Clinical AI Adoption Patterns Among U.S. Hospitals: Comparing Hospital Adoption and Implementation Using AHA Survey Data

Clinical AI Adoption Patterns Among U.S. Hospitals: Comparing Hospital Adoption and Implementation Using AHA Survey Data | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Clinical AI Adoption Patterns Among U.S. Hospitals: Comparing Hospital Adoption and Implementation Using AHA Survey Data Emma B. Sartin, Elayne L. Wilson, Hugh B. Roland, Eric Wallace, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9140964/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background : Hospitals are investing in clinical artificial intelligence (AI), yet most efforts to classify the types and extent AI uptake rely on ad hoc counts of tools or binary adoption indicators. A scalable, psychometrically grounded measure of “clinical AI implementation maturity” would strengthen benchmarking and improve research on determinants and consequences of adoption. Objective : To create a hospital-level measure of clinical AI use and adoption using staged implementation items from the American Hospital Association’s (AHA’s) Annual Survey. Methods : Ordered response items describing implementation stage across multiple clinical AI functions were modeled using an item response theory (IRT) framework. We evaluated the suitability of a unidimensional maturity construct by examining model fit indicators and the coherence of item behavior. Item discrimination parameters and ordered category thresholds were used to determine how well each AI function differentiated hospitals along the maturity continuum and whether response categories reflected interpretable progression. Results : The fitted model supported a stable maturity gradient with meaningful between-hospital variation. Item discrimination estimates indicated that certain AI functions were substantially more informative for distinguishing hospitals at different maturity levels than others, while some functions contributed limited differentiation. Category thresholds were generally ordered, supporting the interpretation of staged implementation as a progression. Threshold patterns suggested that adoption is not random across functions; some capabilities tend to appear earlier in maturity trajectories, whereas others cluster at later stages and better discriminate among higher-maturity hospitals. The resulting scores can be used as continuous measures with uncertainty intervals for comparisons and stratified analyses. Conclusions : Clinical AI implementation maturity can be measured as a latent construct using routine survey data and IRT methods. A psychometrically validated maturity score provides a foundation for benchmarking, risk adjustment in comparative studies, and policy-relevant monitoring of clinical AI diffusion across healthcare delivery systems. Clinical artificial intelligence Hospital adoption Implementation maturity Item response theory Health informatics Predictive analytics Introduction Clinical artificial intelligence (AI) refers to a broad set of computational tools, ranging from traditional statistical models to modern machine-learning systems, that support decision-making and workflow in health care.[ 1 , 2 ] In practice, clinical AI can analyze large volumes of clinical data (e.g., structured electronic health record fields, imaging, waveforms, and clinical text) to generate predictions, classifications, summaries, or recommendations that clinicians and health systems use to improve care.[ 3 ] When deployed effectively, clinical AI offers several potential benefits: earlier identification of deterioration or high-risk patients, more consistent and timely interpretation of complex data, reduced administrative burden through automation, and improved operational efficiency through better triage and resource allocation; in turn, these gains may translate into safer care, improved patient outcomes, and a more sustainable clinical workforce.[ 1 , 2 , 4 , 5 ] Despite this promise, clinical AI adoption and implementation remain uneven across U.S. hospitals and use cases.[ 4 – 8 ] “Adoption” (selecting, purchasing, or initiating use of an AI capability) does not guarantee “implementation” (integrating the tool into real-world workflows at scale, with durable use, performance monitoring, and governance).[ 6 , 7 ] Many health systems successfully pilot AI tools but struggle to operationalize them, with barriers spanning financial, technical, organizational, and regulatory domains.[ 9 ] Common challenges include uncertainty about return on investment, limited interoperability and integration with existing information systems, data quality and representativeness concerns, clinician trust and usability issues, workflow disruption, insufficient staffing for model maintenance and monitoring, and evolving legal and regulatory expectations.[ 9 , 10 ] These obstacles can prevent promising tools from moving beyond isolated pilots, or can lead to partial implementation that fails to deliver measurable value. With this in mind, we examine the clinical AI landscape through the lens of this “adopted versus implemented” gap. Specifically, using the Fiscal Year 2024 American Hospital Association (AHA) Annual Survey [ 11 ] and Item Response Theory (IRT)[ 12 ], we assess which clinical AI constructs hospitals report adopting and which are reported as fully implemented in routine operations.[ 13 ] Additionally, we identify constructs that appear more likely to stall between adoption and implementation. By distinguishing early uptake from sustained operationalization, our goal is to clarify where hospitals are successfully translating clinical AI into practice and where persistent barriers may be limiting real-world impact. METHODS Study design and data source We conducted a cross-sectional, secondary analysis using the Fiscal Year (FY) 2024 AHA Annual Survey Database. The AHA Annual Survey is administered nationally and is refined on an annual basis to align with emerging issues in hospital operations and policy-relevant measurement needs. The FY2024 release included newly added items related to AI implementation. The focus of the current analyses was exploring the adoption and implementation of clinical AI packages. Measures Clinical AI implementation items. Our primary measures were seven clinical AI implementation items, answered on a 5-level ordered response scale: 1 = not implementing, 2 = exploring, 3 = piloting/testing, 4 = expanding, 5 = fully integrated. “Don’t know” responses were coded as missing for analysis. Items specifically probed hospitals’ adoption and implementation of: (1) AI-assisted diagnostics (DIAGAI), (2) predictive analytics for patient care (PAPCAI), (3) clinical decision support tools (CLINAI), (4) AI-assisted surgery (SURGAI), (5) patient communication and education (PCOEAI), (6) population health management (POPHAI), and (7) resource allocation during emergencies (PDMRAI). Additionally, we examined open-text entries for additional AI use cases (see below for approach). Hospital characteristics. Hospital characteristics were used to describe the analytic sample and to compare hospitals included in the clinical AI IRT estimation sample with those not included. Ownership/control codes were recoded into four categories: government non-federal, nonprofit, for-profit, and federal. Bed size was operationalized as small, medium, or large using a three-category bed size variable derived from AHA bed size code. Region was operationalized as Northeast, South, Midwest, and West using a Census region variable derived from state code. Additional characteristics included teaching proxy, urban or rural proxy, health system membership, and Joint Commission accreditation. Analytic approach We followed the logic of Chan et al.’s application of item response theory (IRT) to AHA survey items, with modifications appropriate for ordered response categories.[ 14 ] Unidimensionality assessment. A prerequisite for unidimensional IRT modeling is evidence that the item battery reflects a single latent construct. Exploratory factor analysis using principal factors and an unrotated solution was used to evaluate whether the clinical AI items formed a sufficiently unidimensional scale. Our criteria for “sufficient” unidimensionality were based on prior IRT guidance used by Chan et al.,[ 14 ] including a dominant first factor and strong first-factor loadings across items. IRT model. Because clinical AI implementation was measured using ordered stages rather than binary responses, we estimated a graded response IRT model to obtain an item discrimination parameter for each clinical AI use case and category threshold parameters corresponding to transitions across implementation stages. Discrimination parameters characterize how well each item differentiates hospitals along the latent clinical AI implementation maturity continuum. Threshold parameters characterize the latent maturity level at which the probability of endorsing an implementation stage or higher reaches 50 percent. We generated a hospital-level latent clinical AI implementation maturity score (theta) from the fitted model. Sample inclusion and missingness. “Don’t know” responses were treated as missing. IRT estimation used hospitals contributing non-missing information to the clinical AI battery under the model’s likelihood-based estimation, yielding the clinical AI IRT estimation sample. Hospitals outside the estimation sample were retained for descriptive comparisons to characterize selection into the analytic sample. Descriptive statistics and correlations. Item means and a correlation matrix for the clinical AI items were computed within the IRT estimation sample. We calculated Pearson correlations using pairwise available observations, consistent with common reporting conventions for survey-based scale development. Open-text processing. Open-text “other clinical AI” responses were extracted and reviewed to identify recurring themes. We normalized and categorized text using a keyword-based coding approach to support transparent frequency reporting. We defined a documentation-support category a priori to capture “digital scribe” and closely related terms (for example, “AI scribe,” “ambient scribe,” “ambient listening,” “dictation,” and “transcription”), given the prominence of documentation-oriented entries in the open-text field. RESULTS The FY2024 AHA file contained 6,173 hospitals; of these, 2,423-2,640 hospitals answered the newly added clinical AI items (39.3% – 42.8% of the full sample). Predictive analytics for patient care had the highest response rate ( n = 2,640; 42.8%). Resource allocation during emergencies and AI-assisted surgery had