AI Fits the Learning Health System Only Partway So Far
A scoping review in U.S. health care finds a widening gap between what learning health systems are designed to do and what most artificial intelligence projects actually accomplish. The review mapped frameworks for learning health systems and then compared them with studies that used AI or data science, asking a simple but consequential question: do these tools support the full learning cycle, or only the easiest slice of it?
They mostly support the easiest slice.
The analysis identified 54 framework studies and 30 AI or data science studies published between January 2007 and August 2025. Within the framework literature, 24 studies were conceptual adaptations and 30 described real-world implementations. The authors then organized the recurring components of those frameworks into five domains: data infrastructure and interoperability, governance and stakeholder engagement, learning and improvement cycles, evidence generation and translation, and equity and ethics. Those elements sketch the anatomy of a functioning learning system. They also expose why a model that performs well in a data environment can still fail in practice if it never connects to clinicians, patients, or the workflows that determine whether evidence changes care.
Data Rich, Organizationally Thin
The AI studies leaned heavily toward the data-to-knowledge phase of the learning cycle. Out of 30 AI or data science studies, 27 focused on generating knowledge from existing data. Only three reached the knowledge-to-practice phase. None described a practice-to-data feedback loop in which post-deployment outcomes informed model updating.
This is the central fault line. A predictive model can classify risk, flag patients, or estimate probability, but a learning health system requires more than a strong algorithmic output. It needs governance, trust, implementation pathways, and a feedback structure that captures what happens after the model enters routine care. Without that machinery, AI can become a static analytic product rather than a learning engine.
The review also found a mismatch in the kinds of components emphasized across the two bodies of literature. Compared with framework studies, AI studies less often addressed governance and stakeholder engagement, at 17% versus 35%, and equity, at 13% versus 22%. In practical terms, the AI literature was more comfortable with computation than with coordination. It was also less likely to confront who benefits from model deployment, who might be missed, and how bias or uneven access could shape downstream use.
Equity and Governance Sit Outside the Algorithm
That imbalance matters because the success of a learning health system is not determined by model accuracy alone. Clinical systems are social systems. Data are collected inside institutions with uneven resources, different documentation habits, variable patient access, and local rules that influence which signals are visible and which are buried. A model trained without attention to governance and equity may perform elegantly in a narrow setting while failing to generalize, amplifying disparities instead of reducing them.
The review’s framework synthesis makes this point indirectly but forcefully. Learning health system design repeatedly returned to interoperability, stakeholder engagement, evidence translation, and ethics, not just data pipelines. In contrast, AI studies concentrated on technical production of knowledge. The result is a partial overlap between two fields that ought to be mutually reinforcing. One field defines the system needed to turn routine care into evidence. The other often optimizes tools that stop before they can be absorbed into that system.
For hospitals and health systems, the bedside implication is straightforward. A model that never reaches practice does not alter care delivery, and a model that reaches practice without a feedback pathway cannot adapt when patient populations shift, documentation patterns change, or performance drifts over time. In a learning health system, deployment is not the end of analysis. It is the beginning of the next data cycle.
What the Scoping Review Revealed About the Field’s Direction
The study does not argue against AI. It places AI in its proper position inside a broader institutional architecture. The message is not that data science has failed; it is that the current literature has overinvested in the computational portion of learning while underbuilding the organizational scaffolding required to make that learning durable.
That distinction shows up clearly in the review’s emphasis on the five recurrent framework components. Data infrastructure and interoperability are necessary, but not sufficient. Governance and stakeholder engagement are necessary, but not decorative. Learning and improvement cycles are the mechanism by which evidence becomes operational. Evidence generation and translation determine whether findings leave the page and reach clinicians. Equity and ethics decide whether the resulting system serves the full patient population or only the subset easiest to measure.
The review’s time window, spanning nearly two decades, also suggests that this is not a new concern. The literature has had years to converge, yet AI studies still rarely describe the full learning arc from data to knowledge to practice and back to data. That leaves a fundamental design question in place for health systems investing in digital transformation: are they building instruments that analyze care, or systems that learn from it?
Why the Feedback Loop Still Rarely Exists
A fully functioning feedback loop demands more than technical integration. It requires post-deployment surveillance, ownership of model performance, clear responsibility for updates, and enough institutional maturity to act when a tool begins to underperform. It also requires clinical users to trust the outputs enough to change behavior, while retaining enough skepticism to challenge them when the real-world signal diverges from the training environment.
That is where many AI efforts stall. The model may be validated, even externally tested, but if it is not embedded in governance structures and clinical workflows, the learning process ends at publication or pilot deployment. Health systems then inherit a familiar problem: promising analytic work that never matures into continuous improvement.
The review offers a useful corrective for researchers, administrators, and informaticians. Future work in this space will need to report not only what a model predicts, but how it is governed, who is responsible for monitoring it, whether it changes care, and whether outcomes are fed back into the next iteration. Without those details, a project may be impressive as data science and still incomplete as a learning health system.
The literature is clear on the technical promise. It is less developed on the institutional mechanics that make the promise matter.
The Narrower Evidence Base Around Implementation
The authors limited their search to U.S.-based, English-language studies, which keeps the review tightly focused but also narrows the evidence base to one national context. The selection window began in 2007 and extended through August 2025, capturing the modern rise of health data science and machine learning in clinical settings. Two reviewers independently screened and extracted data, resolving disagreements by consensus, which strengthens the reliability of the mapping process.
No funding was received for the research. The absence of funding support does not weaken the core message; it simply leaves the findings anchored in a descriptive synthesis rather than a programmatic initiative tied to one institution or vendor. The more consequential limitation is structural. If the literature continues to emphasize algorithmic output over deployment, governance, and post-deployment learning, then the field will keep producing tools that are sophisticated in isolation and incomplete in care delivery.
What remains unresolved is not whether AI can analyze health data. It can. The unresolved issue is whether U.S. health systems are building the organizational conditions that allow those analyses to feed back into care, adapt after deployment, and function as part of a genuine learning health system rather than a one-time analytics exercise.