital-Level Stratified Analysis
To test whether the SEP-1 findings are robust across different hospital types and characteristics, we performed stratified analyses. Table S5 shows the effect of AI Staff Scheduling on Sepsis Appropriate Care (SEP-1) across key hospital stratification variables. The effect remains positive and statistically significant in acute care hospitals, high-volume facilities (Q4), and rural hospitals, suggesting the mechanism is not driven by a single hospital archetype.
S5 County-Level Outcome Sensitivity: DV21 vs. CT5
To examine temporal alignment and measurement sensitivity, we compared results using two different mortality measures: (1) DV21: County Health Rankings YPLL per 100,000 (2020–2022, age-adjusted, multi-year blend); and (2) CT5: CDC-derived age-adjusted YPLL per 100,000 for deaths under 75 years in 2023 (contemporaneous with the 2023 exposure measurement). For CT5, suppressed death cells (typical in low-population counties) were imputed to 5 (the midpoint of the 1–9 suppression range) to maximize sample coverage in the baseline AIPW table below. The purpose of this comparison is not to claim that DV21 and CT5 are interchangeable, but to assess whether the county signal remains directionally similar when moving from a historical multi-year YPLL outcome to a contemporaneous 2023 measure. As shown in Table S6, the revised complete-case estimates remain negative for both MO14 and MO21 on both outcomes, but CT5 is materially less precise than DV21 even before overlap-focused reweighting.
We therefore treat DV21 as a lagged benchmark and CT5 as supportive but noisier contemporaneous evidence.
S6 Moderation Analysis: CT6 Hospital-Setting Mortality
To understand the conditions under which workflow AI automation (MO14) is most protective against hospital-setting mortality, we tested whether hospital process and safety metrics moderate the MO14 → CT6 association. Table S7 presents the complete moderation results, including null interactions reported for transparency.
County-level AIPW associations for workflow AI (MO14) and robotics (MO21) on premature mortality outcomes. DV21 is County Health Rankings YPLL per 100,000 (2020–2022). CT5 is CDC-derived YPLL per 100,000 for 2023 using deaths by age band < 75; suppressed death cells are imputed to 5 (midpoint) for CT5.
S7 Methodological Triangulation and Spatial Sensitivity
To further address concerns about estimator choice, dichotomization, and spatial dependence, we report complementary sets of checks spanning triangulation, continuous treatment, explicit spatial econometric models, and threshold sensitivity.
S7.1 Triangulation for the primary hospital mechanism
For the primary hospital confirmatory endpoint (AI Staff Scheduling → SEP-1), AIPW, DML, and TMLE yielded nearly identical estimates, reinforcing that the hospital-level mechanism is not an artifact of a single estimator choice.
S7.2 Triangulation for the primary county estimand
To complement the hospital-mechanism triangulation, we compared estimators for the focal county association MO14 → DV21 (Table S9). Across cross-fitted AIPW, non-cross-fitted AIPW, DML, TMLE, overlap weighting, and spatial block bootstrap, every point estimate is negative. This pattern is the core robustness result for the county analysis: the negative DV21 association does not depend on a single nuisance-fitting strategy or a single causal estimator family.
S7.3 County-level continuous-treatment and spatial robustness
For MO14 → DV21, the primary cross-fitted AIPW estimate is strongly negative (− 975.8 YPLL), the non-cross-fitted baseline is more conservative (− 457.5), and TMLE / spatial-block-bootstrap estimates remain directionally concordant. Continuous-treatment DML and stabilized GPS-IPW are also negative but less precise. These results suggest that the county-level mortality association is not solely an artifact of dichotomization.
A two-part (hurdle) specification further suggests that the extensive margin drives the county result: for DV21 the participation effect is − 510.3 YPLL per 100,000, whereas the intensive-margin slopes among adopters are approximately zero (OLS = +1.02, p = 0.47; GPS = − 1.75, p = 0.32). This pattern is consistent with the zero-inflated empirical distribution of county access and supports use of the binary main-text estimand.