the lowest response rate ( n = 2,423; 39.3%; n = 2,470; 40.0%, respectively). We did not include 3,445 hospitals (55.8%) from the graded response IRT model because they did not contribute usable clinical AI item information under the model. Thus, the model was estimated on 2,728 hospitals (44.2% of the full sample). Table 1 compares hospitals included in the clinical AI IRT estimation sample with those not included. Hospitals included in the estimation sample were larger and more complex. Mean inpatient days (IPDTOT) were higher in the estimation sample (47,407 vs. 28,702), as were mean inpatient discharges (DCTOT; 8,182 vs 3,991) and commercial discharges (THRTDC; 2,075 vs 887). Bed size categories also differed, with “large” hospitals comprising 20.0% of the estimation sample versus 9.5% of those not included. Additionally, more teaching programs were represented in the estimation sample (7.6% vs 2.1%). Ownership differed substantially: nonprofit hospitals comprised 69.1% of the estimation sample compared with 37.4% of those not included. Conversely, for-profit hospitals comprised 13.5% of the estimation sample and 36.3% of the hospitals not included. Regional distribution also differed, with a higher share of Midwestern hospitals in the estimation sample (32.3% vs 22.9%). CBSA type (metro/micro/rural) did not differ significantly between groups ( p = 0.63), despite significant differences for bed size, teaching status, ownership, and region (Table 1 ). Table 1 Hospital AI Sample Characteristics Hospital Characteristics AI Sample Non-respondents Bed Size Small ( 300)* 546 328 Teaching status Yes* 2,522 3,373 No 206 72 Ownership Nonprofit* 1,884 1,288 For-profit 369 1,252 Public 475 905 Accreditation Yes* 2,522 3,373 No 206 72 Region 2 Northeast 369 400 South 984 1,486 Midwest* 881 788 West 488 704 CBSA Type Metro (> 50K) 1,849 2,374 Micro (10-50K) 385 467 Rural (< 10K)* 494 604 Notes: 1. * p < .001 indicates that AI more likely in use or being implemented 2. Northeast (CT|ME|MA|NH|RI|VT|NJ|NY|PA); South (DE|DC|FL|GA|MD|NC|SC|VA|WV|AL|KY|MS|TN|AR|LA|OK|TX); Midwest (IL|IN|MI|OH|WI|IA|KS|MN|MO|NE|ND|SD); West (AZ|CO|ID|MT|NV|NM|UT|WY|AK|CA|HI|OR|WA) Table 2 reports item means and inter-item correlations for the clinical AI battery within the IRT estimation sample. Item means indicated the greatest implementation maturity for clinical decision support tools (CLINAI; mean 2.98) and predictive analytics for patient care (PAPCAI; mean 2.87), followed by AI-assisted diagnostics (DIAGAI; mean 2.78). The lowest mean stage was observed for resource allocation during emergencies (PDMRAI; mean 1.62), consistent with low uptake for that use case. Pairwise correlations were uniformly positive and moderate-to-strong (range 0.37–0.81), with the strongest correlation observed between predictive analytics and clinical decision support ( r = 0.81). The correlation structure supported a single underlying construct of clinical AI implementation maturity, consistent with the factor analytic evidence reported in the Appendix. Table 2 Item Means and Correlations AI Tools Mean 1 2 3 4 5 6 1. Resource allocation during emergencies 1.622 2. AI assisted surgery 1.774 0.478 3. Patient communication and education 2.295 0.365 0.448 4. AI assisted diagnostics 2.776 0.371 0.451 0.576 5. Population health management 2.186 0.537 0.551 0.639 0.558 6. Clinical decision support tools 2.977 0.390 0.445 0.618 0.633 0.637 7. Predictive analytics for patient care 2.865 0.434 0.460 0.631 0.675 0.666 0.809 Table 3 presents graded response IRT results. All items demonstrated monotonic threshold progression (Diff ≥ 2 < Diff ≥ 3 < Diff ≥ 4 < Diff = 5), supporting ordered-category functioning and model fit. Resource allocation during emergencies (PDMRAI) exhibited the highest “full integration” threshold (Diff = 5: 2.72), indicating the greatest difficulty reaching full integration on the latent maturity continuum. AI-assisted surgery (SURGAI) showed the second highest full-integration threshold (2.40). In contrast, clinical decision support tools (CLINAI; 0.89) and predictive analytics (PAPCAI; 0.97) had the lowest full-integration thresholds, indicating comparatively easier progression to advanced implementation. Several adjacent items exhibited non-overlapping confidence intervals for key threshold estimates, indicating distinct tiers of implementation difficulty rather than a smooth continuum. Discrimination estimates were highest for predictive analytics (a = 4.56) and clinical decision support (a = 3.93), indicating strong differentiation of hospitals along the latent maturity trait, and were lowest for emergency resource allocation (a = 1.59) and AI-assisted surgery (a = 1.89). Table 3 Item Discrimination and Difficulty Scores item (full label) Discrimination (a), 95% CI Diff ≥ 2 Diff ≥ 3 Diff ≥ 4 Diff = 5 Resource allocation during emergencies – (PDMRAI) 1.589 (1.439 to 1.738) 0.256 1.680 2.072 2.721 AI assisted surgery – (SURGAI) 1.886 (1.725 to 2.048) 0.224 1.113 1.530 2.404 Patient communication and education – (PCOEAI) 2.653 (2.468 to 2.837) -0.609 0.453 1.073 1.957 AI assisted diagnostics – (DIAGAI) 2.763 (2.575 to 2.951) -0.761 0.095 0.496 1.137 Population health management – (POPHAI) 3.346 (3.094 to 3.598) -0.366 0.583 1.131 1.466 Clinical decision support tools – (CLINAI) 3.926 (3.633 to 4.219) -0.847 0.009 0.281 0.889 Predictive analytics for patient care – (PAPCAI) 4.564 (4.192 to 4.936) -0.751 0.063 0.408 0.969 Among hospitals providing an open-text entry ( n = 256), documentation-oriented AI dominated the responses. Digital scribe and closely related ambient documentation concepts (e.g., “ambient listening,” “ambient documentation,” “AI scribe,” dictation/transcription, and vendor tools such as DAX/Dragon/Nuance and DeepScribe) appeared in 167 entries (65.2%), with explicit “ambient” terminology present in 95 entries (37.1%). Generative AI and large-language-model-style assistants (e.g., “copilot,” “GPT,” “RAG,” handoff tools, and chart summarization) appeared in 34 entries (13.3%). Smaller clusters referenced revenue cycle/coding (15 entries; 5.9%), robotics (11 entries; 4.3%), and patient messaging/portal automation (11 entries; 4.3%). A residual set of 26 entries (10.2%) was too vague for reliable thematic classification (e.g., “AI” without further specification). DISCUSSION The FY2024 AHA clinical AI items support a coherent, staged construct of “clinical AI implementation maturity” and also demonstrate that clinical AI adoption is not a single phenomenon. Factor and correlation results indicate a single underlying dimension, and the graded response IRT model produced well-behaved item thresholds across all seven items, with monotonic progression for each stage boundary. These results provide two actionable insights: (1) some clinical AI capabilities are systematically easier to move from exploration to piloting and from expansion to full integration, and (2) a subset of capabilities remain meaningfully more difficult to implement, even among hospitals with evidence of broader AI implementation maturity. Patterns in item difficulty align with the operational and governance burden implied by each use case. Resource allocation during emergencies exhibited the highest thresholds for advanced implementation, consistent with both low response and low reported maturity. These results validate the ethical complexity and general reluctance for states and healthcare systems to establish clear guidance on scarce resource allocation.[ 15 ] While research suggests that predictive modeling can optimize patient flow and facilitate triage decisions, the lack of foundational decision-making guidance through Crisis Standards of Care may limit the practical utility of these tools by healthcare systems and providers. Thresholds should also be determined for when implementation advances from operational purposes of patient flow forecasting, efficiency, and supply chain optimization to clinical purposes of emergency resource allocation. Defining these AI implementation categories in the AHA survey may result in higher response for this survey item. AI-assisted surgery also required higher maturity to reach piloting, expansion, and full integration. These use cases typically impose heavier requirements for safety management, cross-disciplinary workflow redesign, and organizational accountability than tools such as predictive analytics and decision support, which can be implemented earlier as “augmented decision” functions embedded within existing clinical pathways. Evidence from implementation-focused syntheses repeatedly emphasizes that clinical AI impact is mediated by sociotechnical fit, governance, and the ability to sustain performance monitoring over time, not by model availability alone.[ 16 ] The threshold ordering observed here offers quantitative reinforcement of that point and suggests that “difficult” AI use cases represent a higher bar of organizational capability rather than simply a slower vendor market. A value proposition that combines clinical effectiveness with operational efficiency likely accelerates adoption. Predictive analytics for patient care and clinical decision support tools were among the least difficult items to integrate at advanced stages and exhibited the strongest discrimination. Tools in these categories can plausibly be justified on safety and effectiveness grounds while also being framed as productivity or throughput interventions; for example, earlier deterioration detection to reduce ICU transfers, standardization to reduce variation, and risk stratification to focus clinician time. A national health-system survey conducted in the early generative AI period similarly found widespread activity in clinical risk stratification and imaging-related AI, but variable success and persistent barriers such as tool immaturity, financial concerns, and regulatory uncertainty.[ 17 ] The same survey identified “ambient notes” as the one use case with adoption activity reported by all respondents, reinforcing the idea that immediate workflow relief and productivity benefits can drive rapid uptake even when longer-term evaluation remains incomplete. Notably, emergency resource allocation differs from the higher-uptake use cases in one critical respect: value realization often depends on coordination beyond the walls of a single hospital.[ 18 ] Preparedness and surge response become meaningfully more effective when response capacity is defined across an interconnected network rather than within one facility. Regional coordination requires reliable data sharing, patient tracking, and operational visibility across institutions and emergency management partners. Peer-reviewed work in disaster medicine has described the heterogeneous, multi-system landscape of regional information-sharing and coordination platforms during large-scale medical surge events and has highlighted persistent barriers to effective information sharing, including fragmentation and interoperability gaps.