AI heat-health early warning systems combine machine-learning weather forecasts with health data to predict when and where heat will cause illness and death, often a week or more ahead. The forecasts are now good enough to act on. Lives are saved only when warnings trigger concrete local actions, such as checking on isolated older people, cooling spaces and hospital preparation.
In the summer of 2024, heat killed an estimated 62,775 people in Europe, according to a study by the Barcelona Institute for Global Health published in Nature Medicine. Across the three brutal summers from 2022 to 2024, the total was more than 181,000. Europe has some of the best weather services, richest health systems and most detailed heat plans in the world.
Most of those deaths were not sudden collapses in the street. They were older people, often living alone, whose hearts and kidneys failed over days in a hot flat. The heat that killed them was forecast. Many of those deaths were, in principle, preventable.
Heat-health forecasting has improved dramatically in the last few years, partly thanks to AI. But the lesson of every major heatwave is the same: heat-health warnings fail at the last hundred metres, not the first hundred kilometres. The forecast reaches the city. It does not reach the woman on the fourth floor with no air conditioning. AI makes the forecast better. It does not knock on the door.
How is AI improving heat forecasts?
Weather forecasting has been transformed by machine learning in a remarkably short time. In 2023, Google DeepMind’s GraphCast, described in Science, outperformed the European Centre for Medium-Range Weather Forecasts’ flagship deterministic model on most of the verification targets it was tested on, while running in a fraction of the time. In February 2025, ECMWF itself put its Artificial Intelligence Forecasting System, known as AIFS, into operation alongside its traditional physics-based model. ECMWF reported gains of up to 20% on some measures, such as tropical cyclone tracks, and roughly a thousandfold reduction in the energy needed to produce a forecast.
Cheaper, faster forecasts matter for heat because the useful unit is local. Heat risk depends on the specific neighbourhood: dense concrete, few trees and poor ventilation can make one district several degrees hotter than another a short distance away. AI models make it more feasible to downscale forecasts to that resolution and to run many scenarios.
Can we predict heat deaths, not just heat?
The newer and more important step is forecasting health impact directly. A companion study in Nature Medicine, also from the Barcelona team, showed that heat-health emergencies across Europe could be reliably forecast at least one week in advance. Instead of simply warning that temperatures will reach 38°C, such systems estimate how much mortality risk a given region faces, based on how its population has responded to heat in the past.
That distinction matters. The same temperature is far more dangerous early in summer, before people have adapted, or in a city where heat is unusual. A health-based forecast tells officials where the deaths are likely to be, not just where the thermometer will be highest.
This is where AI can add real value, by combining weather forecasts, historical mortality, population age structure, housing data and even real-time emergency calls to estimate risk in near real time.
But look at what the forecast does not do. It does not tell anyone who is at risk inside the region. It does not identify the man with heart failure taking diuretics, living alone on a top floor. It does not open a cooling centre or change hospital staffing. Those require people, lists and plans.
What actually saves lives in a heatwave?
The evidence on what works in heat-health action is surprisingly consistent, and most of it is low-tech.
Knowing who is vulnerable before the heat arrives. After the August 2003 heatwave, which killed around 15,000 people in France, the country created a national heat plan that includes municipal registers of isolated and vulnerable people who can be phoned or visited when an alert is issued. The idea is simple: when the heat comes, it is too late to find out who lives alone.
Human contact. Research on the 1995 Chicago heatwave found that social isolation was a major factor in who died. Phone calls and visits from neighbours, volunteers and health workers remain among the most effective interventions, because they catch people who do not recognise their own risk.
Clear, early, practical messages. Warnings that tell people what to do, such as drinking water, checking on neighbours, adjusting medication with advice and moving to cooler spaces, work better than warnings that simply state a temperature.
Health system preparation. Hospitals and primary care need to adjust staffing, review patients on medications that increase heat risk, and prepare for surges in admissions.
The best-studied example in a lower-income setting is Ahmedabad in India, which launched South Asia’s first heat action plan in 2013 after a deadly heatwave in 2010. The plan combines early warnings with public messaging, training of health workers, cool roofing and coordination across agencies. A 2018 study by Jeremy Hess and colleagues estimated that the plan was associated with about 1,190 fewer deaths a year in 2014 and 2015 compared with a pre-plan baseline.
Ahmedabad’s plan did not depend on a better forecast. It depended on a warning that set off a chain of named actions by named agencies.
A worked example: the forecast and the fourth floor
This is a hypothetical scenario, built to show where warnings succeed and fail.
A mid-sized city receives a health-based heat forecast eight days before a heatwave. The AI system estimates a sharp rise in mortality risk in three older districts with dense housing and few green spaces.
