Forecasting heat deaths two weeks ahead worked better on mild days than extreme ones

European heat warning systems forecast temperature and leave the health inference to the reader. A study published on 6 October tests a different architecture: forecasting the deaths directly, up to two weeks ahead [s1].

The result is useful, and the pattern inside it is not intuitive.

What the study did

The authors applied a statistical epidemiological framework to temperature forecasts extending up to two weeks in advance, producing daily forecasts of heat-related mortality [s1]. They then evaluated forecast skill against two recent and exceptional European summers — 2022 and 2023 — and assessed how that skill related to temperature [s1].

Those two summers are the natural test case. Roughly 60,000 heat-related deaths occurred in Europe in the summer of 2022 and roughly 50,000 in 2023 [s1]. The paper frames Europe as a heatwave hotspot where numerous temperature records have been broken in recent summers, and notes that summers like 2022 are projected to become the new norm [s1].

The skill pattern

For most of Europe, forecasts were more skilful at milder temperatures — those close to the minimum mortality temperature — than at high ones [s1].

That is the opposite of what a warning system needs, taken at face value. The minimum mortality temperature is the point on a city's temperature-mortality curve where deaths are lowest; it is the comfortable middle of the distribution, not the dangerous end. A forecast that performs best there is performing best where the stakes are lowest.

But the paper reports a second pattern that changes the conclusion. Some of the hottest regions in Europe instead showed enhanced forecast skill associated with higher temperatures [s1].

That is the finding that makes the approach operationally interesting: forecast skill is highest at high temperatures precisely in the places where high temperatures kill the most people. The authors' conclusion follows from that — heat-related mortality forecasts can provide valuable information in European regions associated with high levels of heat-related mortality, and they advocate for local health authorities to incorporate such forecasts into heat warning systems [s1].

Why skill would behave this way

The paper does not attribute the pattern to a single cause, and neither should a reader. But two mechanisms are worth naming as plausible rather than established.

Statistical forecasts are easiest where the underlying relationship is well sampled. Near the minimum mortality temperature, a European city has decades of days to fit the curve on. Far into the heat tail, it has very few — and the exposure-response relationship there is estimated from a handful of extreme events, with wide uncertainty. In southern European cities where extreme heat is routine, that data sparsity problem is smaller, which would explain why skill holds up at high temperatures there and not elsewhere.

There is also a variance argument. Days near the minimum mortality temperature produce mortality close to baseline, which is stable and therefore predictable. Extreme heat days produce mortality that depends on duration, night-time relief, humidity, and whether the population has been primed by earlier heat in the season — all of which add variance that a temperature-driven forecast cannot resolve.

The clinical companion

A review published on 29 October covers the biology that the forecasting model treats as a black box [s2].

It summarises the mechanisms by which heat exposure induces systemic and organ-specific damage, arguing that understanding of these mechanisms remains insufficient and that this is hindering the development of targeted interventions for heat-related illness [s2]. The mechanisms it describes drive shifts in disease profiles toward acute heat-related illness, cardiovascular disease, renal disorders and other conditions, particularly in vulnerable groups [s2].

Two of its framings matter for how a mortality forecast should be used. First, susceptibility to heat spans the entire life course, from prenatal stages to old age, and is amplified by socioeconomic disparities [s2]. Second, the review advocates integrating heat exposure into clinical practice guidelines [s2].

Together those point at what a two-week heat-mortality forecast is actually for. It is not a public alert — the public does not act on a probabilistic death count. It is a health-system input: staffing, bed capacity, outreach to people on medications that impair thermoregulation, and pre-positioning of care for the conditions the review identifies as the actual causes of death during heat.

What the study does not establish

The forecasts were evaluated retrospectively against two summers [s1]. Retrospective skill on known events is a weaker claim than prospective performance, because model choices can be informed, even inadvertently, by what happened.

Two summers is also a small evaluation set, and both were exceptional [s1]. Skill measured on 2022 and 2023 may not transfer to an ordinary summer, and the study makes no claim that it does.

Most importantly, forecast skill is not the same as prevented deaths. Nothing in the paper measures whether issuing a heat-mortality forecast changes what health authorities do, or whether what they do changes outcomes. The gap between a good forecast and an averted death runs through dissemination, institutional response and individual behaviour, none of which was studied.

The underlying mortality models also rest on temperature-mortality relationships estimated from historical data — and as the BMJ review notes, susceptibility is modified by socioeconomic factors [s2] that shift over time and are not usually in the model.

What to watch

The operational test is whether any European public health agency adopts mortality-based rather than temperature-based triggers, and whether the regions where skill was highest are the ones that do. The authors' recommendation is directed at local health authorities [s1]; the next summer is when it becomes visible whether any of them acted on it.

Sources

• [s1] Skillful heat-related mortality forecasting during recent deadly European summers — Proceedings of the National Academy of Sciences, published 6 October 2025. https://doi.org/10.1073/pnas.2426516122
• [s2] Effects of extreme heat on physiology, morbidity, and mortality under climate change: mechanisms and clinical implications — The BMJ, published 29 October 2025. https://doi.org/10.1136/bmj-2025-084675