How AI Helps Cooling Centers Reach Vulnerable Patients During Heat Waves

Cooling centers are one of the few heat-wave interventions that can move a person’s core risk in the right direction quickly: get out of the hot apartment, sit in a cooler public space, drink fluids, avoid the next several hours of physiologic strain. The problem is scale. When Bedi and colleagues estimated the number needed to treat for cooling centers to prevent one heat-related death, the range ran from 439,222 people using statistical excess-death estimates to 9,900,990 people using CDC death certificate data.[1] That wide range is not a rounding error; it reflects the difficulty of measuring heat mortality itself. But either denominator leaves the same operational message: if a city simply tells everyone under a heat alert to go to a cooling center, the public health signal is spread across an enormous population, while the lethal risk is concentrated in a much smaller group.

That is where AI becomes clinically interesting in heat-wave public health response and cooling-center referral. The useful version of the technology is not a model that announces summer will be hot. Clinicians, emergency managers, and older adults in unairconditioned housing already know that. The useful version is a model that narrows the outreach universe before the heat wave peaks: which patients have heart failure, chronic kidney disease, chronic lung disease, diuretics, beta-blockers, psychotropics, or anticholinergic burden; which patients have had prior emergency use; which addresses sit in neighborhoods where heat emergencies cluster; and, where the information exists, which homes lack air conditioning.

Why broad heat advisories miss the clinical problem

A heat advisory is a population message. Heat illness, especially fatal heat illness, is often a patient-specific event. Two people may live on the same block under the same temperature warning. One is a healthy adult who can work remotely in an air-conditioned room. The other is an older patient with chronic kidney disease, congestive heart failure, limited mobility, a loop diuretic, and no reliable ride to a cooling center. Treating those two people as identical recipients of the same public message is administratively simple and clinically imprecise.

The Bedi cooling-center analysis matters because it puts a hard number on that imprecision. A number needed to treat in the hundreds of thousands or millions does not mean cooling centers are useless. It means that offering them at undifferentiated population scale is a very blunt intervention for preventing death. The smaller the truly high-risk group that can be identified and reached, the more meaningful the referral becomes for the people most likely to benefit.

This is also why the discussion should not stop at age. Older adults are a central risk group, but age alone is a poor operating list. A 78-year-old with a cool home, strong family support, and stable transportation does not create the same outreach burden as a 63-year-old with renal disease, psychiatric medications, unstable housing conditions, and no air conditioning. A health system trying to act before the emergency department fills needs a ranked work queue, not a weather flyer.

What an EHR-based heat risk model can actually see

Electronic health records are not a complete map of vulnerability, but they contain signals that are directly relevant to heat physiology. Cardiovascular disease can limit the body’s ability to increase cardiac output under heat stress. Kidney disease can reduce physiologic reserve when dehydration begins. Respiratory disease can make heat and poor air quality more dangerous. A review of 134 studies identified older adults, outdoor workers, and people with cardiovascular, renal, and respiratory disease as groups bearing a substantial share of heat-related morbidity and mortality.[4]

Medication lists are just as important. Diuretics can reduce intravascular volume. Beta-blockers can blunt the cardiovascular response to heat. Anticholinergics can reduce sweating. Psychotropic medications may affect thermoregulation, cognition, sedation, or a person’s ability to recognize and respond to danger. None of those facts requires exotic data collection. They are ordinary medication and problem-list data, although they are often messy, duplicated, outdated, or scattered across encounters.

A practical model would not merely count diagnoses. It would combine multiple weak-to-moderate signals into an actionable risk tier: comorbidities, medications, age, prior ED or inpatient use, mobility limitations if documented, home air-conditioning status if available, and neighborhood-level vulnerability. The output should be built for a nurse, care manager, primary care practice, or population health team that has to decide whom to contact first when the forecast shows several dangerous days ahead.

The hard part is that each of these variables has a different reliability profile. Medication data may be present but stale. Air-conditioning status may be absent. Social needs may be documented only after a crisis. Race, language, disability, and housing insecurity may be incompletely captured or captured in ways that are not fit for automated action. A model can still be useful under those conditions, but the people deploying it need to know which signals are strong, which are proxies, and which are missing because the patient has had less access to care.

The Mass General Brigham pilot shows the right shape of the intervention

The most concrete example to date is Mass General Brigham’s work with IBM on an AI system intended to identify patients vulnerable to extreme heat using EHR data, including health conditions and medications, and then send personalized warnings with cooling-center information through a mobile app. Mass General Brigham serves 2.5 million patients, giving the pilot enough health-system scale to matter operationally.[2]

It is important to keep the tense straight. This is a pilot in development, not published proof that AI-directed cooling-center referrals have prevented deaths during real heat events. Its value, as of Q3 2026, is that it demonstrates a plausible operational direction: use the EHR to create a patient-level heat vulnerability list, connect that list to patient communication channels, and include concrete cooling-center information rather than generic advice to “stay cool.”

