Heat illness has 33 names in Chicago's records, and most aren't 'heat illness'

Counting heat deaths and heat illnesses depends on a decision made long before any counting starts: which diagnoses get to be called heat-related. That list is short, largely inherited, and — on the evidence of a study published on 26 August — badly incomplete.

Letting the data name the diagnoses

Researchers analysed 916,904 acute-care visits, including emergency department and urgent care encounters, among 372,140 adults across four health care systems in Chicago from 2011 to 2023 [s1]. Rather than start from a predefined list of heat diagnoses, they ran what they call a heat-wide association study, borrowing the logic of a genome-wide scan: test every diagnosis code against temperature and see which ones surface.

The design ran in two stages. Quasi-Poisson regression first screened diagnosis codes for association with heat, retaining 44. Those 44 then went through distributed lag non-linear models in a time-stratified case-crossover design to characterise the exposure-response relationship and the lag structure, leaving 33 clinically reportable diagnoses [s1].

The retained list is the finding. Elevated cumulative odds over lags 0 to 3 days appeared for fluid-balance and renal disorders — expected — but also for dermatologic and pressure-related conditions, multiple sclerosis, a cannabinoid-related mental and behavioural disorder, varicose veins of the lower extremities, and multiple injuries and external causes [s1].

Fluid-balance and renal disorders are the diagnoses a conventional heat surveillance list would already contain. The rest are not. Several of them — the cannabinoid-related disorder, varicose veins — are the kind of signal a hypothesis-free scan will surface and that needs replication in other cities before anyone builds a surveillance category around it. The authors present the work as improving the clinical characterisation of heat-associated morbidity in acute-care settings, which is a claim about description rather than causation [s1].

The methodological point stands regardless of which individual codes survive replication. Prior studies relied on restricted diagnoses and diagnostic categories, and that reliance may miss or misclassify heat-related illness [s1]. Any national tally built from the conventional shortlist is counting a subset of an unknown size.

The visits that never happen

A second paper, published 29 August, addresses a failure that no diagnosis-code method can detect: people who do not present at all.

Working in Phoenix, the authors combined persistent change analysis, year-on-year difference maps, and a Poisson Pseudo-Maximum Likelihood spatial-lag-of-X model with two-way fixed effects, to identify which sociodemographic and environmental characteristics amplify or buffer heat risk over time [s2]. Heat-related emergency calls rose in south and central Phoenix and fell in the north [s2].

The census tracts where calls were increasing had higher rates of uninsurance, poverty and disability, and lower educational attainment. They also had higher proportions of non-US citizens, individuals with limited English proficiency, households without a vehicle, and Hispanic residents, with standardised mean differences of 0.55 to 0.75 [s2]. Tracts where calls were decreasing had more non-Hispanic White residents and higher housing costs (SMD 0.57 to 0.77) [s2].

The authors' interpretation of the decreasing side is the part that should give any surveillance system pause. They read the negative association between emergency calls and certain social attributes as evidence of barriers to service use, rather than as evidence of fewer heat-related emergencies [s2]. A tract with falling call volumes is not necessarily a tract that is getting cooler or better adapted. It may be a tract where fewer people are willing or able to call.

What neither study can do

Both are observational and both are single-city. The Chicago analysis covers four health systems in one metropolitan area over 13 years, and a case-crossover design controls for stable individual characteristics by construction while remaining vulnerable to time-varying confounders that track temperature [s1]. The Phoenix analysis is ecological — it describes tract-level characteristics associated with tract-level call trends, and cannot attribute any individual call, or any absent call, to any individual circumstance [s2].

Neither establishes a mechanism. Neither says how many heat-related presentations national surveillance is missing, only that the question is live from two directions at once: the diagnoses that are not on the list, and the people who never enter the records.

What to watch

The immediately testable thing is whether the Chicago diagnosis list replicates in other cities with different climates, populations and coding practices. A heat-wide scan is cheap once the data infrastructure exists, and 33 candidate diagnoses is a specific enough hypothesis set to test rather than argue about.

The harder question is the Phoenix one. If call volumes fall in the neighbourhoods with the least insurance and the fewest vehicles, then every heat adaptation programme evaluated against emergency call data is being scored on a measure that improves when access gets worse.

Sources

• [s1] Identifying heat-related diagnoses in emergency department visits among adults in Chicago: A heat-wide association study. Science Advances, 26 August 2026. https://doi.org/10.1126/sciadv.aed3977
• [s2] Structural inequality shapes spatiotemporal patterns of heat-related emergency calls. Health & Place, 29 August 2026. https://doi.org/10.1016/j.healthplace.2026.103735