HealthPulse: Community Care Demand Forecasting

Inspiration

Singapore became a super-aged society in 2026: 825,300 residents (19.5%) are now 65 or older, and Healthier SG is moving care from hospitals into the community. But ageing is uneven — 29.1% of Outram residents are 65+, versus 19.5% nationally — so national projections cannot tell planners where new community care capacity is needed. We asked: where will community healthcare demand outgrow local capacity over the next five years?

What it does

HealthPulse forecasts 2031 community care demand for all 55 planning areas and sets it against current and planned capacity. For MOH and regional health planners it shows:

2031 demand for community hospital beds, nursing home beds and polyclinic access by planning area
the gap per 1,000 seniors after the 31 announced healthcare facilities open
a ranked map of the 5 most underserved areas, kept only if they stay top-5 across all scenarios
a p-median test of where one more polyclinic would cut travel distance most
a Genie space so planners can ask follow-up questions in plain English

How we will build it

Ingestion: open-data files land in a Unity Catalog Volume and load incrementally into Bronze Delta tables with Auto Loader, scheduled by Lakeflow Jobs.
Transform: Lakeflow Spark Declarative Pipelines build Silver/Gold tables — one planning-area key across URA boundary versions, geocoded facilities and data-quality rules.
Model: cohort-survival projection of single-age population (332 subzones, 2011–2026) using SingStat life tables, multiplied by MOH age-sex utilization rates; backtests against a naive baseline are tracked in MLflow.
Output: an AI/BI dashboard on Databricks SQL (gap map, top-5 areas, scenario toggle) plus Genie for plain-English Q&A.
Governance: Unity Catalog lineage, table tags and access control across every layer.

Data

SingStat population by subzone, single year of age and sex (2011–2026) and life tables; MOH hospital admission rates by age and sex and inpatient beds; the MOH list of 31 upcoming facilities; facility locations from AIC, MOH and CHAS (data.gov.sg); URA Master Plan boundaries; OneMap geocoding.

Challenges

Utilization data exist only at national level, so area demand applies national age-sex rates; we state this openly and test ranking stability across scenarios.
Facility capacity records are incomplete (many nursing homes have no published bed count), so unverified capacity is flagged rather than hidden.
Planning-area boundaries changed between Master Plan 2014, 2019 and 2025, so all data had to be aligned to one geography.

What we learned

Good planning tools need to answer where, not just how much — and they need to be honest about uncertainty. Assembling a clean, traceable multi-source dataset was as important as the model itself.

What's next

In the two-week sprint: the full pipeline, the 2031 projection with backtests, the gap ranking, the dashboard and Genie. Stretch goals: an ageing-scenario toggle and a Healthier SG prevention scenario.