Artificial Intelligence Platform Architecture for Hospital Systems: Systematic Review

Introduction

Background

Artificial intelligence (AI) is often seen as a game changer for health care, yet many hospitals face challenges in scaling it beyond pilot projects due to limited resources and infrastructure. Rising chronic disease burden due to aging populations and increased demand for personalization has exacerbated these challenges [1]. Traditional hospital information systems (Hospital Information System [HIS], Laboratory Information Systems [LIS], and Picture Archiving and Communication System [PACS]) have become administrative tools without analytic or decision-support value [2]. Wide varieties of multimodal health data are now available along with advanced AI methods, such as image recognition, natural language processing, and predictive analytics, that are maturing rapidly [3]. However, most hospitals face difficulties in integrating these approaches within routine clinical workflows [4].

Research Gap

Many hospitals have not built a system-level roadmap for scaling up and governance structure to support the use of AI pilot projects [5]. Fragmented infrastructures worsen this challenge: siloed PACS and LIS [6] limited interoperability across electronic health record (EHR) vendors due to inconsistent Health Level 7 (HL7) Fast Healthcare Interoperability Resources (FHIR) adoption [7], and heterogeneous Internet of Things (IoT) data streams [8]. Reported implementations typically target isolated tasks, such as radiology computer-aided diagnosis [9] or triage algorithms. Focusing on isolated tasks does not facilitate enterprise-wide implementation of AI, as it fails to address the need for integrated and scalable solutions. In this context, hospital executives do not have sensible roadmaps to translate fragmented digital environments into integrated AI platforms that are based on evidence [10].

Existing Frameworks and Development of a Hospital-Specific Architecture

The hospital system-wide AI platform should be built over an existing framework and not a new one. To build a strong theory base, 4 representative frameworks were reviewed: (1) the Healthcare Information and Management Systems Society Digital Health Framework [11], (2) the World Health Organization (WHO) HIS framework [12], (3) sociotechnical theory [13], and (4) the National Institute of Standards and Technology (NIST) Big Data Reference Architecture (BDRA) [14]. Each offers distinct perspectives. The Healthcare Information and Management Systems Society emphasizes digital maturity and interoperability; the World Health Organization HIS highlights governance and service delivery; sociotechnical theory stresses the interaction of people, processes, and technologies; and BDRA provides a detailed blueprint for big data infrastructures. Comparison shows overlap in infrastructure and data integration, but divergences in analytics, governance, and algorithm lifecycle (Table 1).

Despite important similarities, none of these frameworks fully address the unique requirements for hospital-wide AI implementation, particularly in terms of comprehensive integration and lifecycle management. To bridge this gap, we applied a merge-and-normalize approach by extracting core constructs, clustering overlapping elements, and adding components related to AI where the original was missing, especially on algorithm lifecycle management. This process produced 5 interoperable layers tailored to hospital AI contexts: infrastructure, data, algorithm, application, and security and compliance (Figure 1). Specifically, (1) the infrastructure layer covers compute, storage, and network foundations; (2) the data layer ensures standards such as HL7 FHIR and Digital Imaging and Communications in Medicine, multimodal integration across EHR, LIS, and IoT, and quality management; (3) the algorithm layer introduces the AI lifecycle, including development, validation, monitoring, and operations; (4) the application layer emphasizes workflow integration, such as decision support, triage, and patient-facing tools; and (5) the security and compliance layer ensures privacy, accountability, and governance.

‎

Objectives

On the basis of identified gaps and synthesized frameworks, this study intends to propose a 5-layer AI platform architecture specific to hospitals, to systematically review the empirical studies by mapping evidence onto the framework, and to synthesize theoretical and empirical insights to offer practical advice for hospital administrators, policymakers, and developers who want to build a scalable, secure, and clinically integrated AI ecosystem. Two research questions (RQs) guide this process: (1) RQ1: What empirical evidence supports each layer of the proposed 5-layer architecture? and (2) RQ2: How do the interrelationships among layers reveal strengths and gaps in hospital-level AI research?

