📄 public and global health

Integrating respiratory infection surveillance and temperature data improves all-season mortality reconstruction

This study demonstrates that a hybrid neural-network model integrating high-resolution temperature data with respiratory infection surveillance significantly improves the accuracy of daily all-cause mortality reconstruction in Germany compared to models relying solely on weather or infection data, while also revealing that temperature-only models may overestimate the mortality risk attributed to cold exposure by failing to account for overlapping winter infection effects.

Original authors: Wang, J., an der Heiden, M., Irrgang, C.

Published 2026-08-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wang, J., an der Heiden, M., Irrgang, C.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Every year, millions of people die, and public health officials need to understand why. Is it the cold winter air, a spreading virus, or something else entirely? For decades, scientists have tried to sort out these causes, but the task is notoriously difficult. In many places, death certificates list a single underlying cause, yet a person might die from a heart attack triggered by a flu infection on a freezing day. When researchers look only at the weather, they often assume that the spike in deaths during winter is caused by the cold. When they look only at viruses, they might miss how the cold air weakens the body's defenses. This confusion makes it hard to plan for the future, especially as the climate changes and new diseases emerge. To make better decisions about hospitals, vaccines, and heat warnings, we need a clearer picture of how temperature and infection work together to affect human life.

A team of researchers in Germany has built a new way to untangle these overlapping effects. They created a sophisticated computer model that acts like a high-resolution microscope for daily death counts across the country. Instead of guessing which factor is responsible, the model looks at three distinct streams of information at once: the daily temperature and humidity, the number of people hospitalized with severe respiratory infections, and the underlying population trends. By feeding data from 2014 through 2025 into this system, the researchers could reconstruct the daily death toll with remarkable precision. Their goal was not just to predict the future, but to understand the past: to see exactly how much of the winter death spike comes from the cold, how much comes from the flu, and how much comes from other, slower-moving changes in the population.

The results of this study reveal that looking at temperature alone tells an incomplete story. When the researchers compared their new, all-in-one model against older models that only looked at the weather, they found a significant difference in accuracy. The old weather-only models made large errors, missing the mark by an average of 170 deaths per day. The new hybrid model, which included the hospitalization data for severe respiratory infections, reduced that error to just 80 deaths per day. This improvement suggests that the old models were mistakenly blaming the cold for deaths that were actually caused by viruses. When the new model accounted for the presence of respiratory infections, the amount of risk it assigned to cold temperatures dropped. This implies that in previous years, scientists may have overestimated the danger of the cold simply because they did not have a way to separate the virus from the weather.

The study also provided a detailed look at who is most at risk and when. The researchers found that the vast majority of deaths linked to either extreme temperatures or respiratory infections occurred in people over the age of 70. While older women often make up a larger share of the very oldest population, the study showed that when you look at specific five-year age groups, men generally face higher risks than women of the same age. This finding challenges the common assumption that older women are uniformly the most vulnerable group. The model also tracked the unique patterns of the pandemic years. During the height of the pandemic, when people stayed home and wore masks, the model detected a sharp drop in non-COVID respiratory infections, which corresponded with a quieter winter death toll. As restrictions lifted, the model captured the return of seasonal viral peaks, showing how closely the virus and the weather are intertwined.

One of the most important contributions of this work is how it handles the "baseline" of daily deaths. Usually, scientists try to find a quiet period in the year with no extreme weather or viruses to use as a reference point. However, the researchers realized that such a perfect quiet period rarely exists anymore. Instead, they built a system that constantly updates its understanding of the background death rate as it learns from the data. This adaptive approach allowed them to strip away the short-term spikes caused by heatwaves or flu seasons, leaving a clearer view of the long-term trends in population health. The model confirmed that while heatwaves cause sudden, intense spikes in deaths during the summer, the cumulative toll of cold weather and respiratory infections over the winter is far larger.

The researchers are careful to note that their model does not prove that a specific virus killed a specific person. Instead, it shows how much of the total death count can be explained by the presence of these factors. It is a tool for understanding the big picture, not for assigning blame in individual cases. By combining weather data with hospital records, the study offers a more honest accounting of why people die in different seasons. This clarity is vital for public health planning. If we know that a significant portion of winter deaths is driven by infections rather than just the cold, we can focus our resources on better surveillance, vaccination, and hospital preparedness. As the climate continues to change and new respiratory threats emerge, having a system that can distinguish between the weather and the virus will be essential for keeping communities safe.