[ 19 ] Health information exchange itself remains a multi-stage organizational assimilation process with distinct initiation, adoption, implementation, and institutionalization phases.[ 20 ] Systematic reviews and empirical studies indicate that HIE can reduce duplicative testing and emergency department costs and can improve certain operational outcomes, but routine use remains uneven and often context-dependent.[ 21 ] The EMS and emergency care continuum provides a concrete illustration: prehospital-to-hospital data linkage is frequently limited by interoperability, connectivity, and patient matching constraints, despite clear potential value for system-level emergency coordination.[ 22 ] These realities create a plausible structural explanation for why emergency resource allocation AI appears “hard” in the IRT results. Even high-capability hospitals may be constrained by incomplete regional data integration, limited cross-institutional governance, and the episodic nature of the operational value proposition outside active surges. Hospital characteristics associated with greater clinical AI uptake reinforce concerns about capability stratification. Larger hospitals, those with teaching programs, and nonprofit institutions were more likely to appear in the clinical AI estimation sample, suggesting that adoption and maturity are concentrated among organizations with greater capital, informatics capacity, and change infrastructure. Prior work on the diffusion of advanced EHR functions has documented an “advanced use” divide, in which certain hospital types lag in adopting higher-order capabilities even when basic EHR adoption becomes widespread.[ 23 ] Studies focused on critical access hospitals similarly report persistent difficulty keeping pace with more advanced health IT functions, even after broad EHR adoption.[ 24 ] AI maturity results in the present analysis are consistent with this trend, which may produce a practical bifurcation between AI-rich and AI-poor delivery settings. Such divergence has direct implications for health equity, as differential access to advanced clinical decision support, risk stratification, and workflow augmentation can translate into differences in quality, safety, and access to capacity over time. Conversely, a lack of association for CBSA type is notable in this context. Urban versus rural market classification did not differentiate clinical AI uptake in the descriptive comparison, suggesting that organizational resources and governance capacity may dominate simple market-type proxies in predicting measured adoption. A sharper market structure signal may require more granular measures, such as system-level consolidation, vendor footprint, or resource pooling across networks, particularly when AI tools are increasingly procured and governed at the health system level rather than at the facility level. Finally, open-text “other clinical AI” responses provide an additional signal about where hospitals are investing: ambient documentation and digital scribes. This is consistent with external evidence that ambient notes has become a highly salient and widely pursued use case in US health systems. Early empirical evaluations of ambient AI scribe interventions have reported reductions in documentation burden and burnout measures, although evidence remains emergent and implementation introduces new governance needs related to privacy, consent, and clinical documentation quality.[ 25 ] Open-text dominance of documentation support suggests that staged survey batteries that omit documentation AI risk undercounting a major pathway for AI adoption, particularly in the early generative AI era. Survey evolution toward structured staged items for ambient documentation would improve comparability across organizations and over time would facilitate more precise measurement of adoption maturity in a category that appears to be diffusing rapidly. This study has several important limitations. First, it is a cross-sectional, secondary analysis of hospital-level survey data and therefore cannot assess temporal progression or causal drivers of movement across implementation stages. Second, the clinical AI measures are self-reported, and treatment of “don’t know” responses as missing may introduce reporting error and selection bias. Third, the FY2024 AHA clinical AI items capture only a limited set of structured use cases, and the open-text findings suggest that prominent applications such as ambient documentation may be incompletely represented. Conclusions There are several practical implications for this study. The ordered-response IRT approach preserves the staged nature of implementation reporting and translates it into item-level thresholds that identify where hospitals are most likely to stall, such as remaining in exploration rather than progressing to piloting or expansion. It also produces a standardized latent maturity score that can support benchmarking and future work linking maturity to organizational characteristics, interoperability capacity, and downstream outcomes. Chan and colleagues previously demonstrated that IRT can reveal differential implementation difficulty in health IT using AHA survey items, and the present study extends that logic to clinical AI using an ordered-response model aligned with the FY2024 staging format.[ 26 ] Monotonic thresholds across all items further support the integrity of the ordered categories and reduce concern that response options were functioning erratically. Overall, these findings suggest that hospital clinical AI adoption is best understood not as a simple yes-or-no phenomenon, but as a staged process of implementation maturity in which some use cases advance more readily than others. Clinically, this matters as uneven implementation may shape where patients and clinicians actually experience the benefits of AI in practice, particularly in decision support, predictive analytics, and workflow augmentation, while more operationally, ethically, and organizationally complex applications remain less mature. At the same time, interpretation should remain grounded in the practical meaning of the IRT parameters: higher thresholds indicate that a use case requires substantially greater organizational maturity to reach advanced stages, whereas higher discrimination indicates a use case that more strongly differentiates hospitals along the maturity continuum. Together, these results provide a framework that can help health systems, policymakers, and implementation leaders identify where adoption stalls and target the governance, infrastructure, and workforce supports needed to translate AI implementation into reliable and equitable clinical impact. Statements and declarations Author contributions : All authors contributed to the study conception and design. Data analysis, and the drafting of the Methods and Results sections, were performed by Eric W. Ford. Drafts of the Background, Introduction, and Conclusion sections were prepared by Emma B. Sartin. Elayne L. Wilson, Hugh B. Roland, and Eric Wallace drafted the Discussion sections relevant to their respective areas of expertise. All authors contributed to revision of the manuscript and edited the final version for clarity and coherence. All authors read and approved the final manuscript. Funding Declaration: There was no external or internal funding for this research. Software and reporting. All analyses were conducted in StataNow 19.5 (StataCorp, College Station, Texas). Item parameters (discrimination and thresholds), item means, and the correlation matrix were reported to support interpretation and reproducibility. Human subjects considerations. This study is not considered human subjects research as it is a secondary analysis using hospital-level survey data without patient identifiers. Clinical trial number: Not applicable. Ethics and Consent to Participate declaration : Not applicable. References Rajkomar, A., J. Dean, and I. Kohane, Machine learning in medicine. New England Journal of Medicine, 2019. 380 (14): p. 1347-1358. Organization, W.H., Ethics and governance of artificial intelligence for health: large multi-modal models. WHO guidance . 2024: World Health Organization. Spear, J., J.M. Ehrenfeld, and B.J. Miller, Applications of Artificial Intelligence in Health Care Delivery. Journal of Medical Systems, 2023. 47 (1): p. 121. Shortliffe, E.H. and M.J. Sepúlveda, Clinical decision support in the era of artificial intelligence. Jama, 2018. 320 (21): p. 2199-2200. Sendak, M.P., et al., A path for translation of machine learning products into healthcare delivery. EMJ Innov, 2020. 10 : p. 19-00172. Kelly, C.J., et al., Key challenges for delivering clinical impact with artificial intelligence. BMC medicine, 2019. 17 (1): p. 195. Kueper, J.K., et al., Artificial intelligence and primary care research: a scoping review. The annals of family medicine, 2020. 18 (3): p. 250-258. Kanakubo, T. and H. Kharrazi, Comparing the Trends of Electronic Health Record Adoption Among Hospitals of the United States and Japan. Journal of Medical Systems, 2019. 43 (7): p. 224. Sculley, D., et al., Hidden technical debt in machine learning systems. Advances in neural information processing systems, 2015. 28 . Carvalho, J.V., Á. Rocha, and A. Abreu, Maturity Models of Healthcare Information Systems and Technologies: a Literature Review. Journal of Medical Systems, 2016. 40 (6): p. 131. Association, A.H., AHA Annual Survey of Hospitals , A.H. Association, Editor. 2025. Embretson, S.E. and S.P. Reise, Item Response Theory . 2013: Psychology Press. Association, A.H., AHA Annual Survey Information Technology Supplement 2024 (instrument) , A.H. Association, Editor. 2025. Chan, K.S., et al., Assessing electronic health record implementation challenges using item response theory. Am J Manag Care, 2016. 22 (12): p. e409-e415. Services, U.D.o.H.a.H., Technical Resources, Assistance Center, and Information Exchange: Mass Casualty Hospital Capacity Expansion Toolkit , A.f.S.P.a. Response, Editor. 2025. Sharma, M., et al., Artificial Intelligence Applications in Health Care Practice: Scoping Review. Journal of Medical Internet Research, 2022. 24 (10): p. e40238. Poon, E.G., et al., Adoption of artificial intelligence in healthcare: survey of health system priorities, successes, and challenges. Journal of the American Medical Informatics Association, 2025. 