That direction is much more promising than a standalone public dashboard. A dashboard can tell an agency where heat is coming. A clinical workflow can tell a population health team which patients to call, which primary care panels need attention, which messages should be sent before the hottest hours arrive, and which patients may need human outreach rather than an app notification. If the model produces only a score and no one owns the next step, it has not yet become a heat-health intervention.

The referral has to be more specific than “find a cool place”

A useful message to a high-risk patient should answer the questions that determine whether action happens: where the nearest cooling center is, when it is open, whether transportation is available, whether the patient can bring a caregiver, whether medications or medical devices create special needs, and whom to call if symptoms begin. AI can help choose the recipients. It cannot assume that the recipient can convert a warning into a ride, a plan, and several hours in a cooler room.

For population health teams, this changes the measure of success. The first metric is not model elegance. It is whether the risk list can be worked before the heat wave peaks. If a clinic receives thousands of names on the morning of a dangerous heat day, the list may be statistically impressive and operationally unusable. A smaller, tiered list with clear escalation rules may protect more people than a broader list that no one can finish.

Social vulnerability is not an optional add-on

An EHR-centered approach can become too neat if it treats heat risk as though it lives only inside diagnosis codes and medication lists. Heat emergencies also have geography. In a county-level analysis of heat-related EMS activations, Ramesh and colleagues found that counties in the highest Social Vulnerability Index quintile had 1.64 times the adjusted odds of substantially high heat-related EMS activation, defined as more than 200% of the national average; the 95% confidence interval was 1.14 to 2.37. Heat-related emergencies were concentrated in the South, Midwest, and Southwest.[3]

That finding does not prove that adding Social Vulnerability Index data to an EHR model will automatically improve outcomes. It does show that community conditions are part of the emergency pattern. Poverty, housing quality, transportation access, social isolation, language barriers, and local infrastructure shape whether a patient can act on advice. A health system that ignores those factors may identify medically fragile patients while missing the reason they remain in a dangerous apartment.

The equity problem cuts both ways. EHR models can help find patients whose clinical profiles are invisible to citywide messaging. They can also miss people who are poorly connected to care: uninsured patients, people who move frequently, people who use emergency departments without stable primary care follow-up, and residents who distrust health-system outreach. Social vulnerability mapping can partially compensate for that blind spot, but it is not a substitute for community partnerships or up-to-date local knowledge.

What current AI evidence can and cannot prove

There is adjacent evidence that machine learning can discriminate severe risk in heat illness once patients are already in the hospital. Models have shown high discriminative performance for predicting mortality among hospitalized patients with heat-related illness.[5] That is encouraging for clinical AI generally, but it is not the same task as preventing outpatient heat injury through cooling-center referral.

The outpatient task is messier. The model must run before a patient presents. It must use imperfect historical data. It must decide whom to contact without overwhelming staff. It must trigger a message, call, home-health contact, transportation offer, or community handoff. Then the patient has to receive the message, believe it, understand it, have somewhere to go, and be able to get there during the hours of risk.

This is why adoption should not be confused with effectiveness. A health system can deploy an AI heat-risk score and still fail to increase cooling-center use among the patients who need it most. Conversely, a modest model embedded in a disciplined workflow may outperform a more sophisticated model that produces a list no one can operationalize.

From risk list to cooled room

The practical chain is short on paper and difficult in real life: identify high-risk patients, contact them before the dangerous window, provide cooling-center information, solve the barriers that keep them from using it, and confirm that the intervention happened. Each link has an owner. The model owner may sit in informatics. The outreach owner may be population health. The transportation solution may require a city agency or community organization. The patient bears the consequence if those handoffs fail.

A workable program needs rules for escalation. A low-risk patient might receive an automated portal or text message. A patient with multiple high-risk medications, renal disease, and no documented air conditioning may need a phone call. A patient with limited mobility may need transportation coordination. A patient who does not answer may need outreach through a caregiver, home health agency, primary care practice, or community partner. These are not refinements after the AI work is done; they are the intervention.

Cooling centers also have their own constraints. Hours may not match the hottest or most dangerous periods. Locations may be hard to reach. Patients may not know whether they are welcome, whether the site is safe, or whether their medical needs can be accommodated. Public health messaging often assumes that information produces action. Heat response for clinically vulnerable patients has to assume that information is only the first barrier removed.

AI-driven EHR stratification can make cooling-center referrals far more clinically meaningful than population-level advisories by shrinking the outreach universe to patients whose bodies, medications, homes, and neighborhoods make heat especially dangerous. The strongest real-world implementation remains pilot-stage as of Q3 2026, and the central test is still ahead: whether a risk score can be converted into completed cooling interventions for the people least likely to be reached by ordinary warnings.

References

1. The Role of Cooling Centers in Protecting Vulnerable Individuals from Extreme Heat. Epidemiology. 2022.
2. AI Could Help Save Patients from Extreme Heat. Scientific American. July 2025.
3. County-Level Disparities in Heat-Related Emergencies. JAMA Network Open. 2024.
4. A comprehensive study of health impacts of heatwaves and AI-based preparedness for sustainable healthcare. Discover Public Health. 2026.
5. Machine Learning Prediction of Mortality for Patients Diagnosed With Heat-Related Illness. 2021.