BreatheWise

Finding mould-risk homes before they make people ill.

Inspiration

In 2020, two-year-old Awaab Ishak died from prolonged exposure to mould in his home. His death led to Awaab's Law, which now requires social landlords to fix damp and mould within fixed timeframes. It also exposed a harder truth: the harm from cold, damp housing is predictable days in advance, yet the system acts only after someone reports a problem or arrives sick. Around two million people in England live in homes with serious damp and mould, and the burden falls hardest on children, older people and disadvantaged communities. We wanted to move that response from reactive to preventative, and to reach the people who carry the most risk and hold the least power to avoid it. The Edinburgh Homes, Heat and Healthy Kids study sharpened the idea, especially its finding that insulating a home without ventilating it can trap moisture and make the home damper.

What it does

BreatheWise scores each area for its risk of damp, mould and the respiratory illness they cause. It takes a regional cold-weather alert and turns it into a ranked, borough-level warning of where harm will concentrate over the next week, before symptoms appear. NHS planners, Integrated Care Boards, councils and social landlords use the score to act early: prioritise homes, pre-position care, and reach vulnerable residents through channels that already hold them, such as energy suppliers' Priority Services Register. Every home is rated from Low to Very High.

How we built it

The core insight is about timing. A cold spell raises respiratory illness fast, driven by cold air on the airways and the winter viruses that spread when people shut their windows. Damp and mould work slowly, building over a season and leaving a population primed to crash when the cold arrives. So the cold alert is the trigger, and housing is what decides who gets hurt.

We joined four open data layers on borough codes:

• Weather and forecast: Open-Meteo and the live UKHSA cold alert.
• Housing: the GLA London Building Stock Model and EPC ratings.
• Vulnerability: the Index of Multiple Deprivation and BEIS fuel poverty statistics.
• Health outcome: OHID Fingertips respiratory admissions and OpenPrescribing.

Each borough gets a vulnerability score, which we multiply by the live cold-alert severity to produce a time-varying risk:

$$V_b = w_1\,\text{IMD}_b + w_2\,\text{FuelPoverty}_b + w_3\,\text{COPD}_b + w_4\,\text{HousingRisk}_b$$

$$\text{Risk}_{b,t} = V_b \times S_t$$

where $V_b$ is the vulnerability of borough $b$, $S_t$ is the cold-alert severity at time $t$, and the health response is lagged by 3 to 7 days to match how respiratory admissions follow a cold snap. The method builds on Rudge and Gilchrist's 2007 Newham study, which predicted excess winter respiratory admissions from a fuel poverty risk index. We modernised it with real energy data and a forward forecast, the two things its authors said they lacked. We built the dashboard on Lovable, and an LLM layer turns each borough's numbers into a plain-language action brief.

Challenges we ran into

Our first model was wrong in an instructive way. We assumed a cold spell creates damp, which creates mould, which makes people ill, all in the same week. The timing did not hold, because mould takes weeks to build and to act on the lungs. Separating the fast trigger from the slow amplifier reshaped both the model and the story we tell.

Beyond that, the datasets sit on different geographies, so joining postcode, LSOA and borough records took real care. We had no house-level admissions data, so we validated at borough level and calibrated to the published effect sizes from the Newham study where our own data was thin. We also worked to keep the score explainable rather than a black box, since clinical users will not trust a number they cannot interrogate, and we were careful to frame it as a population-level risk signal rather than an individual prediction.

Accomplishments that we're proud of

We built a forecast-driven, explainable risk engine end to end in a day, on open data alone. We grounded it in a peer-reviewed London method and gave it the live data that method never had. And we drew a clear line from a blunt regional alert to action that can reach the actual vulnerable homes.

What we learned

We learned the retrofit paradox, and why housing records on their own can mislead. We learned to separate trigger from amplifier, the idea that makes the whole timing make sense. We learned that the data needed to act already exists but is scattered across health, housing and energy, so the real work is connecting it. And we learned how to keep a composite index explainable, and how to structure prompts so an LLM's output maps cleanly to the fields a dashboard needs.

What's next for BreatheWise

We want to move from borough to house-level resolution using linked health and housing data, and to scale the London MVP across the UK. We plan a resident-facing chatbot so people can check and act on their own home's risk. We want to formalise the cross-sector loop, especially with energy suppliers' Priority Services Register, so a warning reaches a named vulnerable household. And we want to run a proper validation study using the Edinburgh data-linkage methodology.

Built With

• claude