Learning Health System Frameworks and AI in US Care: A Scoping Review
November 16, 2026


Learning health system frameworks describe how routine care can generate evidence that flows back into practice. A scoping review published as an Article in Press in npj Digital Medicine asks how those frameworks have been built and implemented in the United States, and where artificial intelligence (AI) and data science fit in. The review finds that AI work has concentrated on turning data into knowledge, while the organizational and feedback steps that carry that knowledge into practice have received much less attention.
The paper’s first authors, Rotana M. Radwan and Hao Dai, contributed equally. Jingchuan Guo and Jiang Bian jointly supervised the work. The team is based at Purdue University, Indiana University, the Regenstrief Institute and the University of Florida.
What the review covered
The authors followed the Arksey and O’Malley scoping review framework and the PRISMA extension for scoping reviews. The framework search began with 911 publications. After 378 duplicates were removed, 171 full texts were read and 54 framework studies were included: 24 (44%) proposed adapted conceptual frameworks and 30 (56%) reported real-world implementation. The framework studies were published between January 2007 and August 2025.
The AI component identified 835 publications. After 395 duplicates were removed, 147 full texts were read and 31 empirical AI or data science studies were included. Two publications qualified for both components, so the review covers 83 unique publications. Every study was U.S.-based and published in English.
Five components recur across learning health system frameworks
Five components recurred across the 54 framework studies. Data infrastructure and interoperability appeared in 37% of them, governance and stakeholder engagement in 35%, learning and improvement cycles in 33%, evidence generation and translation in 28%, and equity and ethics in 22%.
The AI studies matched the frameworks on the computational components but fell behind on the organizational ones. The table compares coverage in the two literatures.
| Core LHS component | Framework studies (n = 54) | AI studies (n = 31) |
|---|---|---|
| Data infrastructure and interoperability | 20 (37%) | 13 (42%) |
| Governance and stakeholder engagement | 19 (35%) | 5 (16%) |
| Learning and improvement cycles | 18 (33%) | 11 (35%) |
| Evidence generation and translation | 15 (28%) | 10 (32%) |
| Equity and ethics | 12 (22%) | 4 (13%) |
| None of the five components addressed | Not applicable | 12 (39%) |
Source: Radwan RM, Dai H, et al., npj Digital Medicine (Article in Press), Table 2, https://doi.org/10.1038/s41746-026-03452-1. Components were not mutually exclusive, so percentages do not sum to 100%.
Most AI work stops at knowledge generation
Of the 31 AI studies, 28 operate in the data-to-knowledge phase of the learning cycle. Fourteen build predictive or prognostic models, eight build infrastructure or platforms, four extract information from clinical text, and two train federated models without requiring centralized data sharing.
Three studies reach knowledge-to-practice. They are an evaluation of AI-generated suggestions for clinical decision support logic, a prospective deployment of a sepsis detection platform in routine emergency department care, and a case series on a deterioration-risk analytic in routine intensive and acute care. No study describes a practice-to-data mechanism in which outcomes observed after deployment were used to update the model.
Validation follows the same pattern. About 94% of studies relied on retrospective internal validation, two reported external validation, and one reported prospective deployment in a clinical workflow. For a prospective example, see the randomized phase II trial of AI preoperative communication in men with prostate cancer, which Syenza News has covered. Designs like it were rare in the review.
Data sources follow a similar pattern. Most AI studies relied on structured EHR data, and only 7 of 31 used public datasets. About 55% (17 of 31) also drew on unstructured data such as clinical notes, usually processed with natural language processing, and 16% used imaging. One example of that approach is the EpiScreen epilepsy detection model, which reads the notes neurologists write after referral.
The organizational gap
AI studies addressed governance and stakeholder engagement in 16% of cases, compared with 35% of framework studies. Equity appeared in 13% of AI studies and 22% of frameworks. The authors describe the difference as one of emphasis. Framework studies more often specify the leadership, participation and oversight that learning requires, while AI studies more often specify computational mechanisms. Twelve of the 31 AI studies (39%) addressed none of the five components and presented as retrospective, internally validated model development.
On the governance side, a clinical AI governance framework from Southern California Permanente Medical Group scores AI tools across people, process and technology at each stage, from vendor evaluation through maintenance. A separate scoping review of AI governance frameworks found that most lack the comprehensive structure needed for safe, real-world deployment in hospitals and acute care.
What the authors conclude
The authors write that advancing AI-enabled learning health systems will depend on more than increasing analytic capability. They suggest that greater attention may be needed to connect knowledge generation with clinical action and prospective evaluation, and to strengthen feedback, governance, stakeholder engagement, equity, and the integration of clinical and multimodal data. In their words, strengthening these connections may help move learning health systems “from systems that generate knowledge toward systems that can consistently apply and learn from it.”
The governance question also arises in low-resource settings. A Perspective on responsible AI regional networks proposes local governance, data sovereignty and community-driven infrastructure as its three pillars.
Source: Radwan RM, Dai H, et al., “Learning health system frameworks and the integration of artificial intelligence and data science in United States health care: a scoping review,” npj Digital Medicine (Article in Press), https://doi.org/10.1038/s41746-026-03452-1
Let Google know we are your trusted source.
Add our editorial as a preferred source in your search results.
Value Science Weekly
The week in health technology assessment, market access and health economics — what moved, and what it means.