In version one, the city issues a press release and posts on social media. The local hospital hears about it on the news. Nothing else changes. The heatwave comes. In the three high-risk districts, older people living alone do not see the social media posts, and some do not consider themselves at risk. Emergency calls spike on day three. The death toll is roughly what the forecast predicted. The forecast was accurate and useless.
In version two, the forecast triggers a pre-agreed plan. The city’s social services pull the register of isolated older people in the three districts, and volunteers begin phone calls five days out. Primary care clinics review patients on diuretics and other heat-sensitive medications. Community centres in the target districts extend opening hours as cooling spaces. On the first hot day, volunteers visit people who did not answer the phone. The hospital adds staff to its emergency department.
The forecast was identical in both versions. The difference was a register, a phone tree and a plan that turned a probability into a list of doors to knock on.
The strongest counterargument: doesn’t better forecasting enable better action?
A fair critic would say this is a false contrast. A forecast eight days out instead of three gives time to organise volunteers, open cooling centres and adjust staffing. Neighbourhood-level forecasts let cities focus scarce resources on the districts most at risk, instead of spreading them thinly. Health-based forecasts are more persuasive to decision-makers than temperature warnings. Far from being a distraction, better forecasting makes door-knocking possible at scale.
This is largely right, and it is why investment in AI forecasting is worthwhile. Lead time and precision are genuinely valuable.
But they are valuable only to a system ready to use them. A city without a register of vulnerable people cannot use eight days of lead time to call them. A health system with no heat protocol cannot use a mortality forecast to change staffing. The critic’s argument is true conditionally, and the condition is exactly what most places lack. As our earlier piece on why predicting hospital capacity is the easy part argued, the forecast is the easy half of the problem.
What this means for health ministries, cities and developers
For health ministries, the first step is to treat heat as a health emergency with a health plan, not only a weather event. That means heat-health action plans with named responsibilities, registers of vulnerable people maintained year-round, and protocols for primary care and hospitals. Our piece on how heat waves kill through the heart makes the case for targeting warnings at cardiovascular and kidney risk, not only at heatstroke.
For cities, the investment case is in the last hundred metres: volunteer networks, phone trees, cooling spaces within walking distance, and data sharing between social services and health services. AI forecasts can tell cities where and when to mobilise, but only if there is something to mobilise.
For national meteorological services and AI developers, the priority is making forecasts usable. That means health-based outputs at neighbourhood scale, delivered in formats that trigger plans automatically, with honest uncertainty. The goal should be a forecast that a social services manager can act on at 7 a.m., not one that impresses at a conference.
For donors and development banks, heat is a growing health threat in rapidly urbanising, low-income cities where air conditioning is rare. Fund heat-health action plans alongside forecasting systems, because climate policy is health policy. The Ahmedabad model shows that plans can work in lower-income settings. The forecasting part is increasingly cheap. The organising part never will be, and that is where money is needed.
Frequently asked questions
What is a heat-health warning system?
A heat-health warning system links weather forecasts to public health action. When forecast heat is expected to pose a health risk, it triggers warnings to the public and to health and social services, and activates measures such as checking on vulnerable people, opening cooling spaces and preparing hospitals. The best systems base warnings on health risk thresholds, not temperature alone, and specify who does what.
How far ahead can heat-related deaths be predicted?
A study published in Nature Medicine found that heat-health emergencies in Europe could be reliably forecast at least one week in advance. That gives time to prepare, but it forecasts risk across regions and populations, not for individuals. Identifying and reaching the specific people at risk still depends on local registers, health records and outreach.
Who is most at risk of dying in a heatwave?
Older people, especially those living alone, face the highest risk, along with people with heart, lung or kidney disease, people taking certain medications such as diuretics, young children, pregnant women, outdoor workers and people without access to cooling. Social isolation is a major risk factor because it delays recognition and help. Anyone concerned about their own risk should talk to a clinician before hot weather arrives.
What is the Ahmedabad heat action plan?
Ahmedabad, India, launched South Asia’s first heat action plan in 2013 after a deadly heatwave in 2010. It combines early warnings with public awareness campaigns, health worker training, interagency coordination and measures such as cool roofs. A 2018 study estimated the plan was associated with about 1,190 fewer deaths each year in 2014 and 2015 compared with a baseline before the plan.
The takeaway
AI has given us heat forecasts that would have seemed extraordinary a decade ago. It has not given anyone a list of who lives alone on the fourth floor, or a volunteer to climb the stairs. The cities that stop heat deaths will be the ones that spend at least as much on the door-knock as on the forecast.
Does your city or health system know who its most heat-vulnerable residents are before the heatwave starts?
