Inspiration
In 2020, Singapore had its worst dengue year on record: 35,315 cases and 32 deaths.
NEA's cluster map shows where dengue has been, but a cluster only turns red after 10 cases are reported. And reported cases are just the tip of the iceberg. In an NEA study during the 2007 epidemic, 78% of recent infections caused no symptoms. Only 5 of the 61 people who did feel sick saw a doctor.
If you wait for reported cases, you're always behind. So we built DengueRadar: dengue, two weeks early.
How we built it
DengueRadar forecasts dengue risk two weeks ahead, rates each area Low, Medium or High, and ranks the blocks officers should check first. A Time Machine replays past outbreaks to show what we would have predicted against what actually happened.
Our model predicts weekly cases in each area from the last four weeks of cases plus rainfall and temperature from 1 to 6 weeks earlier. The weather is lagged because rain takes weeks to turn into mosquitoes and then into cases. Since dengue counts are small whole numbers, we use LightGBM with a Poisson objective.
Every forecast is scored against a simple baseline: the average of the last four weeks. If our model can't beat that, it isn't worth using.
On Databricks, Lakeflow Jobs and Auto Loader ingest live weather and cluster data into Delta tables. A Lakeflow pipeline cleans the data, and MLflow tracks our models. A Databricks App, a Genie space and an AI/BI dashboard serve the results. Block-level data sits behind Unity Catalog access controls.
For Round 1, we explored the data and built a clickable phone mockup of the app.
Challenges
Misreading the data. We misread two columns in the cluster archive at first and only caught it by going back to the source. A gap in history. The archive stops in 2020, right before Wolbachia releases scaled up and DENV-3 took over, so we'll check whether old patterns still hold. Live data doesn't match old data. We had to build features that exist in both. Keeping the Time Machine fair. It can't peek at the future, so every replay is precomputed from data available on that date.
What we learned
Most dengue is invisible, so forecasting risk matters more than mapping cases. One infection isn't full immunity. Dengue has four types, which is why a shift to DENV-3 can come before an outbreak. A simple baseline is hard to beat. If our model can't beat it, we'll say so.
Built With
- canva
Log in or sign up for Devpost to join the conversation.