Bob's DengueRadar

Our Team

Lincoln You Liheng - NUS, Finance & Quantitative Finance, Year 1
Park Jungwoo - NUS, Chemical Engineering & Quantitative Finance, Year 1
Noriko Khang - NUS, Mathematics & Computer Science, Year 1

Inspiration

Singapore's worst dengue year was 2020: 35,315 cases, after the virus switched to the DENV-3 serotype, which few residents were immune to. In 2026 the same signal is back. DENV-3 rose from 19.1% of cases in January to 59.9% in June.

Meanwhile Project Wolbachia, which cuts dengue risk at release sites by over 70%, reaches half of all households this year. That leaves about 800,000 households unprotected, with no sites announced beyond 2026. Public dengue data shows clusters only after cases appear. We wanted a tool that looks ahead, and that points Singapore's best weapon where it will prevent the most cases.

What it does

DengueRadar forecasts dengue-cluster risk for each of Singapore's 55 planning areas, 2 and 8 weeks ahead. It adjusts for where Wolbachia already protects, and ranks unprotected estates by the cases a release would avert. Six tools sit in one dashboard:

Live cluster map: today's NEA clusters by planning area, with alert level and Wolbachia coverage.
Fortnight risk forecast: Low / Medium / High per area for the next 2 weeks, plus an 8-week outlook.
"Why here?" area brief: the top three drivers behind each score, in plain English.
Wolbachia expansion planner: ranks unprotected areas by expected cases averted per 1,000 households, with a what-if toggle.
Hit list and alerts: a weekly top-10 for source reduction, and an alert when an area turns High.
Trust panel and 2020 replay: forecast vs actual, skill against a naive baseline, and a replay of the record outbreak.

Expected cases averted in area a, where ŷₐ is the forecast cases, E ≈ 0.7 is NEA's reported risk reduction at release sites, and Hₐ is the number of households in the area:

$$\text{averted}_a = E \cdot \hat{y}_a, \qquad \text{per 1,000 households} = \frac{1000 \cdot \text{averted}_a}{H_a}$$

How we're building it

Ingest: Lakeflow Jobs pull data.gov.sg dengue clusters daily and rainfall, temperature and humidity hourly. Auto Loader lands them in bronze Delta tables. History is a one-off bulk load: NEA's yearly weather files (about 1.1 GB each) are aggregated to daily station totals on arrival.
Transform: Lakeflow Declarative Pipelines take the data from bronze to silver to gold with data-quality expectations. They keep every daily cluster snapshot as history and join stations and clusters to the 55 URA planning areas.
Model: gradient-boosted trees on rainfall, temperature and humidity lagged 1–16 weeks, past cluster activity, a DENV-3 flag and Wolbachia status. MLflow tracks every run against recent-history and seasonal baselines in walk-forward backtests. The best model is registered in Unity Catalog and scored weekly.
Serve: an AI/BI dashboard, a Genie space for plain-English questions, SQL alerts, and area briefs from a Foundation Model API endpoint.
Govern: Unity Catalog for every table and model, with lineage from API to forecast.

Challenges we expect

No history in the live feed. The cluster dataset shows only active clusters. We use SGCharts' archive of NEA cluster snapshots (twice weekly, May 2013 to November 2020) for area-level labels, and archive the live feed daily from now on.
Network limits on Free Edition. Outbound internet is restricted to trusted domains. If data.gov.sg is blocked, a scheduled external poller drops files into a Unity Catalog Volume for Auto Loader.
A quiet year. 2026 had 2,424 cases to 22 September, so our demo replays 2020 to show lead time on real data.
Wolbachia isn't an open dataset. We enter release areas by hand from NEA's website and flag them as manual input.

What we've learned so far

Serotype switches, not just weather, set up Singapore's biggest outbreaks.
Containing an outbreak takes about two months, so a useful warning has to come weeks before cases appear.
NEA already forecasts dengue nationally. What's missing in the open is risk by planning area that accounts for Wolbachia and explains itself.

What's next

In the two-week sprint, week 1 covers ingestion, the cluster archive, bronze-to-gold tables, a baseline and model v1. Week 2 adds the Wolbachia layer, the expansion planner, the dashboard, Genie, alerts and the 2020 replay. Our team of three splits into data engineering, modelling, and dashboard and story.