32 (7): p. 1093-1100. Robles Mendo, I., et al., Machine Learning in Medical Emergencies: a Systematic Review and Analysis. Journal of Medical Systems, 2021. 45 (10): p. 88. Lee, C.J., et al., Use of Information Technology Systems for Regional Health Care Information-Sharing and Coordination During Large-Scale Medical Surge Events. Disaster Medicine and Public Health Preparedness, 2023. 18 : p. e1. Esmaeilzadeh, P. and M. Sambasivan, Health Information Exchange (HIE): A literature review, assimilation pattern and a proposed classification for a new policy approach. Journal of Biomedical Informatics, 2016. 64 : p. 74-86. Hersh, W.R., et al., Outcomes From Health Information Exchange: Systematic Review and Future Research Needs. JMIR Medical Informatics, 2015. 3 (4): p. e39. Martin, T.J., et al., Health Information Exchange in Emergency Medical Services. Applied Clinical Informatics, 2018. 9 (4): p. 884-891. Adler-Milstein, J., et al., Electronic health record adoption in US hospitals: the emergence of a digital "advanced use" divide. Journal of the American Medical Informatics Association, 2017. 24 (6): p. 1142-1148. Apathy, N.C., A.J. Holmgren, and J. Adler-Milstein, A decade post-HITECH: Critical access hospitals have electronic health records but struggle to keep up with other advanced functions. Journal of the American Medical Informatics Association, 2021. 28 (9): p. 1947-1954. Olson, K.D., et al., Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout. JAMA Network Open, 2025. 8 (10): p. e2534976. Chan, K.S., et al., Assessing electronic health record implementation challenges using item response theory. American Journal of Managed Care, 2016. 22 (12): p. e409-e415. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9140964","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":621674536,"identity":"4f5f18c5-e5b2-4ba0-97fe-e0ae6b29a2e1","order_by":0,"name":"Emma B. Sartin","email":"","orcid":"","institution":"University of Alabama at Birmingham","correspondingAuthor":false,"prefix":"","firstName":"Emma","middleName":"B.","lastName":"Sartin","suffix":""},{"id":621674537,"identity":"47b31dd9-d337-49f8-8152-92bc977f0eab","order_by":1,"name":"Elayne L. Wilson","email":"","orcid":"","institution":"University of Alabama at Birmingham","correspondingAuthor":false,"prefix":"","firstName":"Elayne","middleName":"L.","lastName":"Wilson","suffix":""},{"id":621674538,"identity":"c64ad7bd-68f1-4598-be98-bbf6124987a7","order_by":2,"name":"Hugh B. Roland","email":"","orcid":"","institution":"University of Alabama at Birmingham","correspondingAuthor":false,"prefix":"","firstName":"Hugh","middleName":"B.","lastName":"Roland","suffix":""},{"id":621674539,"identity":"ce4d2acd-336c-4a52-80db-8825a20c40b2","order_by":3,"name":"Eric Wallace","email":"","orcid":"","institution":"University of Alabama at Birmingham Hospital","correspondingAuthor":false,"prefix":"","firstName":"Eric","middleName":"","lastName":"Wallace","suffix":""},{"id":621674540,"identity":"5bdc5b9e-5d35-4a14-acc5-4d178b53df91","order_by":4,"name":"Eric W. Ford","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAApElEQVRIiWNgGAWjYBAC9gYGBgmGCgYGAwYQgxjAcwCk8gzJWhjbSNLC3vvwxsd5h+XNGZgP3uYhSgvPcWPLmdsOG+5sYEu2JkqLvUQamzTvtsOMGw7wmEkTZwtIy985h+03HOD/RoIWxobDiUBb2IjUwnOM2bLnWHryzmY2Y8s5RGlhb2O88aPG2nY7e/PDG2+I0YIAzKQpHwWjYBSMglGADwAANjssDga8ha0AAAAASUVORK5CYII=","orcid":"","institution":"University of Alabama at Birmingham","correspondingAuthor":true,"prefix":"","firstName":"Eric","middleName":"W.","lastName":"Ford","suffix":""}],"badges":[],"createdAt":"2026-03-16 18:08:57","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9140964/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9140964/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107486168,"identity":"c6d2e71d-fa9d-46fe-9662-e78decd15e1e","added_by":"auto","created_at":"2026-04-22 02:37:34","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":505626,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9140964/v1/51db52bc-600b-4c24-9463-d1d5e6ce4db1.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Clinical AI Adoption Patterns Among U.S. Hospitals: Comparing Hospital Adoption and Implementation Using AHA Survey Data","fulltext":[{"header":"Introduction","content":"\u003cp\u003eClinical artificial intelligence (AI) refers to a broad set of computational tools, ranging from traditional statistical models to modern machine-learning systems, that support decision-making and workflow in health care.[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e] In practice, clinical AI can analyze large volumes of clinical data (e.g., structured electronic health record fields, imaging, waveforms, and clinical text) to generate predictions, classifications, summaries, or recommendations that clinicians and health systems use to improve care.[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e] When deployed effectively, clinical AI offers several potential benefits: earlier identification of deterioration or high-risk patients, more consistent and timely interpretation of complex data, reduced administrative burden through automation, and improved operational efficiency through better triage and resource allocation; in turn, these gains may translate into safer care, improved patient outcomes, and a more sustainable clinical workforce.[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eDespite this promise, clinical AI adoption and implementation remain uneven across U.S. hospitals and use cases.[\u003cspan additionalcitationids=\"CR5 CR6 CR7\" citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] \u0026ldquo;Adoption\u0026rdquo; (selecting, purchasing, or initiating use of an AI capability) does not guarantee \u0026ldquo;implementation\u0026rdquo; (integrating the tool into real-world workflows at scale, with durable use, performance monitoring, and governance).[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] Many health systems successfully pilot AI tools but struggle to operationalize them, with barriers spanning financial, technical, organizational, and regulatory domains.[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] Common challenges include uncertainty about return on investment, limited interoperability and integration with existing information systems, data quality and representativeness concerns, clinician trust and usability issues, workflow disruption, insufficient staffing for model maintenance and monitoring, and evolving legal and regulatory expectations.[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] These obstacles can prevent promising tools from moving beyond isolated pilots, or can lead to partial implementation that fails to deliver measurable value.\u003c/p\u003e \u003cp\u003eWith this in mind, we examine the clinical AI landscape through the lens of this \u0026ldquo;adopted versus implemented\u0026rdquo; gap. Specifically, using the Fiscal Year 2024 American Hospital Association (AHA) Annual Survey [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] and Item Response Theory (IRT)[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], we assess which clinical AI constructs hospitals report adopting and which are reported as fully implemented in routine operations.[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] Additionally, we identify constructs that appear more likely to stall between adoption and implementation. By distinguishing early uptake from sustained operationalization, our goal is to clarify where hospitals are successfully translating clinical AI into practice and where persistent barriers may be limiting real-world impact.\u003c/p\u003e"},{"header":"METHODS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy design and data source\u003c/h2\u003e \u003cp\u003eWe conducted a cross-sectional, secondary analysis using the Fiscal Year (FY) 2024 AHA Annual Survey Database. The AHA Annual Survey is administered nationally and is refined on an annual basis to align with emerging issues in hospital operations and policy-relevant measurement needs. The FY2024 release included newly added items related to AI implementation. The focus of the current analyses was exploring the adoption and implementation of clinical AI packages.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eMeasures\u003c/h3\u003e\n\u003cp\u003e \u003cb\u003eClinical AI implementation items.\u003c/b\u003e Our primary measures were seven clinical AI implementation items, answered on a 5-level ordered response scale: 1\u0026thinsp;=\u0026thinsp;not implementing, 2\u0026thinsp;=\u0026thinsp;exploring, 3\u0026thinsp;=\u0026thinsp;piloting/testing, 4\u0026thinsp;=\u0026thinsp;expanding, 5\u0026thinsp;=\u0026thinsp;fully integrated. \u0026ldquo;Don\u0026rsquo;t know\u0026rdquo; responses were coded as missing for analysis. Items specifically probed hospitals\u0026rsquo; adoption and implementation of: (1) AI-assisted diagnostics (DIAGAI), (2) predictive analytics for patient care (PAPCAI), (3) clinical decision support tools (CLINAI), (4) AI-assisted surgery (SURGAI), (5) patient communication and education (PCOEAI), (6) population health management (POPHAI), and (7) resource allocation during emergencies (PDMRAI). Additionally, we examined open-text entries for additional AI use cases (see below for approach).\u003c/p\u003e \u003cp\u003e \u003cb\u003eHospital characteristics.\u003c/b\u003e Hospital characteristics were used to describe the analytic sample and to compare hospitals included in the clinical AI IRT estimation sample with those not included. Ownership/control codes were recoded into four categories: government non-federal, nonprofit, for-profit, and federal. Bed size was operationalized as small, medium, or large using a three-category bed size variable derived from AHA bed size code. Region was operationalized as Northeast, South, Midwest, and West using a Census region variable derived from state code. Additional characteristics included teaching proxy, urban or rural proxy, health system membership, and Joint Commission accreditation.\u003c/p\u003e\n\u003ch3\u003eAnalytic approach\u003c/h3\u003e\n\u003cp\u003eWe followed the logic of Chan et al.\u0026rsquo;s application of item response theory (IRT) to AHA survey items, with modifications appropriate for ordered response categories.[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/p\u003e \u003cp\u003e \u003cb\u003eUnidimensionality assessment.\u003c/b\u003e A prerequisite for unidimensional IRT modeling is evidence that the item battery reflects a single latent construct. Exploratory factor analysis using principal factors and an unrotated solution was used to evaluate whether the clinical AI items formed a sufficiently unidimensional scale. Our criteria for \u0026ldquo;sufficient\u0026rdquo; unidimensionality were based on prior IRT guidance used by Chan et al.,[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] including a dominant first factor and strong first-factor loadings across items.