Methods

Study Design

This systematic review followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 and PRISMA-P (Preferred Reporting Items for Systematic Review and Meta-Analysis Protocols) guidelines [15] and was prospectively registered in PROSPERO (CRD420251133590). A completed PRISMA checklist is provided in Checklist 1. The 5-layer architecture derived from existing frameworks was adopted as an a priori coding structure for evidence mapping for identifying studies on hospital AI implementation and aligning their findings with this architecture to link theoretical constructs with real-world practice.

Search Strategy

A comprehensive search was conducted in Web of Science, Embase, PubMed, and Scopus from their inception to May 23, 2025. To maximize sensitivity, the strategy combined free-text keywords with controlled vocabulary terms. Search terms included “Hospital Management,” “Healthcare Management,” “Hospital Operations,” “Healthcare Administration,” “Medical Administration,” “Hospital Information Systems,” “AI Deployment,” “AI Implementation,” “AI Integration,” “Artificial Intelligence,” “AI Applications,” “Large Language Models,” “Transformer Applications,” “Hospital Operations Optimization,” “Clinical Workflow Improvement,” “Resource Allocation,” and “Patient Flow Management.” The full search strings for all databases are detailed in Multimedia Appendix 1.

In addition, reference lists of all included articles were manually screened to identify further eligible studies not captured by the electronic search.

Eligibility Criteria

Included studies were in a hospital setting and documented real-world AI use (eg, machine learning, deep learning, natural language processing, predictive analytics, and expert systems) that presented measurable outcomes, including enhanced efficiency, decreased costs/errors, optimized resource usage, improved patient experience, or accurate diagnosis/decision-making enhancement. Studies that did not use AI were excluded, along with reviews, commentaries, conference abstracts, non-English studies, and studies without an abstract. The summary of inclusion and exclusion criteria can be seen in Textbox 1.

Screening

Two reviewers independently screened titles and abstracts, removed duplicates, and assessed preliminary eligibility. The full-text evaluation of the relevant studies was conducted based on the inclusion and exclusion criteria. Reviewers were not blinded to study authorship or outcomes. Discrepancies were resolved by discussion and consensus. The reliability of screening was evaluated through the Cohen κ statistic, and the results were summarized in a 2×2 contingency table.

Data Extraction and Synthesis

Data were extracted using a standardized form, including bibliographic details, study design, clinical domain, objectives, clinical applications, and limitations. Explicit coding definitions were used to enable mapping of each study to the 5-layer architecture (infrastructure, data, algorithm, application, and security and compliance). To promote transparency, the coding criteria for each layer are given in Table 2. Each layer was scored on a 0 to 5 point ordinal maturity scale conceptually aligned with the Capability Maturity Model Integration framework [16] with level definitions and examples detailed in Table 3. Two reviewers independently performed the scoring, and discrepancies were resolved by consensus.

Quantitative synthesis summarized the average maturity scores and SDs for each layer across studies and by study design. Weighted co-occurrence matrices were computed to quantify the cumulative maturity shared between layers, whereas weighted Jaccard similarity indices measured the strength of cross-layer coupling. These metrics were visualized in heatmaps to illustrate maturity distribution and interlayer relationships. Qualitative synthesis identified thematic patterns, gaps, and illustrative cases to demonstrate interactions among layers in real-world contexts.

Quality Assessment

The Critical Appraisal Skills Programme (CASP) tool [17] was applied to assess the quality of included studies because it can be used across various study types. Using checklists for qualitative, quantitative, and mixed methods research, the Critical Appraisal Tools (CATS) tool is used more often than other important assessments, such as the Joanna Briggs Institute, Cochrane, and GRADE. All articles were evaluated for methodology, reliability, interpretation, and usability.

Statistical Analysis

Descriptive statistics were used to summarize the maturity distribution of the 5 layers, expressed as mean scores and SDs across all included studies. Studies were further stratified by methodological category to compare average maturity levels within and across study designs.

