Heat Does Not Reach the Emergency Department Equally.

The aggregate conceals the patient

The first heat episode of the German summer began on 18 June. The national median across a fixed panel of 20 weather stations remained at or above 20°C for 13 consecutive days and reached 28.35°C on 27 June. Three days later, the seven-day share of heat-associated presentations in the RKI emergency-department sentinel reached 4.28%.

That aggregate is real, but it is also the least interesting part of the result.

Among patients aged 80 or older, the corresponding share reached 12.19%. For ages 60–79 it peaked at 6.08%; for ages 20–39, at 1.83%. These are not counts or directly comparable risks across the German population, because each percentage uses presentations within its own age group as the denominator. Even so, one national temperature shock produces very different patterns inside participating emergency departments.

Timing matters because the temperature index peaked on 27 June, while most seven-day syndrome shares peaked between 28 June and 2 July. Part of that delay may be biological or operational; part is arithmetic, because a seven-day moving average cannot turn as quickly as daily weather. Accordingly, the chart establishes co-movement and age heterogeneity, not a precisely identified clinical lag.

Below the main panels sits a less dramatic but indispensable strip: the number of emergency departments contributing to the all-age series changes over time. A movement in the sentinel can reflect a real change in presentations, a change in which departments report, or both. Treating that strip as metadata would be a mistake; it is part of the outcome.

Heat is a curve, not a coefficient

A constant linear slope would imply that another degree matters equally on a cool day and during an extreme episode; the data reject that convenient simplification.

The response curves remain comparatively shallow through ordinary summer temperatures and bend upward in the hot tail. For the 80+ group, the fitted seven-day HEAT share rises from 4.57% at a national median of 18°C to 5.14% at 22°C, then accelerates further among the relatively few hotter observations. This level is already high partly because the syndrome includes volume depletion and hyponatremia. Each age-specific percentage also uses presentations within that age group as its denominator.

The small dots at the right edge are a warning against false precision. Extreme national-median temperatures are rare in a five-summer sample. LOESS is useful here because it lets the shape emerge without imposing a threshold, but it also becomes most fragile exactly where the curve is most interesting.

That is why the analysis keeps three thresholds rather than selecting the one with the most flattering result. Across 2022–22 August 2026 there are 171 days above 20°C, 68 above 22°C, and only 21 above 24°C in the national sentinel index. A steeper coefficient at 24°C may describe genuine nonlinearity, but it rests on much smaller support.

Age changes the cumulative response

The distributed-lag model asks a narrower question: after controls for annual seasonality and weekdays, plus trend and reporting-ED count, how does the age-specific seasonal excess move with heat degrees on the current and previous three days?

At the predeclared 20°C threshold, the cumulative four-day association is 0.21 percentage points per additional degree for all ages. It is 0.12 points for ages 20–39 and 0.29 for ages 60–79. For ages 80+, the estimate is 0.55 points; its 95% Newey–West interval runs from 0.25 to 0.86.

The cumulative estimate is the more credible quantity. Daily temperatures are highly correlated, and the outcome itself is a seven-day moving average. Asking the model to decide exactly which adjacent day deserves the effect invites it to distinguish shadows.

Nor is this a causal estimate: temperature is measured nationally, while the health outcome is aggregated across a changing set of participating departments. Air conditioning and urban form remain outside the model; so do local humidity, behavior and coding delays, along with patient composition. The regression is a disciplined description of timing and nonlinearity, not an experiment conducted by the weather.

More cases are not automatically congestion

The stronger operational claim would be that heat not only changes the case mix but slows emergency care. AKTIN publishes a separate weekly series with mean visits, mean length of stay, a 2017–2019 reference, and the number of reporting departments. That makes the hypothesis testable, but only approximately: the length-of-stay and syndrome sentinels are separate aggregates and need not contain the same hospitals.

Even after controls for mean visits and reporting-department count, as well as seasonality and trend, the direct temperature test remains null. The model associates one weekly heat-degree above 20°C with 0.03 additional minutes relative to the length-of-stay reference; the 95% interval runs from −0.10 to +0.17 minutes.

Weeks with a larger RKI HEAT excess do coincide with longer stays. The coefficient is positive and its interval excludes zero in this specification, but that result should not be promoted over the temperature test. Both outcomes can move with an omitted summer factor, and their participating departments are not linked. The safest conclusion is asymmetric: this prototype finds clear age-specific movement in the syndrome share, but the weather coefficient does not establish that heat slowed throughput.

That distinction matters for capacity planning because a changing case mix can demand different clinical attention without moving the average time spent in the department. Conversely, a length-of-stay shock can arise from staffing, bed availability, or discharge constraints without heat being the cause. Health burden and system congestion are related questions, not synonyms.

What the sentinel can say

RKI data do not estimate Germany’s rate of heat illness. Participation is selected rather than representative, and reporting availability changes. Coding can arrive with delay, while the HEAT definition combines several ICD-10 diagnoses rather than observing a single confirmed condition. Each published share also uses only presentations with sufficiently complete syndrome-defining fields in its denominator.

The defensible statement is narrower: among participating German emergency departments, the share of presentations classified as heat-associated rises with a national temperature index, and the movement is markedly larger among older age groups.

This finding survives three transparent temperature thresholds and a seasonal reference built only from prior summers. Uncertainty grows in the nonlinear tail because extreme days are scarce. The congestion extension remains inconclusive because weather does not recover a stable length-of-stay penalty and the two sentinels cannot be linked at hospital level.

This is still useful evidence: a national average temperature can conceal substantial heterogeneity in who appears at the emergency department. The next step is not a more elaborate national curve but a hospital-linked design. It should pair local weather with stable participation and matched case counts and denominators; humidity, nighttime temperature, coding-lag controls, and throughput measures should come from the same departments.

The aggregate has done its job by showing where not to stop.

Methods and caveats

The weather index uses a fixed 20-station DWD panel selected from a four-by-five geographic grid before analyzing the RKI outcome. Stations had to report from the start of 2022 through 22 August 2026 and contain quality-controlled historical data through 2025. They also had to provide a recent 2026 feed and sit below 500 meters. The daily index is the unweighted median of station mean temperatures, with at least 19 stations available on every day. DWD historical files provide completed quality control through 2025; 2026 comes from the recent directory and is not yet fully quality controlled.

The health outcome filters the RKI syndrome table to HEAT and ed_type = all, including 00+ and all eight published age groups. Its seasonal reference is the median seven-day HEAT share in 2022–2025 within ±7 calendar days of each day of year. Seasonal excess is the observed seven-day share minus that reference in percentage points.

Response curves use local linear LOESS with a 0.30 span; distributed-lag models use heat-degree terms for days 0–3 and two annual harmonics. They add weekday indicators, a linear trend, and reporting-ED count; Newey–West uncertainty uses 14 daily lags. Weekly congestion models use heat burden or aggregate HEAT excess plus mean visits and reporting-ED count, adding two annual harmonics and trend. Newey–West uncertainty uses four weekly lags.

All figures derive deterministically from pinned RKI releases and DWD archives; raw sources remain separate from generated public data. Publication manifests record provenance, licensing, sizes, and SHA-256 hashes.

Sources

• Robert Koch Institute and AKTIN: Emergency-department surveillance, release dated 23 August 2026, CC BY 4.0.
• AKTIN and Robert Koch Institute: Length of stay in emergency departments, release dated 20 August 2026, CC BY 4.0.
• Deutscher Wetterdienst Climate Data Center: Daily climate observations, historical and recent station archives.
• Heimrath et al. (2026): Validation of the heat-associated emergency-presentation syndrome.
• Download the complete derived analysis and publication manifest.