Gawks

Inspiration

Singapore became a super-aged society in 2026, with 19.5% of residents aged 65 and over. While going through SingStat's Key Indicators on the Elderly, one number stood out to us: 88,400 seniors lived alone in 2025, more than double the 41,200 in 2015, and another 179,700 lived only with other seniors.

Social isolation is a stronger predictor of early death than smoking or obesity, yet outreach by the Silver Generation Office and Active Ageing Centres (AACs) is still largely manual. AACs grew from 60 in 2021 to 223 in 2024, but national averages can't show which towns are being missed. We wanted to give community workers a simple answer to one question: where should we go first, and why?

What it does

SilverWatch will rank Singapore's HDB towns by social isolation risk, today and towards 2030. We're planning to build it end to end on Databricks, and for each town it will show:

a risk score and its rank on a heatmap, with a today/2030 toggle
a plain-English reason, e.g. "many seniors, fast-growing elderly population, few AAC places"
a Genie chat space where coordinators can ask questions like "Which towns will have the most seniors living alone by 2030?" without writing code

The risk score will combine standardised features for each town $t$ into one weighted index:

$$ z_{k,t} = \frac{x_{k,t} - \mu_k}{\sigma_k}, \qquad R_t = \sum_{k} w_k \, z_{k,t} $$

using the elderly population, the future-elderly growth ratio $g_t = F_t / E_t$, and AAC supply per senior.

What we learned

Even at the planning stage, digging into the data changed our approach:

Check what the data actually contains. Only the HDB dataset is broken down by town; the rest are national figures. Applying one national living-alone rate to every town would just rescale the elderly count, so we had to rethink which features genuinely differ between towns.
Small data issues can break a ranking. The HDB data stops at 2018, and Hougang appears multiple times in that year. Missing this would have silently pushed it up our list.
Simple stats need careful wording. There are about 3,400 seniors per AAC if spread evenly, but that's an average, not each centre's actual caseload.
No ground truth means choosing explainability. There's no official label for which towns are "isolated", so we chose an explainable composite score over a black-box model, and plan to test that the top 10 stays stable when the weights $w_k$ change.
The future matters as much as the present. Sembawang's future-elderly population is more than double its current elderly population, and Bishan's is 1.8 times. A dashboard that only looks at today would miss these towns.

What's next for Gawks

Build the pipeline: ingest the datasets into Delta with Lakeflow, clean and map them to HDB towns, and build the gold feature table, all governed in Unity Catalog
Build the model and dashboard: compute the risk scores, track experiments in MLflow, and create the AI/BI dashboard and Genie space
Add local data: Census 2020 planning-area data on one-person households and 1–2 room flats, for a truly local living-alone signal
Stretch goal: an AAC placement recommender that finds where new centres would cover the most seniors who are currently missed