Cross-layer relationships were analyzed using weighted co-occurrence and weighted Jaccard similarity metrics derived from the 0 to 5 maturity scores.

The weighted co-occurrence is defined as follows:

$W_{co}\left(A,B\right)={\sum }_{i=1}^n\text{min}\left(A_i,B_i\right)$

where $A_i$and $B_i$represent the maturity scores (0‐5) assigned to layers $A$ and $B$ in study $i$.

The weighted Jaccard similarity is defined as follows:

$J_w\left(A,B\right)=\frac{{\sum }_{i=1}^n\text{min}\left(A_i,B_i\right)}{{\sum }_{i=1}^n\text{max}\left(A_i,B_i\right)}$

where $A_i$and $B_i$represent the maturity scores (0‐5) assigned to layers $A$ and $B$ in study $i$.

Interrater reliability during study selection was quantified using Cohen κ statistic, defined as follows:

$\mathrm{\kappa }=\frac{p_o-p_e}{1-p_e}$

where $p_o$ is the observed agreement, and $p_e$ is the expected agreement by chance.

Ethical Considerations

This systematic review does not involve human participants, identifiable patient data, or protected health information. All data analyzed in this review were obtained from publicly accessible publications. Therefore, an ethical review was not required under Zhongnan Hospital of Wuhan University’s secondary research policies. The study complied with the Declaration of Helsinki and institutional guidelines for secondary data use.

Results

Study Selection

A total of 283 records were identified, of which 257 were from databases and 26 from other sources. After removing duplicates, 255 records were available to undergo title and abstract screening. At this stage, 159 records were excluded, comprising 56 non-AI cases, 25 review articles, and 78 cases not completed in the hospital. Of 96 full texts assessed for eligibility, 67 were ruled out for the following reasons: not original research (n=34), lacked implementation details (n=18), or practical AI use was not described (n=15). In the end, a total of 29 studies [18-46] were included in the review, representing 10.2% of the 283 records. Figure 2 illustrates the selection process. Screening was done by 2 reviewers independently, who demonstrated high interrater reliability (Cohen κ=0.98), suggesting almost perfect agreement. The 2×2 contingency table is shown in Multimedia Appendix 2.

‎

Study Characteristics

The characteristics of the 29 included studies are summarized in Tables 4 and 5. Clinical coverage was broad: nonspecific or cross-specialty applications were most frequent (10/29, 34.5%) [18-21,25-27,30,40,45], followed by radiology (6/29, 20.7%) [22,24,29,43,44,46], emergency medicine (5/29, 17.2%) [28,31,33,37,39], cardiology (2/29, 6.9%) [34,41], and gynecology (2/29, 6.9%) [23,38]. Surgery [36], chronic disease [35], nursing [32], and psychiatry [42] each contributed 1 study.

In terms of study design, diagnostic test studies (8/29, 27.6%) [18,22-24,31,36,44,46] were the most common. Other study designs included cohort studies (6/29, 20.7%) [20,32,34,38,39,43], qualitative studies (6/29, 20.7%) [21,26,30,35,40,45], clinical prediction studies (5/29, 17.2%) [27,28,33,37,41], cross-sectional studies (3/29, 10.3%) [19,25,42], and economic evaluation studies (1/29, 3.4%) [29].

Geographically, the majority of studies originated from high-income countries (Table 6). The United States accounted for nearly half (12/29, 41.3%) [18,20,24,26,29,36,39-41,43-45], reflecting strong emphasis on EHR-linked AI, radiology, cardiology, and workflow optimization. European contributions included the United Kingdom (3/29, 10.3%) [22,34,38], Germany (2/29, 6.9%) [21,30], Italy (2/29, 6.9%) [19,42], France (1/29, 3.4%) [33], the Netherlands (2/29, 6.9%) [28,37], and Switzerland (1/29, 3.4%) [46], with common emphases on General Data Protection Regulation compliance, data sharing, and national cardiovascular initiatives. China contributed 2 studies (2/29, 6.9%) [25,35], reporting multicenter deployments (eg, DeepSeek across 90 tertiary hospitals) and AI integration with blockchain and wearable technologies. Korea contributed 2 studies (2/29, 6.9%) [27,32], focusing on nursing decision support and interoperable clinical decision support systems. Emerging economies were also represented: India (blockchain-enabled gynecology) [23] and Uganda (trauma triage in low-resource settings) [31]. Collectively, these geographic patterns demonstrate United States and European dominance but also highlight distinct implementation trajectories and challenges from Asia and other regions.

