Building Trustworthy Public Health Prediction Models

Public health decisions increasingly draw on predictive models: forecasts of seasonal surges, maps of geographic risk, rankings of communities for intervention. When these models are good, they help limited resources go further. When they are poor - or simply poorly understood - they can mislead with an air of scientific authority.

Trust in a prediction model is not something a vendor can claim. It has to be earned through practice. Here are six practices we believe matter most.

1. Start with the decision, not the model

A model is only useful if it informs a real decision. Before building anything, it is worth asking: Who will use this output? What will they do differently? How far ahead do they need to know? At what geographic resolution do they plan?

A beautifully accurate national forecast is of little use to a district team that needs ward-level guidance three weeks ahead.

2. Beat a simple baseline

Every forecasting model should be compared with simple alternatives: the same week last year, a seasonal average, or a naive persistence forecast. If a complex model cannot reliably beat these baselines on held-out data, it should not be used - however sophisticated it appears.

3. Validate the way it will be used

Validation should mirror operational use. That means evaluating forecasts made only with data that would have been available at the time, including reporting delays. It means testing across multiple seasons and geographies, including unusual years. And it means reporting performance for the places and periods that matter most - such as the run-up to seasonal peaks.

4. Make uncertainty impossible to miss

A single line on a chart invites over-confidence. Prediction intervals should be shown by default, and they should be calibrated: an 80% interval should contain the observed value roughly 80% of the time. Data-quality indicators - such as reporting completeness - should accompany forecasts so users can see when inputs are weak.

5. Explain drivers in plain language

Decision-makers are more likely to use - and appropriately challenge - a forecast when they understand what is driving it. “Risk is rising mainly because of above-average rainfall four to six weeks ago, and recent case momentum” is more useful than a score alone. Explanations should be honest about their limits: they describe the model’s behaviour, not necessarily causal mechanisms.

6. Keep people accountable and in the loop

Models should support decisions, not make them. That means clear ownership of each decision, the ability for experts to override model outputs with a recorded rationale, and regular review of how model outputs were used and with what results.

Beyond the model: governance

Trustworthy prediction also depends on governance around the model:

• Data agreements that respect ownership and sovereignty.
• Transparent reporting of methods, following established guidelines for prediction model studies where applicable.
• Monitoring for drift as reporting systems, populations and climate change.
• Equity review to check whether data gaps systematically disadvantage some communities.

Our commitment

At Epinova we treat these practices as requirements. Our models are in development, and we will publish validation results before any operational use. We would rather show a modest model honestly than an impressive one that decision-makers cannot trust.

If you are working on public health prediction - as a researcher, programme manager or funder - we would welcome the chance to compare notes.