District Dengue Early Warning from Weather and Case Counts for Health Departments

The problem: dengue response starts after the outbreak has begun

Dengue returns to Pakistan's cities every year after the monsoon, and some years bring large outbreaks that fill hospitals. Health departments can do a lot: larval surveillance, spraying, clearing standing water, preparing beds and platelet supplies. Each of these takes weeks to organise. When the response starts only after case numbers rise, it is already late.

Mosquito breeding follows rain and temperature with predictable delays. Combining that with recent case counts makes it possible to forecast where cases are heading several weeks ahead. The challenge is building a model that is honest about uncertainty, tested on past seasons, and presented in a form a district health officer can use.

What already exists and the gap

• Punjab's Dengue Activity Tracking System, developed by the Punjab Information Technology Board, tracks larval surveillance, spraying and confirmed patients across districts. It records activity and cases; it does not forecast them.
• OpenDengue (Clarke and colleagues, Scientific Data, 2024) is a global open database of public dengue case counts, with weekly or monthly data and subnational records for many countries.
• Research models for dengue forecasting exist for Brazil, Vietnam and other countries, usually as academic code.

The gap is an operational, district-level forecast for South Asian settings that is backtested, explained and delivered as risk levels rather than raw model output.

Inspiration

Lowe and colleagues' paper "Dengue outlook for the World Cup in Brazil: an early warning model framework driven by real-time seasonal climate forecasts" (Lancet Infectious Diseases, 2014) issued probabilistic dengue risk forecasts for 553 Brazilian microregions ahead of the tournament, using climate forecasts and the latest case data in a hierarchical model. It is a model of how to communicate forecast risk. Your project adapts the core idea with a simpler model first.

What you build

• The data pipeline collects weekly district case counts from available public sources and weather from reanalysis data, aligned to the same weeks.
• The model forecasts cases 2 to 8 weeks ahead with intervals, and estimates the probability of crossing each district's outbreak threshold.
• The dashboard shows a map of risk levels, the forecast curve for each district and the main drivers, such as "heavy rain three weeks ago".
• A health officer uses the weekly view to decide where to send spraying teams first.

The technically interesting part

• Count data. Case counts are overdispersed: many quiet weeks and a few explosive ones. A negative binomial model handles this far better than ordinary regression.
• Distributed lags. Rain affects cases with a delay of several weeks, spread over time. Model lagged weather as a smooth function rather than picking one lag by hand.
• Thresholds. Define an outbreak per district from historical levels (for example, a moving percentile), so a forecast becomes a decision signal.
• Data gaps. Public case data in South Asia is often incomplete, delayed or aggregated. Detect reporting gaps, handle them explicitly, and be clear where data is too thin to forecast.
• Fair baselines. A forecast is only useful if it beats "same week last year" and "average of past seasons". Always show those baselines.

How to evaluate it

• Accuracy: mean absolute error and a proper scoring rule such as the continuous ranked probability score, at each forecast horizon, on held-out seasons.
• Warning quality: hit rate and false alarm rate for outbreak threshold crossings.
• Calibration: whether 80 percent intervals contain the true value about 80 percent of the time.
• Usefulness: show the dashboard to one public health professional and record what they would act on.

Suggested plan

• Weeks 1 to 3: gather case and weather data, explore seasonality, choose districts with the best data.
• Weeks 4 to 7: baselines, negative binomial model, lagged weather features.
• Weeks 8 to 10: forecasts with intervals, thresholds, backtesting.
• Weeks 11 to 13: dashboard and explanations.
• Weeks 14 to 15: expert feedback, report and demo. A second semester could add the hierarchical model and seasonal climate forecasts.

Path to a product

Provincial health departments, international health programmes and hospital networks all plan around dengue season. An open forecasting service with a hosted dashboard, run in partnership with a public health body, is the most realistic route. The project is also strong research material and leads into epidemiological data science roles.

Scope warning

Do not promise to predict exact case numbers; forecast ranges and risk levels. Do not use individual patient data. If district data is unavailable for Pakistan, demonstrate the method on a country with good OpenDengue coverage and show how it would transfer. Start with simple models and add complexity only if it beats them.

Learn what you need on Sefism

• Data science roadmap: regression, uncertainty and model comparison.
• Data science cheat sheet: evaluation and validation habits.
• OpenIntro Statistics: the statistics behind intervals and calibration.
• An Introduction to Statistical Learning: generalised linear models and validation.
• Our World in Data: context and public health data to support your write-up.

Get help with this idea

The usual failure is a model that fits past data and forecasts poorly. Tauseef can help you set up proper backtests, choose the districts and baselines, and present risk in a way health staff can act on. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session