Quality Assessment

The quality of the 29 studies was evaluated using the CASP standard. The studies comprised 6 study types: economic evaluations (n=1) [29], clinical prediction studies (n=5) [27,28,33,37,41], diagnostic test studies (n=8) [18,22-24,31,36,44,46], cohort studies (n=6) [20,32,34,38,39,43], qualitative studies (n=6) [21,26,30,35,40,45], and cross-sectional studies (n=3) [19,25,42]. Three studies (3/29, 10.3%) [20,23,34] scored below 40% of CASP items due to insufficient methodological descriptions and unclear recruitment or analysis procedures. Fifteen studies (15/29, 51.7%) [18,21,22,24,25,27,28,33,37,41-46] met between 50% and 70% of the criteria, whereas 11 studies (11/29, 37.9%) [19,26,29-32,35,36,38-40] exceeded 80%. In diagnostic and prediction studies, common limitations included incomplete reporting of recruitment processes, lack of external validation, and absence of blinding. Despite these weaknesses, most studies demonstrated clear research aims and appropriate methodological choices. Detailed CASP scores for each study are presented in Multimedia Appendix 3.

RQ1 Findings

Overall Maturity Across the 5 Layers

To validate the proposed 5-layer architecture, all 29 studies were systematically mapped to the framework based on the maturity levels (Table 7). The application layer (mean 3.17, SD 0.85) and data layer (mean 3.00, SD 0.76) demonstrated the highest maturity, followed by the algorithm layer (mean 2.79, SD 0.77) and infrastructure (mean 2.79, SD 1.70) layers, the latter showing considerable variability across hospitals. Security and compliance layer (mean 1.69, SD 1.89) remained the least mature and most inconsistently addressed across studies. These findings suggest that research has higher maturity on data readiness, model development, and workflow integration, whereas infrastructure and governance considerations showed both lower maturity and greater variability, suggesting that technical capacity, institutional governance, and compliance mechanisms remain unevenly developed and inconsistently reported in primary studies. Detailed evidence for each mapping decision is provided in Multimedia Appendix 4.

Evidence Stratified by Study Design

Given the methodological heterogeneity of the included studies, we conducted a stratified synthesis by study design (Table 8). The results of this analysis show that studies of clinical prediction, diagnostic test, and cohort studies achieved higher maturity in the data, algorithm, and application layers, particularly the clinical prediction studies, which showed the most consistent and advanced technical implementation. In contrast, qualitative and cross-sectional studies exhibited greater maturity variation, contributing more substantially to the infrastructure and security and compliance layers. The single economic evaluation demonstrated moderate maturity, limited mainly to the technical layers. This stratified synthesis highlights how methodological design shapes the visibility of different layers, with quantitative evaluation studies emphasizing technical robustness and data integration, whereas qualitative designs better capture infrastructural and governance maturity essential for sustainable AI platform development.

Mapping the Evidence to the 5-Layer Framework

We mapped the identified evidence to the 5-layer framework: (1) infrastructure layer, (2) data layer, (3) algorithm layer, (4) application layer, and (5) security and compliance layer. To achieve this, we reviewed the specific evidence in Multimedia Appendix 4 and selected high-frequency examples to map into the architecture. Figure 3A shows the detailed evidence mapping, where each layer is populated by distinct categories of evidence, and Figure 3B presents the simplified conceptual pyramid as an overview.

