Learning Health Systems

If the modes of medicine are multiplying, something must keep rewriting them, and that something is the learning health system.

The learning health system was conceived before AI made it practical, and what has changed since is the cycle time. Where the classic formulation imagined evidence accumulating over years, AI-era learning systems can in principle detect a signal, test it against population-scale records, and surface a revised recommendation within a clinical cycle.

The concept is not marginal. The learning health system has been endorsed by the National Academy of Medicine, the Agency for Healthcare Research and Quality, the Joint Commission, and the World Health Organization, and its constituent competencies have been described in detail for the better part of a decade.^10,11

Adoption nonetheless remains limited, and a scoping review of the empirical literature found considerable growth in applications alongside very few high-quality studies of implementation.¹¹ The obstacles that recur are consistent, and conspicuously non-technical: siloed data; absent or ill-adapted governance; misaligned incentives; underdeveloped workforce competencies; and cultural resistance to change.¹⁰

The AI era changes the calculus in a specific way. The challenges facing AI implementation and those facing learning health system development substantially overlap. Both require interoperable, high-quality, real-time data. Both require governance that can adapt faster than an annual cycle. Both require a workforce fluent in evidence and in informatics. The overlap means investment made for one purpose serves the other.¹⁰ The implication is hopeful. The promise of AI may finally supply health systems with a sufficient reason to build the data infrastructure that the learning health system has required, and failed to obtain, since it was first proposed.

Successful systems converge on one organizational form: the deliberate alignment of health information technology, the estate that must be reliable, secure, and cost-effective, with biomedical informatics, the discipline that must be exploratory and is often disruptive. Systems that have integrated the two describe an ambidextrous leadership model, optimizing current operations while driving novel solutions, a posture associated with better organizational performance in healthcare and in technology alike.¹⁰ The pandemic was the natural experiment. Systems that had built that collaboration produced capacity-demand forecasting, laboratory-based deterioration prediction, telehealth-supported remote monitoring, and vaccination delivery at scale in weeks.

A second insight from the same literature dissolves a distinction the academy has treated as fundamental. Research and quality improvement are two sides of one coin. Pre-clinical silent trials, which run a decision-support model in the background without influencing care, improve model performance and de-risk deployment. EHR-integrated A/B and rapid-cycle randomized testing convert quality initiatives into genuine experiments.

Multidisciplinary oversight groups that include review-board members can expedite that work without weakening consent, privacy, or ethics standards, and at least one large academic system reports that savings from its randomized quality initiatives more than covered the operating cost of the program running them.¹⁰ The traditional gap between clinical operations and rigorous evaluation is an artifact of how the two were organized.

None of this is a United States phenomenon, and the international picture adds a dimension the domestic literature underplays. A consensus process convened by the Future of Health with the Duke-Margolis Institute for Health Policy, drawing on senior leaders from more than fifty health systems worldwide, identified three action areas: advancing data-driven decision-making; building an organizational culture of learning; and engaging patients and caregivers.¹²

The case examples are useful because they are ordinary. Sheba Medical Center in Israel uses self-service synthetic data, statistically faithful to real records but carrying no identifying information, to let frontline teams interrogate clinical and operational questions without waiting on governance; one such inquiry changed the administration of an anaesthesia-reversal agent with no effect on outcomes and material annual savings. SingHealth embedded learning cycles into Singapore's national preventive-care initiative and then evaluated its own implementation ethnographically. And the learning health networks convened by Cincinnati Children's now span hundreds of pediatric organizations globally, with patients and families treated as co-producers of evidence and not as subjects of it.¹²

Set against Exhibit 3.1, the learning health system is not a twelfth mode either. It is the mechanism by which the matrix rewrites itself, the loop through which a mode of care is tested, revised, and either propagated or retired. The clusters describe the state of the system at a moment. The learning health system describes the rate at which that state changes, and the rate is now the more consequential quantity.

Two features of that international record inform the argument of this report. The first is that every example is sociotechnical rather than technical. What distinguishes these institutions is not their software but their willingness to restructure who may ask a question, and how quickly. The second is that patient co-production keeps appearing in the systems that work. A learning system whose loop is closed only among clinicians, data scientists, and administrators will learn efficiently about what those parties already consider worth measuring.

Both features, disciplined data infrastructure and genuine co-production with patients, point to the same role for the physician, one the literature rarely states. In a learning health system the physician is the party who formulates the question that is clinically consequential, and who then validates that the loop is answering it rather than something adjacent and easier to measure.

Neither half is automatable. The first requires having stood in the encounter where the uncertainty is felt; the second requires knowing enough about how the evidence was produced to recognize when a convenient proxy has quietly replaced the outcome anyone cared about. A learning system with excellent infrastructure and nobody performing that function will learn quickly, in a direction no one chose.

Modes of medicine will themselves undergo incessant updating and rethinking, which is why this chapter precedes the chapters on physicianship. Any vision of the future physician built around a fixed picture of care will be obsolete before the trainees it describes complete residency.

The recommendation therefore concerns organization rather than technology. A proactive and systematic approach must be adopted in the face of disruptive change looming across the modes of medicine, from care paradigms to technoscientific advances to machine intelligence to global health imperatives. The physician must be positioned as a key player across this emerging biomedical order, with an explicit role in governing what the learning system is permitted to learn.

That last clause carries more weight than its length suggests. A system that learns from care will optimize toward whatever it can measure, and it will do so faithfully. The unmeasured goods of medicine, the reassurance that prevented a return visit, the conversation that changed what a family wanted, the restraint that avoided an intervention, are then not merely absent from the loop. They are actively selected against.

Someone must therefore hold the authority to rule certain inferences out of bounds, to declare certain unmeasured goods protected, and to require that a proposed change to practice be justified against something other than the metric that proposed it. That is a normative function rather than a technical one, and it belongs with the party who answers for the resulting care. It gives the physician a specific and defensible office within the learning system in place of generic standing beside it.

One caution belongs beside the enthusiasm. Some elements of the learning health system are close to cost-neutral, among them the alignment of informatics with information technology, curricular change, and the refinement of review-board boundaries, and require only cultural priority. Others, particularly the data infrastructure that AI-era learning demands, are substantially cost-additive at a moment of contracting research support.

The most interesting proposition in the recent literature is that AI may be the first innovation capable of offsetting the cost of its own adoption, by semi-automating the administrative work of coding, billing, authorization, denial appeals, and regulatory reporting that consumes so much of what health systems spend.¹⁰ That proposition is unproven. It is also the pivot on which the financial viability of nearly everything described in this chapter rests.

The technologies driving that change are the subject of the next chapter.