How Ethiopia is using AI to improve rural health facility placement

Governments in low- and middle-income countries face a persistent dilemma when expanding healthcare access: demand far exceeds available resources, while decisions about where to invest are shaped by both data and deeply contextual human judgment. In rural and hard-to-reach regions, the consequences of poor facility placement are immediate and severe, translating into long travel times, delayed care, and preventable health outcomes. However, the tools traditionally used to guide these decisions often fall short, either optimizing for mathematical efficiency alone or relying heavily on expert opinion that is difficult to standardize, audit, or scale.

A new study titled “Health Facility Location in Ethiopia: Leveraging LLMs to Integrate Expert Knowledge into Algorithmic Planning,” presented at the 25th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2026), introduces a hybrid planning framework that combines classical optimization techniques with large language models to support health facility placement decisions that are both data-driven and aligned with expert guidance. Developed in collaboration with Ethiopian public health institutions, the study positions artificial intelligence not as a replacement for human decision-making, but as a structured bridge between quantitative rigor and qualitative judgment.

Why pure optimization has fallen short in health planning

Health facility location planning is, at its core, an optimization problem. With a fixed budget, planners aim to maximize population coverage while respecting constraints such as distance, capacity, and accessibility. Over the past decade, advances in geospatial data and optimization algorithms have enabled increasingly precise estimates of where new facilities would reach the largest number of people. These approaches provide clear, defensible answers in mathematical terms and offer theoretical guarantees about performance.

In practice, however, such solutions rarely determine final policy decisions on their own. Public health officials, regional planners, and community stakeholders bring additional considerations that are difficult to encode numerically. These include uneven terrain, seasonal accessibility, regional equity concerns, political feasibility, and historical underinvestment in certain districts. Experts may prioritize strengthening services in fragile areas even when population density is lower, or insist on minimum service levels in specific regions for strategic or ethical reasons.

The study notes that this disconnect has limited the real-world impact of optimization-based planning tools. When algorithmic outputs fail to reflect expert priorities, they are often sidelined during negotiations, leaving decisions to be made through informal compromise rather than transparent analysis. At the same time, relying exclusively on expert judgment can obscure trade-offs, reduce accountability, and make it harder to assess whether limited budgets are being used effectively.

The Ethiopian context illustrates this challenge clearly. Since the launch of the Health Extension Program in the early 2000s, the country has expanded basic health posts across rural areas. A new national roadmap envisions upgrading selected posts into comprehensive facilities capable of delivering more advanced services such as childbirth and postnatal care. These upgrades are costly, and only a small fraction of facilities can be improved each year. Choosing which posts to prioritize has become a high-stakes decision involving competing objectives and viewpoints.

A hybrid framework that lets AI interpret human expertise

To address this gap, the researchers propose a hybrid framework known as the LLM and Extended Greedy, or LEG, approach. The system is designed to integrate expert knowledge expressed in natural language directly into an optimization process without sacrificing mathematical guarantees on coverage.

The framework begins with a classical greedy optimization algorithm that selects facility locations to maximize population coverage under a fixed budget. This step provides a strong baseline, ensuring that any proposed allocation performs well in terms of reaching people within acceptable travel times. Importantly, this baseline comes with well-established theoretical guarantees, offering planners confidence in its efficiency.

The innovation comes in the next stage. Instead of treating expert advice as an external constraint or an informal adjustment, the LEG framework uses a large language model to interpret qualitative guidance provided by stakeholders. This guidance can include statements such as prioritizing certain districts, maintaining minimum facility counts in underserved areas, or balancing investments across regions. The LLM translates this unstructured input into structured adjustments at a higher, district-level allocation.

A guided greedy procedure then converts these district-level preferences back into specific facility locations, carefully controlling how much the solution is allowed to deviate from the original coverage-maximizing allocation. Two tunable parameters govern this trade-off. One determines the proportion of decisions driven strictly by the greedy algorithm, while the other sets how much loss in marginal coverage is acceptable when following expert advice. Together, they allow policymakers to explicitly balance efficiency against alignment with human judgment.

Crucially, the framework preserves formal performance guarantees. Even when the LLM influences part of the decision-making process, the final allocation is guaranteed to achieve a defined fraction of the optimal population coverage. This addresses a major concern in high-stakes domains such as public health, where opaque or unconstrained AI systems may produce outputs that appear reasonable but lack reliability or fairness.

Real-world evidence from Ethiopian health planning

The researchers evaluated the LEG framework using real population and accessibility data from three Ethiopian regions: Afar, Somali, and Benishangul-Gumuz. These regions were chosen because they differ significantly in terrain, population density, and infrastructure, creating a realistic test for the system’s flexibility and robustness.

To simulate real planning conditions, the study incorporated multiple sets of expert advice, including deliberately conflicting recommendations. This reflects the reality of public-sector decision-making, where stakeholders often disagree on priorities and trade-offs. The LLM was tasked with iteratively refining allocations based on both quantitative feedback from the optimization process and qualitative alignment with expert guidance.

The results show that incorporating language-based feedback substantially improves alignment with expert preferences compared to purely numerical approaches. Allocations produced with LLM-guided refinement more closely reflected stakeholder priorities, such as equitable regional distribution or targeted support for specific districts. At the same time, population coverage remained high and consistently satisfied the framework’s theoretical guarantees.

As planners adjust the balance parameters, they can observe predictable changes in outcomes. Lower values favor qualitative conformity, allowing expert advice to shape decisions more strongly. Higher values recover behavior closer to classical optimization, prioritizing coverage above all else. This interpretability is critical for policy adoption, as it allows decision-makers to justify choices and explore alternative scenarios openly.

The framework also supports sequential, multi-year planning. Public health investments are rarely made in a single round; budgets arrive incrementally, and earlier decisions constrain future options. The LEG approach can incorporate previous allocations as minimum constraints, ensuring continuity across planning cycles while still allowing new expert input and optimization at each stage. This feature makes it particularly well-suited for national programs that roll out infrastructure upgrades over many years.