Team
Team Name: SMO Project: A2 - DengueRadar - Real-time Outbreak Forecasting
| Name | Institution | Course / Programme | Year | |
|---|---|---|---|---|
| Le Thi Huong Giang | Singapore Management University | Master of IT in Business (Artificial Intelligence) | 2nd Year | [email protected] |
| Bui Tien Thanh | Singapore Management University | Master of IT in Business (Artificial Intelligence) | 2nd Year | [email protected] |
Inspiration
Dengue remains a recurring public-health challenge in Singapore, and its severity can change quickly. Annual reported cases increased more than six-fold from 2021 to 2022, showing how rapidly dengue activity can escalate.
Existing dengue cluster maps are valuable for understanding where active transmission is occurring today, but they provide limited forward visibility into where risk may rise next.
This inspired us to ask:
Can Singapore's open dengue, weather and geospatial data be combined to provide earlier, explainable risk signals?
Our goal with DengueRadar is to move from reactive monitoring toward predictive risk intelligence.
What it does
DengueRadar is an ML-powered dengue risk-intelligence platform designed to forecast where dengue risk may increase over the next 2 weeks.
It follows a five-step decision-support workflow:
- Forecast — estimate dengue activity 14 days ahead using historical dengue and weather signals.
- Classify — translate forecasts into Low / Medium / High risk levels.
- Explain — use SHAP-based explanations to identify the features contributing most to each forecast.
- Prioritize — rank areas so teams can identify where earlier attention may be needed.
- Alert & Q&A — flag risk escalations and allow users to explore forecasts and trends through natural-language queries.
The platform is intended to support three stakeholder groups:
- NEA operations teams — identify where vector-control attention may need to be prioritized.
- Public-health analysts — investigate emerging risk and understand key forecast drivers.
- Community teams — identify areas where preventive outreach may be most useful.
How we are building it
DengueRadar combines several types of Singapore open data:
- Dengue activity — historical dengue cases and current active clusters.
- Weather signals — historical and real-time rainfall and temperature observations.
- Geospatial context — planning-area boundaries for spatial aggregation and visualization.
- Environmental context — green-space and hawker-centre density as exploratory stretch features.
The core hypothesis is that combining historical dengue activity with recent weather conditions can provide more useful forward-looking signals than recent case history alone.
On Databricks, we designed an end-to-end architecture:
Open data → Lakeflow ingestion → Bronze / Silver / Gold Delta pipeline → feature engineering → ML training and evaluation → model registry → inference → Databricks App
The machine-learning layer will compare a recent-history baseline against models such as LightGBM and Random Forest.
We plan to use temporal validation to ensure future observations are not used during model training, while MLflow will track experiments, feature sets, model parameters and evaluation metrics.
The best-performing model will be registered through Unity Catalog and used to generate the 14-day forecast.
We also plan to use:
- SHAP for model explainability and forecast drivers.
- Databricks Alerts for risk-escalation notifications.
- AI/BI Genie Agents for natural-language exploration of forecasts, trends and alerts.
- Unity Catalog for governance, access control and lineage.
- Databricks Apps for the user-facing risk-intelligence experience.
What makes DengueRadar different
DengueRadar is not intended to be another dashboard showing current cases.
Its value comes from moving through:
Monitoring → Forecasting → Explanation → Prioritization → Action
Rather than only showing where dengue activity is occurring today, DengueRadar aims to estimate where risk may rise over the next 2 weeks.
The model will be evaluated against a simple recent-history baseline using forecasting metrics such as MAE and WAPE.
Because the final output is also translated into Low / Medium / High risk levels, we will evaluate operational classification performance using metrics such as High-risk Recall and F1.
We also plan to use MLflow experiments to test whether adding rainfall, temperature and environmental context provides measurable predictive value beyond historical dengue activity alone.
Challenges
One of the most important challenges is data granularity.
Dengue, weather and geospatial datasets may be available at different spatial and temporal levels, so they need to be carefully aligned before modelling.
Other challenges include:
- Handling missing or irregular weather observations.
- Mapping weather stations and dengue clusters to consistent geographic units.
- Avoiding temporal leakage when creating lagged and rolling features.
- Maintaining the same feature-engineering logic during training and inference.
- Ensuring current cluster data is not used as a historical ML feature unless historical snapshots are available.
- Providing model explanations without presenting associations as causal relationships.
These limitations are part of why we want to evaluate DengueRadar rigorously rather than simply produce a visually appealing risk map.
What we learned
Designing DengueRadar reinforced that a useful data-science product requires more than training a machine-learning model.
A credible solution also requires:
- Reliable data ingestion and transformation.
- Consistent feature engineering.
- Defensible temporal validation.
- Experiment tracking and reproducibility.
- Model governance.
- Explainability.
- Data-quality monitoring.
- A clear decision that the end user can act on.
The project also gives us an opportunity to explore how Databricks can support the full data and ML lifecycle within one platform — from Lakeflow and Delta Lake, through MLflow and Unity Catalog, to Databricks Apps, Alerts and Genie Agents.
Next steps
Our Round 1 submission presents the proposed solution architecture and an illustrative application prototype.
If selected for Round 2, we plan to:
- Build the end-to-end Bronze / Silver / Gold data pipeline.
- Align historical dengue and weather data for model training.
- Engineer lagged dengue, rainfall and temperature features.
- Train and compare forecasting models against a recent-history baseline.
- Track feature-set and model experiments using MLflow.
- Convert 14-day forecasts into Low / Medium / High risk levels.
- Add SHAP-based explanations for forecast drivers.
- Connect new weather observations for production inference.
- Implement risk-escalation alerts.
- Deploy the interactive DengueRadar experience through Databricks Apps.
Our goal is simple:
Help teams see not only where dengue is today, but where risk may rise next.
Built With
- databricks
Log in or sign up for Devpost to join the conversation.