Probabilistic bias analysis for treatment misclassification
Because MO14 and MO21 are self-reported hospital adoption items, we evaluated nondifferential treatment-misclassification sensitivity using probabilistic relabeling over assumed sensitivity/specificity pairs. For MO14 → DV21, every tested scenario preserved a negative median ATE (roughly − 83 to − 591 YPLL per 100,000), although only one scenario yielded a 95% interval excluding zero. Relative to the no-misclassification cross-fitted estimate, these simulated medians were generally attenuated toward smaller magnitudes, indicating that plausible label error can reduce effect size and statistical precision even when direction is preserved. MO21 → DV21 remained directionally negative across scenarios but was less precise, while CT5 and CT6 were notably less stable under assumed misclassification. We therefore view the misclassification analysis as supporting directional robustness of the focal MO14 → DV21 association, while also showing that county-level magnitudes are sensitive to plausible treatment-label noise.
S7.4 Distance-Based Dose-Response
These results complement the primary AIPW analysis (which uses an access-weighted adoption share) by showing that a geographically grounded exposure measure yields a significant continuous gradient. The stronger 2024 slope likely reflects the 56% increase in AI-enabled hospitals (1,119 to 1,743), which expanded the variation in the positive tail of the exposure distribution. Together with the hurdle-model evidence in S7.2, these findings suggest the extensive margin dominates the population-level signal, but a meaningful intensive gradient exists when exposure is measured through geographic proximity rather than institutional adoption intensity. This discrepancy between geographic and institutional-intensity exposure measures helps reconcile the significant distance-based gradient with the weaker within-adopter intensive-margin estimates in the primary access-share models.
S7.5 Threshold and access-construction sensitivity
To probe whether the county result depends on a single exposure cutoff, we re-estimated AIPW after binarizing MO14 at 10, 25, and 50 percentage-point thresholds; all estimates remained negative. We also examined alternative drive-time thresholds (15, 30, 45, and 60 minutes) and alternative hospital weighting schemes (unweighted, beds, adjusted patient days). Negative associations persisted across these access constructions.
S7.6 Overlap Weighting (ATO) Sensitivity
To address limited overlap and extreme inverse-probability weights, we re-estimated county-level effects using overlap weights (ATO), which downweight observations with propensity scores near 0 or 1. Table S17 reports the non-cross-fitted AIPW baseline and ATO estimates side by side. For the primary historical outcome (DV21), both MO14 and MO21 remain negative and statistically significant under ATO (MO14: ATE = − 201.6, 95% CI [ − 328.3, − 70.2], p = 0.004; MO21: ATE = − 185.5, 95% CI [− 301.6, − 72.0], p = 0.004), with ESS 1,743.6 and 1,893.2, respectively. As expected, ATO magnitudes are attenuated relative to full-population estimands because ATO targets the overlap population. For contemporaneous CT5 and CT6 outcomes, ATO estimates attenuate toward the null, indicating lower effective signal once extreme-propensity counties are downweighted.
S8 Variable Definitions
Table S14 details the construction and sources of the covariates used in the propensity score and outcome models.
Falsification / Negative Controls
Panel B presents outcomes with no plausible clinical pathway through workflow AI or robotics. The absence of significant associations supports the specificity of the primary mechanism findings and argues against residual confounding by general hospital quality.
S8.1 Sensitivity to COVID-19 Confounding
To assess whether differential pandemic burden explains the primary county signal, we re-estimated MO14 → DV21 on the 2,686-county subsample with complete COVID-19 mortality data, with and without cumulative COVID-19 deaths (2020–2022) as an additional covariate. On this restricted sample, the baseline estimate was ATE = − 411.0 (95% CI [ − 788.2, − 98.5]; p = 0.008). Adding the COVID covariate yielded ATE = − 452.1 (95% CI [ − 817.3, − 161.1]; p = 0.004), a less-than-10% change in magnitude. The stability and slight strengthening after adjustment suggest the main association is not driven by pandemic-burden differences.
S9 Explicit Spatial Econometric Models
The SAR spatial lag parameter (ρ = 0.341) is highly significant, confirming that county mortality outcomes are spatially interdependent. After absorbing this dependence, the AI coefficient attenuates from − 702 (OLS) to − 422 (SAR) but remains significant, suggesting that a substantial share of the naive OLS association reflects spatial clustering. The SEM specification yields an intermediate coefficient (− 533) with a large spatial error parameter (λ = 0.487), and conclusions remain directionally consistent across specifications.
IPTW versus overlap-weighted (ATO) estimates for primary county mortality associations. ATO targets the overlap population and downweights extreme propensity scores.
