Inspiration
We spend about 90% of our time indoors, yet funders deciding where to put money into healthier buildings face a scattered evidence base. The System Public Graph links thousands of studies through their statistical findings, so we asked: for indoor air, which actions does the evidence actually support, and where are the gaps worth funding?
What it does
Airways for Impact is an evidence explorer for 8 indoor air quality actions, from damp and mould remediation to radon mitigation and ventilation. For each action it shows two kinds of evidence:
- Problem evidence: studies linking the exposure the action reduces (damp, combustion pollution, radon, allergens) to health outcomes.
- Fix evidence: studies that test the action itself.
Each is graded Strong, Moderate or Limited based on how many studies there are, the share with a statistically significant result, and how many are trials. An interactive graph connects each action to 7 health outcome groups and 6 wider outcomes (energy, affordability, comfort, school and work, outdoor air, healthcare systems). Users can filter by setting (home, school, workplace) and region (all or UK only), click any link to see the underlying findings and DOIs, and combine actions into a package to compare side by side.
The headline insight: we know a lot about what harms people indoors, and very little about what fixes it. About 500 studies show harm, against about 80 that test a fix, of which only 15 are trials. Damp and mould has 102 harm studies but just 2 testing remediation, which makes it the clearest case for funding better evaluation.
How we built it
- Loaded the System Public Graph v2.1 tables into DuckDB and curated 841 studies relevant to indoor air.
- Matched studies to actions with keyword rules on the variable text, because System's concept labels are built for medical terms and often mislabel building terms.
- Graded evidence strength in Python from confidence intervals, p-values, statistic type and System's trial flag.
- Added outside context: the English Housing Survey, Census 2021 overcrowding, Cochrane reviews, NICE cost-effectiveness guidance, radon cost-per-QALY modelling, the HAPIN and HAPI trials, and Global Burden of Disease 2021.
- Exported a JSON graph and built a single static HTML app (plain JavaScript and SVG) on GitHub Pages, combining Christina's interface design with the evidence graph.
- Used Claude, Anthropic's AI model, throughout for data processing, analysis and code.
Challenges we ran into
- Matching studies to actions: concept-level labels put building terms in the wrong place, so we switched to matching on the variable text and fixed misclassifications by hand (radon lung cancer was first filed under "respiratory").
- Defining "strong evidence" honestly: the findings are associations extracted by an AI model, and high significance rates partly reflect publication bias. We rate the evidence base, not effect size, and say "too few studies to judge" below 5 studies.
- Missing context: 28% of studies don't record a country, only 2% are from the UK, and 37.5% don't say the setting. We show these gaps rather than guessing.
- Scope: we narrowed from healthy buildings in general to indoor air, and added crowding as an exposure because it worsens damp and stale air.
- Merging two parallel workstreams (Christina's interface and the data graph) into one app that works across browsers.
Accomplishments that we're proud of
- Turning a huge general-purpose graph into a focused, decision-ready tool for one question.
- Separating "the problem is real" from "the fix works", which reframes where investment and research money should go.
- Every line in the graph traces back to real findings and DOIs.
- Showing gaps (no fix studies for radon, no evidence on outdoor air co-benefits) as clearly as strengths.
What we learned
- Cutting exposure isn't the same as improving health: the HAPIN trial cut household air pollution but didn't change birth weight or pneumonia rates.
- Some actions only pay off when targeted. Untargeted insulation costs over £100k per QALY, and single dust-mite measures show no benefit.
- Prevention can be much cheaper than cure: radon protection built into new homes costs about £11.4k per QALY, against £36.8k for retrofitting.
- With AI-extracted evidence, being open about how it was made matters as much as the findings.
What's next for Airways for Impact
- Add effect sizes and study quality to the grading, not just counts and significance.
- Expand to other healthy-building themes (heat, noise, falls) using the same problem-versus-fix framework.
- Bring in cost data across all actions to estimate return on investment.
- Validate the keyword matching with domain experts and add more UK and local data, such as local housing condition surveys.