‎

RQ2 Findings

While RQ1 examined the maturity of individual layers, RQ2 explored the interrelationships among layers in hospital AI systems. The weighted co-occurrence heatmap (Figure 4) clearly shows that data, algorithm, and application exhibited the strongest interconnections in the center. The weighted Jaccard similarity heatmap (Figure 5) further confirmed this pattern, showing strong maturity overlap among the 3 layers (data-application=0.85, data-algorithm=0.89, and algorithm-application=0.80). This “core triad” indicates that data preparation, model development, and workflow integration are inseparable steps in most implementations.

‎

‎

Cells show weighted co-occurrence scores (sum of per-study maturity shared by each pair of layers, 0‐5 scale), so larger values indicate stronger cross-layer coupling. Diagonal cells give the cumulative maturity for each layer. The central triad, data, algorithm, and application, shows the strongest coupling (eg, data-application=82; data-algorithm=79; and algorithm-application=77), whereas security and compliance is consistently weaker with other layers. Warmer colors denote higher weighted coupling.

In contrast, security and compliance appeared peripheral, with lower scores when coupled with the other layers (Jaccard=0.43‐0.46). Infrastructure demonstrated moderate connectivity with application (Jaccard=0.77) and data (Jaccard=0.73), implying that infrastructural maturity often co-develops with technical capability but is not systematically aligned with governance or oversight mechanisms. These patterns point to the need for earlier integration of governance and compliance into platform design, ensuring they function as core rather than peripheral components.

Implementation Examples

To assess if the proposed 5-layer hospital AI platform (infrastructure, data, algorithm, application, and security and compliance) can work in practice, we narratively synthesize 4 fielded deployments from our 29 included studies, in settings with explicit clinical integration and measurable end points.

Example 1

An AI application for noncontrast head computed tomography (CT) [46] was deployed in an emergency radiology setting to flag multiple intracranial hemorrhage subtypes and reprioritize critical cases in the radiology worklist. The deployed system, using PACS infrastructure (layer 1: infrastructure), which focused on curated CT datasets, engaged difficult cases, such as postoperative changes and artifacts (layer 2: data). A multisubtype detection algorithm (layer 3: algorithm) applied with balanced sensitivity and false positives allowing for better triage. This led to automated elevation of priority levels and reordering of queues clinically, which reduced turnaround times for report generation and emergency department length of stay (layer 4 application). To reduce the chances of false alarms being raised through human verification and not to miss out on any subtype, standard operating procedures were framed at this layer (layer 5: security and compliance).

Example 2

An AI-generated first-reader system was integrated into current workflows of chest radiographs to flag high-confidence tumor cases for earlier review [44]. The deployment was able to link with the digital radiography acquisition, PACS, Radiology Information System, reporting system for batch inference, and rule-based triggers (layer 1: infrastructure). Through the use of large-scale chest X-ray datasets in Digital Imaging and Communications in Medicine format with structured reads to calibrate thresholds and monitor downstream confirmations (layer 2: data). The software gave triage scores for nodules, masses, and hilar enlargement. These scores formed the high-confidence queue (layer 3: algorithm). At the application level, this facilitated reprioritization and consistency-oriented quality management across radiologists (layer 4: application), whereas false-positive audits and drift monitoring ensured compliance and performance stability over time (layer 5: security and compliance).

Example 3

A machine learning–enabled system at primary stroke centers aided in the early detection of large-vessel occlusion and accelerated transfer protocols [39]. Through CT and computed tomography angiography (CTA) acquisition, PACS, alerting on-call, and interhospital coordination platforms, the system was tightly coupled for smooth integration of infrastructure (layer 1: infrastructure). It combined imaging with time-stamped process data (arrival, transfer, reperfusion) to monitor pathway performance (layer 2: data). The algorithm automatically stratified large vessel occlusion cases, triggering alerts and activating predefined stroke pathways (layer 3: algorithm). At the application level, this enabled the rapid activation of “green channel,” transport prioritization, and standardization of decision points in acute stroke care (layer 4: application). As per the established transfer policies and accountability along the care chain, cross-site data sharing took place in a manner that governed the deployment according to the governance and compliance expectations (layer 5: security and compliance).