\u003c/p\u003e \u003cp\u003e \u003cb\u003eIRT model.\u003c/b\u003e Because clinical AI implementation was measured using ordered stages rather than binary responses, we estimated a graded response IRT model to obtain an item discrimination parameter for each clinical AI use case and category threshold parameters corresponding to transitions across implementation stages. Discrimination parameters characterize how well each item differentiates hospitals along the latent clinical AI implementation maturity continuum. Threshold parameters characterize the latent maturity level at which the probability of endorsing an implementation stage or higher reaches 50 percent. We generated a hospital-level latent clinical AI implementation maturity score (theta) from the fitted model.\u003c/p\u003e \u003cp\u003e \u003cb\u003eSample inclusion and missingness.\u003c/b\u003e \u0026ldquo;Don\u0026rsquo;t know\u0026rdquo; responses were treated as missing. IRT estimation used hospitals contributing non-missing information to the clinical AI battery under the model\u0026rsquo;s likelihood-based estimation, yielding the clinical AI IRT estimation sample. Hospitals outside the estimation sample were retained for descriptive comparisons to characterize selection into the analytic sample.\u003c/p\u003e \u003cp\u003e \u003cb\u003eDescriptive statistics and correlations.\u003c/b\u003e Item means and a correlation matrix for the clinical AI items were computed within the IRT estimation sample. We calculated Pearson correlations using pairwise available observations, consistent with common reporting conventions for survey-based scale development.\u003c/p\u003e \u003cp\u003e \u003cb\u003eOpen-text processing.\u003c/b\u003e Open-text \u0026ldquo;other clinical AI\u0026rdquo; responses were extracted and reviewed to identify recurring themes. We normalized and categorized text using a keyword-based coding approach to support transparent frequency reporting. We defined a documentation-support category a priori to capture \u0026ldquo;digital scribe\u0026rdquo; and closely related terms (for example, \u0026ldquo;AI scribe,\u0026rdquo; \u0026ldquo;ambient scribe,\u0026rdquo; \u0026ldquo;ambient listening,\u0026rdquo; \u0026ldquo;dictation,\u0026rdquo; and \u0026ldquo;transcription\u0026rdquo;), given the prominence of documentation-oriented entries in the open-text field.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cp\u003eThe FY2024 AHA file contained 6,173 hospitals; of these, 2,423-2,640 hospitals answered the newly added clinical AI items (39.3% \u0026ndash; 42.8% of the full sample). Predictive analytics for patient care had the highest response rate (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2,640; 42.8%). Resource allocation during emergencies and AI-assisted surgery had the lowest response rate (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2,423; 39.3%; \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2,470; 40.0%, respectively). We did not include 3,445 hospitals (55.8%) from the graded response IRT model because they did not contribute usable clinical AI item information under the model. Thus, the model was estimated on 2,728 hospitals (44.2% of the full sample).\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e compares hospitals included in the clinical AI IRT estimation sample with those not included. Hospitals included in the estimation sample were larger and more complex. Mean inpatient days (IPDTOT) were higher in the estimation sample (47,407 vs. 28,702), as were mean inpatient discharges (DCTOT; 8,182 vs 3,991) and commercial discharges (THRTDC; 2,075 vs 887). Bed size categories also differed, with \u0026ldquo;large\u0026rdquo; hospitals comprising 20.0% of the estimation sample versus 9.5% of those not included. Additionally, more teaching programs were represented in the estimation sample (7.6% vs 2.1%). Ownership differed substantially: nonprofit hospitals comprised 69.1% of the estimation sample compared with 37.4% of those not included. Conversely, for-profit hospitals comprised 13.5% of the estimation sample and 36.3% of the hospitals not included. Regional distribution also differed, with a higher share of Midwestern hospitals in the estimation sample (32.3% vs 22.9%). CBSA type (metro/micro/rural) did not differ significantly between groups (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.63), despite significant differences for bed size, teaching status, ownership, and region (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eHospital AI Sample Characteristics\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHospital Characteristics\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAI Sample\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNon-respondents\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBed Size\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSmall (\u0026lt;\u0026thinsp;99)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,404\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2,164\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMedium (100\u0026ndash;299)*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e778\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e953\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLarge (\u0026gt;\u0026thinsp;300)*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e546\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e328\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTeaching status\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYes*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2,522\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3,373\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e206\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e72\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOwnership\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNonprofit*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,884\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,288\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFor-profit\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e369\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,252\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePublic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e475\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e905\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccreditation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYes*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2,522\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3,373\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e206\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e72\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRegion\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNortheast\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e369\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e400\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSouth\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e984\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1,486\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMidwest*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e881\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e788\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e488\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e704\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCBSA Type\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetro (\u0026gt;\u0026thinsp;50K)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,849\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2,374\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMicro (10-50K)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e385\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e467\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRural (\u0026lt;\u0026thinsp;10K)*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e494\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e604\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"3\"\u003eNotes: 1. * \u003cem\u003ep \u0026lt;\u003c/em\u003e .001 indicates that AI more likely in use or being implemented\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"3\"\u003e2. Northeast (CT|ME|MA|NH|RI|VT|NJ|NY|PA);\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"3\"\u003eSouth (DE|DC|FL|GA|MD|NC|SC|VA|WV|AL|KY|MS|TN|AR|LA|OK|TX);\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"3\"\u003eMidwest (IL|IN|MI|OH|WI|IA|KS|MN|MO|NE|ND|SD);\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"3\"\u003eWest (AZ|CO|ID|MT|NV|NM|UT|WY|AK|CA|HI|OR|WA)\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e reports item means and inter-item correlations for the clinical AI battery within the IRT estimation sample. Item means indicated the greatest implementation maturity for clinical decision support tools (CLINAI; mean 2.98) and predictive analytics for patient care (PAPCAI; mean 2.87), followed by AI-assisted diagnostics (DIAGAI; mean 2.78). The lowest mean stage was observed for resource allocation during emergencies (PDMRAI; mean 1.62), consistent with low uptake for that use case. Pairwise correlations were uniformly positive and moderate-to-strong (range 0.37\u0026ndash;0.81), with the strongest correlation observed between predictive analytics and clinical decision support (\u003cem\u003er\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.81). The correlation structure supported a single underlying construct of clinical AI implementation maturity, consistent with the factor analytic evidence reported in the Appendix.