AI follows $332 billion in health aid and exposes a chronic-disease mismatch

The striking result is a mismatch between two proportions. Non-communicable diseases account for about three-fifths of the global burden of disease—measured in lost healthy years, not case numbers—but received only 2.5% of the disease-specific development assistance identified in a new research analysis. Machine learning made it possible to compare millions of project descriptions with the health problems those projects were intended to address.[2][1]

The study, co-authored by ETH Zurich’s Kenneth Harttgen, appeared in Nature Communications on 18 August and was highlighted by ETH on 30 September. It concerns aid records from 2000 to 2022. Its roughly $332 billion total is the disease-specific funding identified within that historical dataset. It is not a measure of all global healthcare expenditure, nor a tally of funding available in 2026.[1][2]

What the AI actually counted

The researchers used a Llama 3.1-based classification pipeline to read project descriptions and assign disease categories. They checked performance against manually labelled records rather than assuming that readable model output was correct. The denominator matters: the millions of records were development projects, not millions of separate health programmes.[1]

ETH’s summary highlights the imbalance for chronic conditions such as diabetes and asthma. The finding is useful because aid statistics can show large totals while obscuring what those totals address. Categorising the descriptions makes a different question possible: how closely do the priorities visible in donor records correspond to the diseases causing harm?[2]

A funding map is not a complete account of care

The OECD’s reporting framework is organised around development-finance flows, sectors, purposes and policy markers. Its guidance also explains why contributions to multilateral institutions cannot be mechanically equated with spending in a particular country in the same year: timing, lending arrangements, earmarks and data coverage complicate the relationship.[3]

That accounting perspective is important here. An aid database describes reported financial activity. It is not a hospital ledger of every service delivered, and it cannot substitute for a national account of domestic government budgets, private payments and household spending. A country may finance a disease largely at home while relying on aid for another purpose.[3]

The paper explicitly treats disease burden as a reference point, not a formula that dictates the correct allocation. Its authors cannot track subsequent reallocations of the reported disbursements. Where a project covers more than one disease category, they split its funding equally between the applicable categories to avoid double-counting—an analytical convention, not an observed itemised budget.[1]

Consider a hypothetical project described as supporting both diabetes and kidney care. A model can identify the two stated aims, but the description alone may not reveal whether spending was divided evenly or concentrated heavily on one service. Better reading does not create missing financial detail. The estimate becomes more useful when that boundary remains visible.

There is also a policy distinction between unequal shares and demonstrated waste. A funding decision can reflect treatment costs, available domestic resources, spillovers or the feasibility of delivering an intervention. The useful response to the mismatch is to investigate those reasons, not to infer that moving every dollar until the proportions match would maximise health.

The AI contribution is therefore administrative and analytical rather than clinical. It makes a huge body of written aid records more searchable by purpose, revealing questions that broad sector totals can hide. The next decision still belongs to funders and recipient institutions: explain the imbalance, test whether it is justified and identify where an additional investment would do the most good.

Report an error or suggest a correction