Limitations
AI4M gives every LGA a monthly estimate that helps compare areas and see where malaria risk is likely to be highest. Four limits apply. The figures behind them are in the model card.
1. Outbreak prediction is delivered in part, through risk
Section titled “1. Outbreak prediction is delivered in part, through risk”AI4M tells you where and when malaria risk is highest, for every LGA and month, and when in the year cases are expected to peak. Risk indicates where outbreaks are more likely: in the three states with monthly data, the state with the highest risk had by far the most outbreak months. It shows where to prepare, not when an outbreak will begin. In the three states with public monthly case data, it also forecasts case numbers and the chance of an outbreak month. Giving early warning of a new outbreak across the country needs monthly facility data by LGA, which AI4M does not yet have.
Until then, use a score to decide where to look and prepare, and routine surveillance to confirm what is happening.
2. Accuracy is modest
Section titled “2. Accuracy is modest”The model explains about a quarter of the variation in malaria between places. It is good for ranking areas and weak as a precise figure for any one of them. The likely range is wide for that reason: plus or minus 0.13 for a state and 0.24 for an LGA.
3. It works from survey results that are five years old
Section titled “3. It works from survey results that are five years old”The latest national survey with malaria testing was in 2021, and none is currently scheduled. Month-to-month movement in a score comes from rainfall and follows the real season only loosely, so do not read it as measured change in malaria. For when cases are expected to be highest, use the transmission season. A forecast assumes a typical season and will not anticipate floods or drought.
4. Local detail is limited
Section titled “4. Local detail is limited”The surveys were designed to describe states, not LGAs, and 82 of the 774 LGAs had no usable survey location. The model also has no information on local conditions such as net campaigns, drug supply, conflict or displacement, so two areas with the same score can be in very different situations.
What comes next
Section titled “What comes next”All four limits have the same cause: AI4M does not yet have monthly data from health facilities. With it, estimates would rest on recent measurements in every LGA instead of a five-year-old survey, and early outbreak warning becomes possible. AI4M is built to take that data without being rebuilt.
What this means in practice
Section titled “What this means in practice”| Do | Avoid |
|---|---|
| Compare areas and look for consistently high-risk places | Treating small differences between scores as real |
| Use state figures for planning, and LGA figures to guide a closer look | Using an LGA score on its own to decide where resources go |
| Show the likely range beside a score | Reporting a score as a precise fact |
| Combine AI4M with surveillance data and local knowledge | Using AI4M to declare or rule out an outbreak |