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eItem Means and Correlations\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAI Tools\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1. Resource allocation during emergencies\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.622\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2. AI assisted surgery\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.774\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.478\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3. Patient communication and education\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2.295\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.365\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.448\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4. AI assisted diagnostics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2.776\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.371\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.451\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.576\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5. Population health management\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2.186\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.537\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.551\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.639\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.558\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6. Clinical decision support tools\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2.977\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.390\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.445\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.618\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.633\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.637\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7. Predictive analytics for patient care\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2.865\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.434\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.460\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.631\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.675\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.666\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.809\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents graded response IRT results. All items demonstrated monotonic threshold progression (Diff\u0026thinsp;\u0026ge;\u0026thinsp;2\u0026thinsp;\u0026lt;\u0026thinsp;Diff\u0026thinsp;\u0026ge;\u0026thinsp;3\u0026thinsp;\u0026lt;\u0026thinsp;Diff\u0026thinsp;\u0026ge;\u0026thinsp;4\u0026thinsp;\u0026lt;\u0026thinsp;Diff\u0026thinsp;=\u0026thinsp;5), supporting ordered-category functioning and model fit. Resource allocation during emergencies (PDMRAI) exhibited the highest \u0026ldquo;full integration\u0026rdquo; threshold (Diff\u0026thinsp;=\u0026thinsp;5: 2.72), indicating the greatest difficulty reaching full integration on the latent maturity continuum. AI-assisted surgery (SURGAI) showed the second highest full-integration threshold (2.40). In contrast, clinical decision support tools (CLINAI; 0.89) and predictive analytics (PAPCAI; 0.97) had the lowest full-integration thresholds, indicating comparatively easier progression to advanced implementation. Several adjacent items exhibited non-overlapping confidence intervals for key threshold estimates, indicating distinct tiers of implementation difficulty rather than a smooth continuum. Discrimination estimates were highest for predictive analytics (a\u0026thinsp;=\u0026thinsp;4.56) and clinical decision support (a\u0026thinsp;=\u0026thinsp;3.93), indicating strong differentiation of hospitals along the latent maturity trait, and were lowest for emergency resource allocation (a\u0026thinsp;=\u0026thinsp;1.59) and AI-assisted surgery (a\u0026thinsp;=\u0026thinsp;1.89).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eItem Discrimination and Difficulty Scores\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eitem (full label)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDiscrimination (a), 95% CI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDiff\u0026thinsp;\u0026ge;\u0026thinsp;2\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDiff\u0026thinsp;\u0026ge;\u0026thinsp;3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eDiff\u0026thinsp;\u0026ge;\u0026thinsp;4\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eDiff\u0026thinsp;=\u0026thinsp;5\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResource allocation during emergencies \u0026ndash; (PDMRAI)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.589\u003c/p\u003e \u003cp\u003e(1.439 to 1.738)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.256\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.680\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2.072\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2.721\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAI assisted surgery \u0026ndash; (SURGAI)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.886\u003c/p\u003e \u003cp\u003e(1.725 to 2.048)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.224\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.113\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.530\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2.404\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePatient communication and education \u0026ndash; (PCOEAI)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.653\u003c/p\u003e \u003cp\u003e(2.468 to 2.837)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.609\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.453\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.073\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.957\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAI assisted diagnostics \u0026ndash; (DIAGAI)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.763\u003c/p\u003e \u003cp\u003e(2.575 to 2.951)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.761\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.095\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.496\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.137\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePopulation health management \u0026ndash; (POPHAI)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.346\u003c/p\u003e \u003cp\u003e(3.094 to 3.598)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.366\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.583\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.131\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.466\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClinical decision support tools \u0026ndash; (CLINAI)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.926\u003c/p\u003e \u003cp\u003e(3.633 to 4.219)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.847\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.009\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.281\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.889\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePredictive analytics for patient care \u0026ndash; (PAPCAI)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.564\u003c/p\u003e \u003cp\u003e(4.192 to 4.936)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.751\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.063\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.408\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.969\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eAmong hospitals providing an open-text entry (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;256), documentation-oriented AI dominated the responses. Digital scribe and closely related ambient documentation concepts (e.g., \u0026ldquo;ambient listening,\u0026rdquo; \u0026ldquo;ambient documentation,\u0026rdquo; \u0026ldquo;AI scribe,\u0026rdquo; dictation/transcription, and vendor tools such as DAX/Dragon/Nuance and DeepScribe) appeared in 167 entries (65.2%), with explicit \u0026ldquo;ambient\u0026rdquo; terminology present in 95 entries (37.1%). Generative AI and large-language-model-style assistants (e.g., \u0026ldquo;copilot,\u0026rdquo; \u0026ldquo;GPT,\u0026rdquo; \u0026ldquo;RAG,\u0026rdquo; handoff tools, and chart summarization) appeared in 34 entries (13.3%). Smaller clusters referenced revenue cycle/coding (15 entries; 5.9%), robotics (11 entries; 4.3%), and patient messaging/portal automation (11 entries; 4.3%). A residual set of 26 entries (10.2%) was too vague for reliable thematic classification (e.g., \u0026ldquo;AI\u0026rdquo; without further specification).\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eThe FY2024 AHA clinical AI items support a coherent, staged construct of \u0026ldquo;clinical AI implementation maturity\u0026rdquo; and also demonstrate that clinical AI adoption is not a single phenomenon. Factor and correlation results indicate a single underlying dimension, and the graded response IRT model produced well-behaved item thresholds across all seven items, with monotonic progression for each stage boundary. These results provide two actionable insights: (1) some clinical AI capabilities are systematically easier to move from exploration to piloting and from expansion to full integration, and (2) a subset of capabilities remain meaningfully more difficult to implement, even among hospitals with evidence of broader AI implementation maturity.\u003c/p\u003e \u003cp\u003ePatterns in item difficulty align with the operational and governance burden implied by each use case. Resource allocation during emergencies exhibited the highest thresholds for advanced implementation, consistent with both low response and low reported maturity. These results validate the ethical complexity and general reluctance for states and healthcare systems to establish clear guidance on scarce resource allocation.[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] While research suggests that predictive modeling can optimize patient flow and facilitate triage decisions, the lack of foundational decision-making guidance through Crisis Standards of Care may limit the practical utility of these tools by healthcare systems and providers. Thresholds should also be determined for when implementation advances from operational purposes of patient flow forecasting, efficiency, and supply chain optimization to clinical purposes of emergency resource allocation. Defining these AI implementation categories in the AHA survey may result in higher response for this survey item.\u003c/p\u003e \u003cp\u003eAI-assisted surgery also required higher maturity to reach piloting, expansion, and full integration. These use cases typically impose heavier requirements for safety management, cross-disciplinary workflow redesign, and organizational accountability than tools such as predictive analytics and decision support, which can be implemented earlier as \u0026ldquo;augmented decision\u0026rdquo; functions embedded within existing clinical pathways. Evidence from implementation-focused syntheses repeatedly emphasizes that clinical AI impact is mediated by sociotechnical fit, governance, and the ability to sustain performance monitoring over time, not by model availability alone.