Sustainable AI in Global Health Depends on More Than Accuracy
Artificial intelligence is moving quickly into global health, but technical performance alone does not make a tool usable in low- and middle-income countries. In settings where electricity, connectivity, staffing, and data quality can all be uneven, an algorithm that performs well in a controlled setting may still fail at the bedside, in the clinic registry, or across a district health system. The central question is no longer whether AI can be built. It is whether it can be deployed in a way that is durable, locally governed, and worth the resources it consumes.
This framework reframes AI in global health as a systems problem rather than a software problem. The emphasis falls on implementation, sustainability, equity, and environmental cost at the same time, because those pressures collide in real-world health systems. In low-resource settings, even modest computational demands can translate into higher energy use, additional hardware needs, and new maintenance burdens. If a tool depends on infrastructure that does not exist, or on repeated external support to keep running, it may expand the very gaps it is meant to close.
A one-sentence truth sits underneath the entire framework: a model that cannot be maintained locally is not sustainable care.
Green AI, Local Stewardship, and the Limits of Short Pilots
The framework proposes four linked components, beginning with Green AI necessity assessments. Before deployment, teams should ask whether the AI system is needed at all, whether a simpler digital tool would do the same job, and whether the expected benefit justifies the energy, hardware, and operational footprint. In low- and middle-income countries, that question carries clinical weight. Power constraints, limited device replacement capacity, and competing health priorities mean that energy-intensive systems can become fragile or inequitable the moment they leave the demonstration phase.
A second pillar is a One Digital Health systems lens. This approach treats human health, animal health, environmental conditions, and digital infrastructure as connected parts of the same operating environment. In practice, that means AI cannot be evaluated only by its model metrics. The surrounding workflow, data pathways, referral patterns, governance structure, and infrastructure readiness all shape whether the tool improves care or simply adds another layer of complexity. A maternal health triage model, for example, may be clinically sound on paper and still create delays if it is built around data fields that are inconsistently captured in routine practice.
The framework also calls for pragmatic, workflow-integrated evaluation. That means testing AI in the actual clinical process it is meant to support, rather than in a narrow technical sandbox. When evaluation is tied to routine care, implementation barriers become visible early: missing data, user burden, alert fatigue, documentation friction, and local mistrust. Those are not peripheral issues. They determine whether clinicians use the system at all.
Workflow Integration Determines Whether AI Helps or Hinders Care
The way an AI tool fits into a clinic can matter as much as the model itself. If the system slows documentation, increases clicking, or generates recommendations that are difficult to interpret, clinicians will either work around it or abandon it. In resource-constrained settings, even minor inefficiencies can have outsized effects because staff shortages leave little room for extra tasks. The clinical consequence is straightforward: a tool that adds work without improving decisions risks worsening access rather than strengthening it.
The framework’s emphasis on locally relevant evidence addresses another persistent problem in global health technology. Many AI systems are trained, validated, or tuned in environments that differ sharply from the settings where they are later deployed. Differences in language, disease patterns, demographic structure, infrastructure, and care pathways can erode performance. Locally generated evidence is therefore not a luxury. It is the only reliable way to know whether a tool behaves safely and fairly in the population it is supposed to serve.
Equity also sits at the center of this model. If training data underrepresent rural communities, marginalized groups, or populations with limited digital access, the resulting system can amplify existing bias. AI can deepen that divide when its output becomes embedded in decision-making without adequate local oversight. In global health, fairness cannot be inferred from intent. It has to be demonstrated through design, testing, and governance.
Federated Governance and Durable Infrastructure Shape Real-World Impact
The final pillar is federated governance, a structure that supports locally led stewardship while still allowing cross-institutional learning. This is a practical response to a familiar imbalance: tools and standards are often imported, while authority over data use, adaptation, and monitoring remains elsewhere. Federated governance shifts control toward local actors who understand the care context, the available resources, and the consequences of failure.
Local stewardship matters because sustainability depends on ownership. Systems are more likely to endure when clinicians, health ministries, informaticians, and community stakeholders can adapt them without waiting on external teams for every adjustment. That includes decisions about maintenance, retraining, monitoring, and retirement. A model that cannot be updated locally becomes obsolete quickly, especially when disease patterns, workflows, or public health priorities change.
Environmental responsibility is not separate from clinical responsibility here. Energy consumption, compute demands, and hardware turnover all have downstream effects, especially when AI systems are scaled across large public health networks. If deployment is not planned with those realities in mind, the carbon and operational cost can become substantial. The framework argues for durable infrastructure and stronger health information systems because sustainability begins long before model launch and continues long after the first pilot ends.
Evidence Gaps Remain in How Sustainable AI Is Measured
The most important limitation is also the most obvious one: the field still needs rigorous evidence on how to measure success beyond short-term technical performance. Many projects can show feasibility, but far fewer can show durability, affordability, equity, and environmental restraint over time. A system that works for a six-month pilot may still fail when scaled to district or national use.
Decision-makers in global health will need to ask harder questions before approving new AI deployments. What infrastructure is required to keep the system running? Who owns the data and the updates? How much energy does the model consume? Which populations benefit, and which are left out? What workflow burden does the tool add to already stretched staff?
Those questions define the next phase of AI in LMICs. Funders and implementers who focus only on model novelty will keep producing pilot projects that look promising and then stall. Systems that are locally governed, workflow-compatible, environmentally aware, and built on durable health information infrastructure have a better chance of lasting long enough to matter.