Technical Summary: Integrating Respiratory Infection Surveillance and Temperature Data for All-Season Mortality Reconstruction

Problem Statement
Accurate all-season mortality surveillance is critical for public health planning, particularly under climate change and pandemic scenarios. However, attributing mortality variation to specific drivers is complicated by the overlapping effects of temperature, respiratory infections, and non-environmental factors. Traditional statistical frameworks, such as Generalized Additive Models (GAMs) and Distributed Lag Non-linear Models (DLNMs), effectively quantify temperature-associated mortality but face two primary limitations:

Humidity and Cold Attribution: They often omit or simplify humidity, despite its role in modifying heat stress and cold-associated risk. Furthermore, models relying solely on temperature may overestimate cold-associated mortality by attributing winter peaks driven by respiratory infections to low temperatures.
Baseline Estimation: Conventional baselines often rely on reference periods that are not truly exposure-free, leading to contamination in all-season analyses where both winter and summer contain significant exposure-related mortality.
Data Gaps: While influenza surveillance exists, nationwide inpatient surveillance for Severe Acute Respiratory Infections (SARI)—which correlates more closely with mortality than outpatient reports—has historically been limited in temporal depth and spatial resolution.

Methodology
The authors developed a component-based neural-network model to reconstruct daily all-cause mortality in Germany from 2014 to 2025. The model integrates four distinct data streams:

Mortality Data: Daily and weekly all-cause death counts from the Federal Statistical Office (Destatis), stratified by federal state, sex, and age.
Environmental Data: High-resolution (1 km) daily mean temperature and dew-point temperature from the HOSTRADA dataset (German Weather Service).
Infection Data: National weekly hospitalization incidence for non-COVID SARI and COVID-19-associated SARI from the Robert Koch Institute (RKI).
Demographics: Population counts stratified by age and sex.

Model Architecture
The model estimates daily death counts ( $m$ ) by multiplying the population ( $P$ ) by a corrected baseline mortality rate ( $b$ ) and a sum of relative-risk components:
$m = P \cdot b \cdot (1 + r_{temp} + r_{nC-SARI} + r_{COVID})$

Key architectural features include:

Adaptive Baseline: A baseline component ( $b$ ) that is not optimized via gradient descent but is periodically updated (every 50 epochs) using a smoothed residual of observed deaths minus modeled components. This allows the baseline to capture slow demographic shifts and background trends while separating them from short-term seasonal peaks.
Temperature Block: Uses 1-D convolutional layers to map daily mean temperature and dew point to a relative risk ( $r_{temp}$ ), accounting for lagged effects (up to 6 days) and non-linear relationships.
SARI Blocks: Separate convolutional blocks for non-COVID SARI ( $r_{nC-SARI}$ ) and COVID-19-associated SARI ( $r_{COVID}$ ). These blocks use non-negative weights to ensure attribution is non-negative and are time-invariant (no imposed seasonality).
Calibration: Multiplicative correction terms for day-of-week effects and state-level offsets (trained only on pre-2020 data to avoid pandemic distortion).
Training: The model is trained using AdamW with a Poisson deviance loss, augmented by a mean-alignment step to prevent drift in aggregate death counts.

Key Results

Reconstruction Performance: The hybrid model significantly outperformed single-input models. The hybrid model achieved a daily Root-Mean-Square Error (RMSE) of 80.2 deaths, compared to 170.2 for a weather-only model and 117.1 for an infection-only model.
Winter Attribution: Including respiratory infection indicators reduced the risk assigned to cold exposure. The weather-only model showed a steeper increase in mortality risk at low temperatures than the hybrid model, suggesting that temperature-only models partially misattribute infection-related winter mortality to cold.
Humidity Effects: Incorporating dew-point temperature improved the representation of hot periods. At 25°C, the model without humidity indicators predicted a 20.7% risk increase, whereas the hybrid model showed a range of 10.5% to 45.2% depending on relative humidity (40%–80%), highlighting the amplifying effect of humidity on heat stress.
Demographic Distribution: Model-attributed deaths were heavily concentrated in older age groups (>70 years). While absolute death counts were higher in women over 85 due to population size, age-stratified rates (5-year groups) revealed that men generally had higher mortality rates for both respiratory infection and temperature components in most age bands, challenging aggregation biases seen in broader age categories (e.g., 80+).
Baseline Trends: The adaptive baseline successfully isolated long-term trends, showing a general decline in background mortality with specific upward deviations in 2022–2023 for certain age groups post-pandemic.

Significance and Claims
The paper claims that integrating respiratory infection surveillance with environmental data allows for a more consistent reconstruction of all-cause mortality dynamics. The primary contribution is methodological: the framework demonstrates that environmental and infectious inputs explain complementary parts of mortality variation.

The authors explicitly state that their component magnitudes represent conditional model decompositions rather than causal estimates of deaths caused by SARI or temperature. They emphasize that the model cannot determine whether an infection caused, precipitated, or coincided with a death, as it is trained on all-cause mortality without individual-level infection status.

However, the study argues that this approach offers a practical alternative for periods where conventional baseline estimation is compromised (e.g., during the COVID-19 pandemic or severe winters). By separating winter infection peaks from the baseline, the model reduces the risk of misattributing these peaks to temperature or background trends. The authors suggest this framework provides a methodological basis for future integrated mortality surveillance, early-warning systems, and comparative analyses, provided that suitable surveillance data (such as SARI hospitalization) are available. They remain modest regarding the generalizability of the framework to settings lacking comprehensive SARI data.

More like this