[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] The threshold ordering observed here offers quantitative reinforcement of that point and suggests that \u0026ldquo;difficult\u0026rdquo; AI use cases represent a higher bar of organizational capability rather than simply a slower vendor market.\u003c/p\u003e \u003cp\u003eA value proposition that combines clinical effectiveness with operational efficiency likely accelerates adoption. Predictive analytics for patient care and clinical decision support tools were among the least difficult items to integrate at advanced stages and exhibited the strongest discrimination. Tools in these categories can plausibly be justified on safety and effectiveness grounds while also being framed as productivity or throughput interventions; for example, earlier deterioration detection to reduce ICU transfers, standardization to reduce variation, and risk stratification to focus clinician time. A national health-system survey conducted in the early generative AI period similarly found widespread activity in clinical risk stratification and imaging-related AI, but variable success and persistent barriers such as tool immaturity, financial concerns, and regulatory uncertainty.[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] The same survey identified \u0026ldquo;ambient notes\u0026rdquo; as the one use case with adoption activity reported by all respondents, reinforcing the idea that immediate workflow relief and productivity benefits can drive rapid uptake even when longer-term evaluation remains incomplete.\u003c/p\u003e \u003cp\u003eNotably, emergency resource allocation differs from the higher-uptake use cases in one critical respect: value realization often depends on coordination beyond the walls of a single hospital.[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] Preparedness and surge response become meaningfully more effective when response capacity is defined across an interconnected network rather than within one facility. Regional coordination requires reliable data sharing, patient tracking, and operational visibility across institutions and emergency management partners. Peer-reviewed work in disaster medicine has described the heterogeneous, multi-system landscape of regional information-sharing and coordination platforms during large-scale medical surge events and has highlighted persistent barriers to effective information sharing, including fragmentation and interoperability gaps.[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] Health information exchange itself remains a multi-stage organizational assimilation process with distinct initiation, adoption, implementation, and institutionalization phases.[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] Systematic reviews and empirical studies indicate that HIE can reduce duplicative testing and emergency department costs and can improve certain operational outcomes, but routine use remains uneven and often context-dependent.[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] The EMS and emergency care continuum provides a concrete illustration: prehospital-to-hospital data linkage is frequently limited by interoperability, connectivity, and patient matching constraints, despite clear potential value for system-level emergency coordination.[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] These realities create a plausible structural explanation for why emergency resource allocation AI appears \u0026ldquo;hard\u0026rdquo; in the IRT results. Even high-capability hospitals may be constrained by incomplete regional data integration, limited cross-institutional governance, and the episodic nature of the operational value proposition outside active surges.\u003c/p\u003e \u003cp\u003eHospital characteristics associated with greater clinical AI uptake reinforce concerns about capability stratification. Larger hospitals, those with teaching programs, and nonprofit institutions were more likely to appear in the clinical AI estimation sample, suggesting that adoption and maturity are concentrated among organizations with greater capital, informatics capacity, and change infrastructure. Prior work on the diffusion of advanced EHR functions has documented an \u0026ldquo;advanced use\u0026rdquo; divide, in which certain hospital types lag in adopting higher-order capabilities even when basic EHR adoption becomes widespread.[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] Studies focused on critical access hospitals similarly report persistent difficulty keeping pace with more advanced health IT functions, even after broad EHR adoption.[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] AI maturity results in the present analysis are consistent with this trend, which may produce a practical bifurcation between AI-rich and AI-poor delivery settings. Such divergence has direct implications for health equity, as differential access to advanced clinical decision support, risk stratification, and workflow augmentation can translate into differences in quality, safety, and access to capacity over time.\u003c/p\u003e \u003cp\u003eConversely, a lack of association for CBSA type is notable in this context. Urban versus rural market classification did not differentiate clinical AI uptake in the descriptive comparison, suggesting that organizational resources and governance capacity may dominate simple market-type proxies in predicting measured adoption. A sharper market structure signal may require more granular measures, such as system-level consolidation, vendor footprint, or resource pooling across networks, particularly when AI tools are increasingly procured and governed at the health system level rather than at the facility level.\u003c/p\u003e \u003cp\u003eFinally, open-text \u0026ldquo;other clinical AI\u0026rdquo; responses provide an additional signal about where hospitals are investing: ambient documentation and digital scribes. This is consistent with external evidence that ambient notes has become a highly salient and widely pursued use case in US health systems. Early empirical evaluations of ambient AI scribe interventions have reported reductions in documentation burden and burnout measures, although evidence remains emergent and implementation introduces new governance needs related to privacy, consent, and clinical documentation quality.[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] Open-text dominance of documentation support suggests that staged survey batteries that omit documentation AI risk undercounting a major pathway for AI adoption, particularly in the early generative AI era. Survey evolution toward structured staged items for ambient documentation would improve comparability across organizations and over time would facilitate more precise measurement of adoption maturity in a category that appears to be diffusing rapidly.\u003c/p\u003e \u003cp\u003eThis study has several important limitations. First, it is a cross-sectional, secondary analysis of hospital-level survey data and therefore cannot assess temporal progression or causal drivers of movement across implementation stages. Second, the clinical AI measures are self-reported, and treatment of \u0026ldquo;don\u0026rsquo;t know\u0026rdquo; responses as missing may introduce reporting error and selection bias. Third, the FY2024 AHA clinical AI items capture only a limited set of structured use cases, and the open-text findings suggest that prominent applications such as ambient documentation may be incompletely represented.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eThere are several practical implications for this study. The ordered-response IRT approach preserves the staged nature of implementation reporting and translates it into item-level thresholds that identify where hospitals are most likely to stall, such as remaining in exploration rather than progressing to piloting or expansion. It also produces a standardized latent maturity score that can support benchmarking and future work linking maturity to organizational characteristics, interoperability capacity, and downstream outcomes. Chan and colleagues previously demonstrated that IRT can reveal differential implementation difficulty in health IT using AHA survey items, and the present study extends that logic to clinical AI using an ordered-response model aligned with the FY2024 staging format.[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] Monotonic thresholds across all items further support the integrity of the ordered categories and reduce concern that response options were functioning erratically.\u003c/p\u003e \u003cp\u003eOverall, these findings suggest that hospital clinical AI adoption is best understood not as a simple yes-or-no phenomenon, but as a staged process of implementation maturity in which some use cases advance more readily than others. Clinically, this matters as uneven implementation may shape where patients and clinicians actually experience the benefits of AI in practice, particularly in decision support, predictive analytics, and workflow augmentation, while more operationally, ethically, and organizationally complex applications remain less mature. At the same time, interpretation should remain grounded in the practical meaning of the IRT parameters: higher thresholds indicate that a use case requires substantially greater organizational maturity to reach advanced stages, whereas higher discrimination indicates a use case that more strongly differentiates hospitals along the maturity continuum. Together, these results provide a framework that can help health systems, policymakers, and implementation leaders identify where adoption stalls and target the governance, infrastructure, and workforce supports needed to translate AI implementation into reliable and equitable clinical impact.\u003c/p\u003e"},{"header":"Statements and declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e: All authors contributed to the study conception and design. Data analysis, and the drafting of the Methods and Results sections, were performed by Eric W. Ford. Drafts of the Background, Introduction, and Conclusion sections were prepared by Emma B. Sartin. Elayne L. Wilson, Hugh B. Roland, and Eric Wallace drafted the Discussion sections relevant to their respective areas of expertise. All authors contributed to revision of the manuscript and edited the final version for clarity and coherence. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding Declaration:\u0026nbsp;\u003c/strong\u003eThere was no external or internal funding for this research.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSoftware and reporting.