Example 4

A platform-scale AI deployment was rolled out across 90 tertiary hospitals to provide image analysis and clinical decision support at scale [25]. The infrastructure consisted of multisite compute and networking resources with containerized inference engines with full-stack EHR and PACS integration for seamless updates (layer 1: infrastructure). Data was curated across institutions, combining different types and standardization procedures as part of (layer 2: data) for use in other contexts. A portfolio of task-specific models was built as a platform asset with continuous iteration (layer 3: algorithm), providing triage, detection, recommendation, quality control, and more capabilities. The platform created common clinical entry points for decision support (layer 4: application) to streamline workflows across sites. Security and compliance were formalized through access control, audit trails, change management, and staff training, underscoring its readiness for large-scale operations (layer 5: security and compliance).

Across these examples, 3 cross-cutting themes emerged. First, workflow-native integration, such as worklist reprioritization, automated alerts, and transfer coordination, was the critical pathway for translating algorithmic outputs into measurable clinical benefits. Second, infrastructure and data robustness are critical to sustaining application-level improvements. Deployments without standard pipelines are fragile. Third, governance was put into action through standard operating procedures, threshold calibration, and audit mechanisms to control false positives, drift, and intersite variation. The 5-layer model is practically viable and also highlights where further investment is needed, including in infrastructure resilience, data governance, and clinician adoption.

Discussion

Principal Findings

A 5-layer hospital AI platform model (infrastructure, data, algorithm, application, and security and compliance) was developed by synthesizing 4 reference frameworks. From 283 records screened, 29 studies (29/283, 10.2%) were included with high interrater reliability (κ=0.98). Most studies were diagnostic test (8/29, 27.6%) or qualitative (6/29, 20.7%), and almost half were conducted in the United States (12/29, 41.4%). Quality varied: only 37.9% (11/29) of studies achieved more than 80% of CASP items, whereas 10.3% (3/29) scored below 40%.

Evidence mapping showed the application (mean 3.17, SD 0.85), data (mean 3.00, SD 0.76), and algorithm (mean 2.79, SD 0.77) layers have the highest and most balanced maturity, forming a tightly integrated core triad. Weighted co-occurrence analysis demonstrated the strongest interconnections among these 3 layers (data-application=82, data-algorithm=79, and algorithm-application=77), and weighted Jaccard similarity indices confirmed substantial maturity overlap (data-algorithm=0.89, data-application=0.85, and algorithm-application=0.80). The infrastructure layer (mean 2.79, SD 1.70) displayed moderate maturity but high variability. The security and compliance layer (mean 1.69, SD 1.89) remained the least mature and weakly connected to the others (Jaccard=0.43‐0.46). Stratification by study design showed that technical studies (diagnostic, prediction, and cohort) achieved higher maturity within the technical core, whereas qualitative and cross-sectional studies more often addressed infrastructure and governance that remain underdeveloped in most technical evaluations.

Comparison With Prior Work

There was a maturity imbalance across the layers, which were dominated by technical layers. There were similar trends observed in earlier reviews, which showed that algorithm performance and workflow integration achieved higher maturity, whereas infrastructure, ethics, and compliance were less developed [47,48]. Several factors may explain this. The focus of studies and publication bias has favored predictive accuracy and model validation [47,49]. Most studies were led by clinical and technical teams. The governance and IT planning teams were featured less due to disciplinary divides [50]. Many studies described pilot projects, where compliance and infrastructure emerged later in scaling.

