TheUpsideDevs - DengueRadar

Inspiration

Singapore now protects about half its homes with Wolbachia mosquitoes, and 2026 is the quietest dengue year since 2018. That sounds like the problem is solved. It isn't. Quiet years have come before big outbreaks: about 5,300 cases in 2021 were followed by over 32,000 in 2022. Immunity is low right now, and 97% of today's dengue cluster cases sit outside the Wolbachia areas. Dengue hasn't gone away. It has moved, and a forecast that only looks at the weather can't see that. NEA already forecasts how many cases Singapore will have. Officers still need to know where.

What it does

DengueRadar is our proposal for the DAISI Challenge 2026 (Problem A2). The tool we plan to build will give each of Singapore's 55 planning areas a Low, Medium or High dengue risk for the next two weeks, using NEA's own traffic-light colours. It will include:

"Why?" cards that explain each risk level in plain words, such as nearby clusters, recent rain, heat or no Wolbachia coverage.
A new-cluster warning for areas that are clear today but likely to get a cluster soon.
A team planner that tells an officer where to send inspection teams so they reach the most at-risk homes without overlapping.
A weekly summary and resident notices in English, Chinese, Malay and Tamil, where code checks every number against our data before anything is sent.

How we built it

The dashboard is not built yet. What we have done is the data work that the idea depends on, so that every number in our proposal comes from real public data.

Done so far

The public data only shows today's dengue clusters, with no history. We pieced together 314 past copies of NEA's cluster list from 2015 to 2026 using three public archives. That gave us 17,750 clusters.
We read 184 weekly health bulletins to fill in case counts for 2023 to 2026.
We found the start date of every Wolbachia area from NEA announcements.
We boiled down about 1 GB per year of rainfall and temperature readings into daily figures per weather station.
We ran early forecast tests on this data, training on older years and predicting later ones.

Planned for the build sprint

A Databricks pipeline in three stages (raw, cleaned, ready), so every number can be traced back to its original file.
A forecast in three steps: how many cases Singapore-wide, which areas are at risk, and what to do about it.
The dashboard, alerts and summaries described above.

Challenges we ran into

The past teaches the wrong lesson. NEA put Wolbachia where dengue used to be worst, so in old data a Wolbachia area looks like a high-dengue area. Our first test model learned the opposite of the truth. The start dates fixed this, because the model now sees each area before and after protection began.
Messy sources. NEA's cluster page changed format over the years, street names had typos like "Tampines Stret 21", and the bulletins name their files in three different ways.
Too few thermometers. Only 18 stations measure temperature, covering 14 of the 55 areas.
A tough baseline. NEA keeps clusters open for 14 days, so "it will stay the same" is hard to beat.

Accomplishments that we're proud of

These are early results from our tests on public data, not from a finished product.

An 11-year cluster history that the public data doesn't offer. Where our three archives overlap, the case counts match exactly.
In our tests, picking 5 areas to watch every two weeks caught 44% of new clusters in 2024 to 2026, against 7% for random picks. That is about 6x better than guessing, on years the model never saw.
We ran the team planner logic on today's real clusters. Six teams placed this way reach 25,990 at-risk homes, against 9,988 when each team goes to one of the six worst blocks. That is 2.6x more.
We placed 99% of cases on the map despite the typos, and we flag the rest instead of dropping them.

What we learned

What is happening next door matters most. Adding clusters in neighbouring areas lifted new-cluster detection from 38% to 44%.
Simple guesses deserve respect. In the very quiet 2024 to 2026 period, "same as last week" still beat our early model on Singapore-wide case numbers. We plan to show that openly in the app.
Small samples need honesty. Our latest test has only 68 new clusters, so the results are a good sign and not yet proof.

What's next for DengueRadar

Build sprint, 12 to 26 October: set up the Databricks pipeline, train the area forecast, then build the dashboard, email alerts and ask-a-question feature.
A locked forecast. On 12 and 19 October we will save our forecast in a table nobody can edit. On Demo Day (27 October) we will open it and compare it live with what really happened.
Finer Wolbachia detail, measured street by street instead of by whole neighbourhood.
Satellite ground temperature to fill the gaps between weather stations, if time allows.
Saving NEA's cluster list every day, so the history never has a gap again.