\u0026nbsp;\u003c/strong\u003eAll analyses were conducted in StataNow 19.5 (StataCorp, College Station, Texas). Item parameters (discrimination and thresholds), item means, and the correlation matrix were reported to support interpretation and reproducibility.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eHuman subjects considerations.\u0026nbsp;\u003c/strong\u003eThis study is not considered human subjects research as it is a secondary analysis using hospital-level survey data without patient identifiers.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical trial number:\u0026nbsp;\u003c/strong\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics and Consent to Participate declaration\u003c/strong\u003e: Not applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eRajkomar, A., J. Dean, and I. Kohane, \u003cem\u003eMachine learning in medicine.\u003c/em\u003e New England Journal of Medicine, 2019. \u003cstrong\u003e380\u003c/strong\u003e(14): p. 1347-1358.\u003c/li\u003e\n \u003cli\u003eOrganization, W.H., \u003cem\u003eEthics and governance of artificial intelligence for health: large multi-modal models. WHO guidance\u003c/em\u003e. 2024: World Health Organization.\u003c/li\u003e\n \u003cli\u003eSpear, J., J.M. Ehrenfeld, and B.J. Miller, \u003cem\u003eApplications of Artificial Intelligence in Health Care Delivery.\u003c/em\u003e Journal of Medical Systems, 2023. \u003cstrong\u003e47\u003c/strong\u003e(1): p. 121.\u003c/li\u003e\n \u003cli\u003eShortliffe, E.H. and M.J. Sep\u0026uacute;lveda, \u003cem\u003eClinical decision support in the era of artificial intelligence.\u003c/em\u003e Jama, 2018. \u003cstrong\u003e320\u003c/strong\u003e(21): p. 2199-2200.\u003c/li\u003e\n \u003cli\u003eSendak, M.P., et al., \u003cem\u003eA path for translation of machine learning products into healthcare delivery.\u003c/em\u003e EMJ Innov, 2020. \u003cstrong\u003e10\u003c/strong\u003e: p. 19-00172.\u003c/li\u003e\n \u003cli\u003eKelly, C.J., et al., \u003cem\u003eKey challenges for delivering clinical impact with artificial intelligence.\u003c/em\u003e BMC medicine, 2019. \u003cstrong\u003e17\u003c/strong\u003e(1): p. 195.\u003c/li\u003e\n \u003cli\u003eKueper, J.K., et al., \u003cem\u003eArtificial intelligence and primary care research: a scoping review.\u003c/em\u003e The annals of family medicine, 2020. \u003cstrong\u003e18\u003c/strong\u003e(3): p. 250-258.\u003c/li\u003e\n \u003cli\u003eKanakubo, T. and H. Kharrazi, \u003cem\u003eComparing the Trends of Electronic Health Record Adoption Among Hospitals of the United States and Japan.\u003c/em\u003e Journal of Medical Systems, 2019. \u003cstrong\u003e43\u003c/strong\u003e(7): p. 224.\u003c/li\u003e\n \u003cli\u003eSculley, D., et al., \u003cem\u003eHidden technical debt in machine learning systems.\u003c/em\u003e Advances in neural information processing systems, 2015. \u003cstrong\u003e28\u003c/strong\u003e.\u003c/li\u003e\n \u003cli\u003eCarvalho, J.V., \u0026Aacute;. Rocha, and A. Abreu, \u003cem\u003eMaturity Models of Healthcare Information Systems and Technologies: a Literature Review.\u003c/em\u003e Journal of Medical Systems, 2016. \u003cstrong\u003e40\u003c/strong\u003e(6): p. 131.\u003c/li\u003e\n \u003cli\u003eAssociation, A.H., \u003cem\u003eAHA Annual Survey of Hospitals\u003c/em\u003e, A.H. Association, Editor. 2025.\u003c/li\u003e\n \u003cli\u003eEmbretson, S.E. and S.P. Reise, \u003cem\u003eItem Response Theory\u003c/em\u003e. 2013: Psychology Press.\u003c/li\u003e\n \u003cli\u003eAssociation, A.H., \u003cem\u003eAHA Annual Survey Information Technology Supplement 2024 (instrument)\u003c/em\u003e, A.H. Association, Editor. 2025.\u003c/li\u003e\n \u003cli\u003eChan, K.S., et al., \u003cem\u003eAssessing electronic health record implementation challenges using item response theory.\u003c/em\u003e Am J Manag Care, 2016. \u003cstrong\u003e22\u003c/strong\u003e(12): p. e409-e415.\u003c/li\u003e\n \u003cli\u003eServices, U.D.o.H.a.H., \u003cem\u003eTechnical Resources, Assistance Center, and Information Exchange: Mass Casualty Hospital Capacity Expansion Toolkit\u003c/em\u003e, A.f.S.P.a. Response, Editor. 2025.\u003c/li\u003e\n \u003cli\u003eSharma, M., et al., \u003cem\u003eArtificial Intelligence Applications in Health Care Practice: Scoping Review.\u003c/em\u003e Journal of Medical Internet Research, 2022. \u003cstrong\u003e24\u003c/strong\u003e(10): p. e40238.\u003c/li\u003e\n \u003cli\u003ePoon, E.G., et al., \u003cem\u003eAdoption of artificial intelligence in healthcare: survey of health system priorities, successes, and challenges.\u003c/em\u003e Journal of the American Medical Informatics Association, 2025. \u003cstrong\u003e32\u003c/strong\u003e(7): p. 1093-1100.\u003c/li\u003e\n \u003cli\u003eRobles Mendo, I., et al., \u003cem\u003eMachine Learning in Medical Emergencies: a Systematic Review and Analysis.\u003c/em\u003e Journal of Medical Systems, 2021. \u003cstrong\u003e45\u003c/strong\u003e(10): p. 88.\u003c/li\u003e\n \u003cli\u003eLee, C.J., et al., \u003cem\u003eUse of Information Technology Systems for Regional Health Care Information-Sharing and Coordination During Large-Scale Medical Surge Events.\u003c/em\u003e Disaster Medicine and Public Health Preparedness, 2023. \u003cstrong\u003e18\u003c/strong\u003e: p. e1.\u003c/li\u003e\n \u003cli\u003eEsmaeilzadeh, P. and M. Sambasivan, \u003cem\u003eHealth Information Exchange (HIE): A literature review, assimilation pattern and a proposed classification for a new policy approach.\u003c/em\u003e Journal of Biomedical Informatics, 2016. \u003cstrong\u003e64\u003c/strong\u003e: p. 74-86.\u003c/li\u003e\n \u003cli\u003eHersh, W.R., et al., \u003cem\u003eOutcomes From Health Information Exchange: Systematic Review and Future Research Needs.\u003c/em\u003e JMIR Medical Informatics, 2015. \u003cstrong\u003e3\u003c/strong\u003e(4): p. e39.\u003c/li\u003e\n \u003cli\u003eMartin, T.J., et al., \u003cem\u003eHealth Information Exchange in Emergency Medical Services.\u003c/em\u003e Applied Clinical Informatics, 2018. \u003cstrong\u003e9\u003c/strong\u003e(4): p. 884-891.\u003c/li\u003e\n \u003cli\u003eAdler-Milstein, J., et al., \u003cem\u003eElectronic health record adoption in US hospitals: the emergence of a digital \u0026quot;advanced use\u0026quot; divide.\u003c/em\u003e Journal of the American Medical Informatics Association, 2017. \u003cstrong\u003e24\u003c/strong\u003e(6): p. 1142-1148.\u003c/li\u003e\n \u003cli\u003eApathy, N.C., A.J. Holmgren, and J. Adler-Milstein, \u003cem\u003eA decade post-HITECH: Critical access hospitals have electronic health records but struggle to keep up with other advanced functions.\u003c/em\u003e Journal of the American Medical Informatics Association, 2021. \u003cstrong\u003e28\u003c/strong\u003e(9): p. 1947-1954.\u003c/li\u003e\n \u003cli\u003eOlson, K.D., et al., \u003cem\u003eUse of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout.\u003c/em\u003e JAMA Network Open, 2025. \u003cstrong\u003e8\u003c/strong\u003e(10): p. e2534976.\u003c/li\u003e\n \u003cli\u003eChan, K.S., et al., \u003cem\u003eAssessing electronic health record implementation challenges using item response theory.\u003c/em\u003e American Journal of Managed Care, 2016. \u003cstrong\u003e22\u003c/strong\u003e(12): p. e409-e415.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Clinical artificial intelligence, Hospital adoption, Implementation maturity, Item response theory, Health informatics, Predictive analytics","lastPublishedDoi":"10.21203/rs.3.rs-9140964/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9140964/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e: Hospitals are investing in clinical artificial intelligence (AI), yet most efforts to classify the types and extent AI uptake rely on ad hoc counts of tools or binary adoption indicators. A scalable, psychometrically grounded measure of “clinical AI implementation maturity” would strengthen benchmarking and improve research on determinants and consequences of adoption.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eObjective\u003c/strong\u003e: To create a hospital-level measure of clinical AI use and adoption using staged implementation items from the American Hospital Association’s (AHA’s) Annual Survey.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e: Ordered response items describing implementation stage across multiple clinical AI functions were modeled using an item response theory (IRT) framework. We evaluated the suitability of a unidimensional maturity construct by examining model fit indicators and the coherence of item behavior. Item discrimination parameters and ordered category thresholds were used to determine how well each AI function differentiated hospitals along the maturity continuum and whether response categories reflected interpretable progression.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e: The fitted model supported a stable maturity gradient with meaningful between-hospital variation. Item discrimination estimates indicated that certain AI functions were substantially more informative for distinguishing hospitals at different maturity levels than others, while some functions contributed limited differentiation. Category thresholds were generally ordered, supporting the interpretation of staged implementation as a progression. Threshold patterns suggested that adoption is not random across functions; some capabilities tend to appear earlier in maturity trajectories, whereas others cluster at later stages and better discriminate among higher-maturity hospitals. The resulting scores can be used as continuous measures with uncertainty intervals for comparisons and stratified analyses.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e: Clinical AI implementation maturity can be measured as a latent construct using routine survey data and IRT methods. A psychometrically validated maturity score provides a foundation for benchmarking, risk adjustment in comparative studies, and policy-relevant monitoring of clinical AI diffusion across healthcare delivery systems.\u003c/p\u003e","manuscriptTitle":"Clinical AI Adoption Patterns Among U.S. Hospitals: Comparing Hospital Adoption and Implementation Using AHA Survey Data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-13 05:20:54","doi":"10.21203/rs.3.rs-9140964/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a9f34dcf-0f7c-44f6-bbf0-553ab97695e8","owner":[],"postedDate":"April 13th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-20T16:25:46+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-13 05:20:54","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9140964","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9140964","identity":"rs-9140964","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}