Inspiration
Heat waves kill more Americans than any other weather event, and many of the people who die are taking common medicines. Antipsychotics, diuretics, insulin and others can blunt sweating, drain fluids or drop blood sugar just when the body needs to cope with heat. Guidance tells clinicians to "review medications" before summer, but it rarely says which medicines matter most, why, or what to do. We wanted to give prescribers that answer, grounded in the research.
What it does
HeatRx shows which medicines make heat more dangerous and why, then turns that evidence into action for each patient.
- Evidence atlas: pick a medicine and see where it acts in the body during heat. Each effect is backed by studies,
for example:
- antipsychotics: about 2× the odds of dying in the 2003 French heat wave (aOR 2.09);
- antidepressants: about 1.7× (aOR 1.71);
- insulin: about 40% higher odds of a serious low on the hottest days (OR 1.38).
- The graph: every link opens to the papers behind it. You can see whether a medicine was studied in heat directly, or connects to heat through a shared effect in the body (drug → kidney injury ← heat).
- Patient plan: enter a patient's medicines, conditions and home situation, and get a mitigation-first plan. It keeps the medicine where possible, adds the right monitoring, and covers cooling, hydration and sick-day rules, ready to copy into the EHR or hand to the patient.
- Patient panel: on a heat advisory day, a clinic sees which patients to call first, mapped against NYC's neighbourhood heat vulnerability.
- Evidence appendix: the full study list, guideline coverage and gaps for anyone who wants to dig in.
How we built it
- Data: the System Public Graph, using the 2.0M-finding vetted snapshot plus the 5.9M-finding live API, with unvetted rows always tagged.
- Matching: we matched the drug and heat ends of each finding by variable name, and the body effects in the middle by concept. Each drug finding is classed as raised, lowered or unclear risk from its confidence interval.
- Checking: headline numbers were checked against the paper abstracts, and graph labels that didn't match the paper were corrected.
- Pipeline: a Python pipeline (SQLite, standard library) builds one evidence bundle of about 3,300 findings from about 2,000 papers. It is linked to 8 heat-health guidance documents (CDC, NYC Health, UKHSA, Health Canada, ANSM, WHO Europe and others), FDA drug labels, Medicare Part D prescriber data, the NYC Heat Vulnerability Index, CDC PLACES and National Weather Service alerts.
- App: a React 19 + TypeScript + Vite app (Tailwind, Motion, Zustand), with an interactive evidence graph, a transparent rules engine for patient actions, and paper abstracts loaded live from PubMed.
Challenges we ran into
- Thin direct evidence. Only about 20 findings from 15 papers study a medicine and heat together. We used the graph to connect indirect relationships instead: the drug and heat each linked to the same effect in the body.
- Keyword traps. "ARB" matched a river basin, insulin showed up as a lab marker, and fever control looked like "temperature". We had to read the matched rows, not just count them.
- Number versus label. Extraction sometimes got the number right and the label wrong. Checking the abstracts settled it.
- Not over-alarming. A risky medicine is rarely a reason to stop it, so we had to design recommendations that lead with mitigation.
Accomplishments that we're proud of
- Every number on screen traces to a study in the graph or a named public source, and you can click through to the paper.
- One product that goes from evidence to a phone call: atlas → graph → patient plan → clinic call list.
- Plain language that clinicians can act on, with a glossary for every statistic.
What we learned
- A knowledge graph is most useful for the gaps between studies. Two-hop paths surfaced plausible risks that no single study tested.
- Absence from the graph is not evidence of absence: guidance warns about some drugs that have never been studied in heat.
- Clinicians want "what do I do today," not "how strong is the odds ratio". The evidence has to sit one click behind the action.
What's next for HeatRx
- Pilot with a clinic or pharmacy before summer: run the panel against real medication lists in the EHR, triggered by NWS heat alerts.
- An evidence chat assistant that answers clinician questions only from the graph and cites the papers.
- Keep the evidence current as new studies enter the System Public Graph, and validate the patient rules with clinicians.
- Expand beyond NYC with local heat-vulnerability data.