Domain-specific differences were noted. In emergency imaging [39,46], infrastructure and application layers were highly mature due to the urgency of real time, whereas compliance was limited to operating procedures. In routine imaging [22,44], data-algorithm dominated, with application relying on threshold tuning. In chronic disease management [38,41], data governance and compliance were critical for privacy-preserving integration, and application focused on longitudinal risk stratification. In deployments across multiple hospitals [25], infrastructure and data pipelines were key bottlenecks, and compliance was institutionalized through audit trails and training. Negative outcomes were also reported: chest radiograph triage increased false positives in patients with comorbidities [22], ICH detection struggled with specific subtypes [46], and multihospital systems faced high costs and low adoption [25]. These findings confirm that success depends on context, workflow fit, and governance.

Implications for Practice and Policy

Administrators can use the 5-layer model for phased investments, particularly in the areas of infrastructure and compliance, to eradicate blockages at a later stage [51,52]. Policymakers can promote compliance-by-design standards, requiring privacy, accountability, and explainability from the outset [53,54]. Vendors can align products with hospital priorities by embedding interoperability and monitoring tools [52]. Funding organizations and health systems should support the underdeveloped layers, especially governance and infrastructure through training and cross-institutional collaboration [51].

Challenges and Limitations of AI Deployment in Hospitals

Technical and infrastructural barriers were widely reported. Data were often fragmented across EHR, LIS, and PACS, and many systems lacked application programming interface support or sufficient computing capacity [55,56]. These weaknesses limited scalability, whereas bandwidth bottlenecks reduced real-time performance and inadequate monitoring allowed model drift [22,25,46]. Solutions that have been proposed include HL7 FHIR–based data lakes [57], hybrid cloud-edge architectures [58], and continuous monitoring with automated retraining pipelines [59,60].

Organizational and adoption barriers also affected implementation. Clinician skepticism, workflow misalignment [22,46], and lack of structured feedback were common problems [30,35,61,62]. Better outcomes were described when interdisciplinary teams were formed, clinical champions were engaged early, and projects targeted low-risk and high-value use cases [63,64].

Regulatory and compliance gaps were identified as the weakest layer. Privacy safeguards, governance frameworks, and liability protocols were rarely embedded into early projects [21]. Compliance-by-design approaches have been recommended, supported by federated learning [65], explainability mechanisms, and blockchain-based audit trails [66], together with early engagement of regulators [67].

Economic and resource barriers were another critical concern. High upfront and maintenance costs, uncertain return on investment, and misalignment with hospital budget cycles were repeatedly described [25,68]. Strategies such as phased investments, AI-as-a-service models, and shared consortia were suggested to reduce costs and support sustainable deployment [69].

Limitations of This Review and Future Research

This review has several limitations. A limitation is that only peer-reviewed English-language publications were included, which may have excluded valuable implementation reports published in other languages or as gray literature. This restriction was applied to maintain consistency and reliability in the 0 to 5 ordinal maturity scoring, as translation variability could compromise coding accuracy and interrater agreement. In addition, most gray literature lacks formal peer review or standardized reporting, which may introduce methodological inconsistency and compromise the overall reliability of evidence synthesis. Therefore, it was intentionally excluded to preserve data quality and comparability. Another limitation is that studies were mapped to the 5-layer framework using an ordinal maturity scoring, which may not fully reflect the complexity of AI applications. A further limitation is that the framework was validated only through literature synthesis and has not been prospectively tested in real hospital environments, such as in resource-limited settings or smaller hospitals.

Future research should extend evidence retrieval to include non-English and gray literature through calibrated multilingual screening and curated institutional sources. Further refinement of maturity metrics and cross-layer evaluation methods is warranted to better represent the dynamic evolution of hospital AI systems. Furthermore, prospective multicenter studies are also required to validate the framework in practice and to test its scalability across diverse hospital settings.

Conclusions

The integration of hospital information systems with AI is essential for the digital transformation of health care. Evidence from 29 empirical studies was synthesized with established frameworks to validate a 5-layer architecture. This model provides both theoretical and practical guidance for platform-level AI in hospitals. The framework can be applied by researchers, practitioners, and policymakers to support the development of scalable, secure, and clinically integrated AI platforms.