Spatial Durbin perspective
While we do not estimate a full Spatial Durbin Model in this version, the OLS + neighbor-exposure specification provides an analogous decomposition. The direct AI coefficient is − 662 YPLL (p < 0.001) and the neighbor-exposure coefficient is − 391 YPLL (p = 0.09), suggesting that a non-trivial share of the total association may reflect spillovers through shared referral networks or regional health infrastructure. If spillovers contaminate the control group, the direct AIPW estimate is biased toward zero, implying that the primary estimates may be conservative. A formal interference-aware estimand (e.g., EATE) that explicitly separates direct and spillover effects would require stronger assumptions about interference structure (for example, partial interference within referral networks) and is left for future work. As complementary spatially robust inference, the state-level block bootstrap (49 clusters, 300 resamples) yields ATE = −498.4 with 95% CI [−902.4, −173.4].
S10 Multi-Year Panel Comparison, 2022–2024
Table S19 summarizes cross-wave changes in technology diffusion, geospatial access, inequality, and county mortality model performance. The consistent negative AI coefficients across years and model types provide cross-wave replication. The divergence between coverage gains (+9.0 pp) and Gini deterioration (+0.028) indicates that new AI-enabled hospitals are disproportionately located in areas that already had relatively good access.
S11 Covariate Balance Diagnostics and Overlap
We assessed the quality of the propensity score weighting by examining Standardized Mean Differences (SMD) before and after weighting. An SMD with an absolute value below 0.1 is often used as a rough heuristic, but in this setting the more important question is whether weighting compresses the largest confounders while the doubly robust outcome model still carries the lagged outcome and other residual imbalances directly.
Fig. S2 illustrates the balance for the primary treatment variable: AI for Automating Routine Tasks. Prior to weighting (red circles), significant imbalances existed, particularly for population size, baseline YPLL, and neighboring adoption lags. After weighting (blue squares), the largest imbalances compress sharply toward zero, although the lagged outcome and health-behavior index remain modestly imbalanced—consistent with the main-text argument for retaining those terms directly in the outcome regression.
Table S21 provides diagnostics on the propensity-score weights and effective sample sizes (ESS) for the primary county estimand. Overlap was sufficient after clipping extreme weights at the 1st and 99th percentiles, and the overlap-weighted ATO estimates in Table S17 further reduce sensitivity to limited common support.
S11.1 SUTVA Diagnostics (Spillover and Interference)
To test the Stable Unit Treatment Value Assumption (SUTVA), we examined two potential violations: spatial spillovers (interference) and treatment variation.
Spillover Effects: We compared outcomes for treated counties surrounded by high-adoption neighbors versus those surrounded by low-adoption neighbors. High-High (HH) clusters showed better raw outcomes (approx. 1,188 YPLL lower) due to baseline socioeconomic advantages, but High-Low (HL) “Islands” showed the strongest adjusted marginal association (β ≈ − 395), implying that spillovers likely contaminate the control group in clustered areas (conservative bias).
Treatment Heterogeneity: We tested if the treatment effect varied by hospital size (Small vs. Medium vs. Large AI-enabled hospitals). ANOVA results indicated no significant heterogeneity (p ≈ 0.60), supporting our use of a binary treatment definition rather than an intensity-based metric, which was found to be collinear with urbanicity.
S12 Missingness and Reproducibility
Missing data was handled using listwise deletion for the primary analysis (N = 3, 143 counties with complete data). A sensitivity analysis using missingness indicators for key covariates (e.g., granular housing data) did not alter the direction or significance of the primary results. For hospital-level analysis, hospitals without valid 2024 CMS outcome data were excluded, but weights were recalculated to preserve representativeness. All code scripts and aggregated data files are provided in the GitHub repository to facilitate reproduction.
The plot displays the Standardized Mean Difference (SMD) between MO14-exposed and unexposed counties for key covariates. The red circles indicate the initial imbalance (unweighted), showing profound differences in population size, baseline mortality, and neighboring adoption. The blue squares indicate the balance after weighting. Most covariates move sharply toward zero, though baseline YPLL and health behaviors remain modestly imbalanced—one reason the doubly robust outcome model retains those terms directly. Overlap diagnostics: companion diagnostics in Tables S20–S21 summarize balance, effective sample sizes, and weight tails.
References
