Round 1 proposal: A2 DengueRadar
Which five Singapore planning areas should community teams prioritise for dengue prevention over the next two weeks?
DengueRadar will turn a two-week forecast of reported dengue cluster activity into a practical prevention worklist for town councils and volunteer coordinators. Users will see a map, Low/Medium/High priority bands, recent activity and weather trends, and five areas to prioritise for breeding-site checks and resident outreach.
Data and approach
We will combine NEA Dengue Clusters, SGCharts' historical NEA locality snapshots, NEA historical and live rainfall and air temperature data, and URA planning-area boundaries.
Historical locality coordinates will be mapped to planning areas. We will check duplicates, invalid coordinates, missing snapshots and weather coverage. Missing observations will remain unknown rather than being treated as zero activity.
Proposed Databricks architecture
Scheduled Python ingestion will load observations into Delta tables. Cleaned weekly features will support a four-week mean baseline, an interpretable regression model and pretrained Chronos-2. MLflow will track model comparisons, and a Databricks App will display the forecast and prevention worklist.
Chronos-2 CPU inference will be tested early. The best feasible model will power the demo. Fine-tuning is outside the core sprint scope.
Validation and intended impact
We will use chronological backtests to measure forecast error and the share of future reported cluster activity captured by the five prioritised areas.
Our hypothesis is that the selected five areas can capture 10 percentage points more activity than the baseline ranking. This is a target to test, not an achieved result or a claim about infections prevented.
Two-week build scope
Week 1: validate data, map localities, build the baseline and test CPU inference. Week 2: compare models, build the map and worklist, and prepare an end-to-end demo.
The historical archive ends in 2020, so historical replay will provide the main evaluation. Any live forecasts will be clearly labelled experimental.
This is an idea submission. Implementation and evaluation are planned for the build sprint.
Log in or sign up for Devpost to join the conversation.