Machine learning reveals drivers of cold-related illness during energy infrastructure attacks in Wartime Ukraine

Abstract

During the Russian invasion of Ukraine, targeted attacks on energy infrastructure exposed civilians to heightened cold-related health risks. This study aimed to: (1) characterize respiratory infections and cold-related injuries during the conflict; (2) identify vulnerable sociodemographic groups; (3) assess household adaptations; and (4) evaluate how winter preparation influenced health outcomes. We surveyed 2311 households across 24 Ukrainian oblasts during the winter of 2022–2023. One adult per household provided data on demographics, winter preparations, housing, heating, and access to services. Machine learning models were used to predict respiratory infections, symptoms, and cold injuries, based on sociodemographic and household factors. Respiratory infections affected 75.2% of participants, and 3.76% reported cold injuries, rising to 10% among older adults. Larger households experienced more respiratory infections, while areas under Russian control reported higher rates of cold injury. Key predictors of respiratory infections included age, household size, financial stability, and heating practices; cold injuries were predicted by age, region, anxiety, and household size. This is the first study to apply machine learning in examining the health impacts of cold-related events following energy infrastructure attacks in an active conflict zone. Our findings underscore the vulnerability of older adults and the widespread burden of respiratory infections, highlighting the need for targeted cold injury prevention in conflict-affected and cold-climate regions.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-32344-9.

Introduction

The Russian invasion that began in February of 2022 not only resulted in deaths and injuries but also impacted the lives and safety of Ukrainian civilians who experienced “wide area effects,” like shelling, multiple-launch rocket systems, artillery, missiles, and air strikes^1–3. The coverage highlights the challenges of dealing with the multifaceted nature of warfare, which affects both rural and urban areas through various battles. As the conflict progressed, Russian strategy evolved to include launching missiles, rockets, and loitering munitions, targeting energy facilities⁴. These facilities provide essential services, including heating and electricity, water supplies, domestic heating, hospitals, and other critical infrastructure, such as street lighting, communication systems, and banking systems. Russian attacks are estimated to have severely damaged 40% of Ukraine’s thermal generation, 90% of its wind power, and over 40% of its solar and nuclear energy sources, which were either under occupation or damaged⁴. These relentless assaults on the power grid have resulted in millions of Ukrainians experiencing shortages of electricity, heat, water, and other essential services⁴. While such disruptions pose an inconvenience in summer months, which typically have a humid continental climate with warm summers, they are most dangerous in winter, when temperatures are typically − 3 to − 6 °C, but can drop as low as − 20° or even − 30 °C⁵.

Conflict and forced displacement have long been linked to respiratory infections among displaced populations⁶. However, the deliberate destruction of civilian energy infrastructure during harsh winters presents an unprecedented threat, significantly heightening the risk of cold-related injuries and respiratory illnesses. To our knowledge, no prior studies have examined the impact of such targeted attacks on civilian health during a war, highlighting a critical gap in understanding the full extent of suffering inflicted in these conditions. Here, cold injury refers to tissue damage caused by exposure to cold⁷. However, it is not the only factor determining injury severity. Cold injury occurs when the body’s ability to adapt to low temperatures is compromised, influenced by factors such as exposure duration, humidity, wind, altitude, clothing, underlying medical conditions, and individual variability. Cold injuries can be pre-freezing (trench foot and chilblains/pernio) or freezing (hypothermia and frostbite) and can affect specific body parts or the entire body⁷. In this study, we refer to self-reported respiratory infections and symptoms, which encompass conditions affecting the nose, sinuses, throat, and lungs. As these reports were not clinically verified, we did not distinguish between specific causes such as viral infections or influenza.

Given the nature of an unprecedented attack on energy infrastructure during the winter season, it is imperative to understand the public health impact of cold injuries and the adaptation and mitigation strategies adopted by the affected population to reduce injuries and infections. Thus, there is a clear need to conduct a study on respiratory infections and symptoms, as well as cold injuries.

We aimed to identify the household and environmental features most predictive of these outcomes. Specifically, we sought to: (1) Characterize patterns of respiratory infections and cold-related injuries among Ukrainians during the Russian invasion by applying machine learning models to survey data. (2) Identify key sociodemographic factors associated with cold-related illness by examining how variables such as psychological distress and financial instability contribute to model-predicted importance. (3) Explore adaptation strategies (e.g., energy-saving behaviors, coping with displacement) and assess how these strategies appear in the models as protective factors against cold-related health outcomes. (4) Evaluate the role of winter preparation at the individual and household levels, using predictive modeling to highlight which preparedness features are most influential in the occurrence of adverse health outcomes, particularly among populations facing infrastructure failure and extreme cold.

Methodology

Data collection and preparation

TGM Research, a nationally representative survey panel in Ukraine, recruited participants from April 5, 2023, to May 15, 2023, via emails, in-app notifications, or text messages. Using a quota sampling design, a representative sample was obtained from Ukraine’s 26 administrative regions, comprising 24 oblasts and two special-status cities (Kyiv City and Sevastopol City)⁸. The online survey data were collected from one adult per household, aged 18–72, in 2311 households across 24 oblasts in Ukraine.

All methods were performed in accordance with the relevant guidelines and regulations.

Demographic data were collected on participants’ age, sex, status as an internally displaced person (IDP), education, marital status, and occupation (Tables S1, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14, S15, S16, S17). The questionnaire gathered data on respondents’ living and health conditions, including diseases such as respiratory infections and symptoms, cold injuries, and heart attacks (Tables S1, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14, S15, S16) during the winter season from December 2022 to February 2023. The dataset included health issues among all household members, including internally displaced persons living in the same household, children (< 18 years), older adults (> 60 years), and pregnant individuals within the household, along with the overall health status of the household. A final list of features is provided in the supplement (Excel data file).

Participants responded to 39 yes/no questions about whether they had experienced any of the following cold-related conditions during the Russian invasion. When they endorsed yes, they were asked, 'How many people in your home experienced the event?' and 'Do they have access to the medications or treatments of the diseases or symptoms?'. Self-reported respiratory infections and symptoms, including flu, pneumonia, rhinorrhea, cough, sneezing, fever, and sore throat, were combined due to the small number of positive responses. Participants reported experiencing hypothermia, trench foot (also known as immersion foot), chilblains, frostbite, frostnip, Raynaud’s disease, and corneal freezing. Due to the relatively small number of participants reporting hypothermia, trench or immersion foot, chilblain, frostbite, frostnip, Raynaud’s disease, and cornea freezing, we created a dichotomous cold injuries variable, indicating whether anyone in the household, including vulnerable members (such as children, pregnant women, and older adults), experienced any cold injuries during December 2022–February 2023.

Preprocessing

The collected data underwent extensive preprocessing to ensure quality and suitability for analysis. A total of 2311 survey responses were received. Among these, 183 (7.9%) contained placeholder values (e.g., “999” for unknown household size). Because responses were forced-choice, participants could not skip questions and were instructed to use “999” if unsure. For machine learning (ML) models, these placeholder values were replaced with − 1, while for statistical analyses, the corresponding surveys were censored.

Another five (0.2%) responses indicated household sizes ranging from 50 to 300, likely representing individuals in shelters. These were recoded as 55 for ML analyses. Due to data quality issues, only 17 (0.7%) surveys were excluded from the Chi-square and t-test analyses. Ultimately, 2111 (91.3%) valid responses were retained for statistical analysis and visualization (Figs. S1, S2, S3, S4).

The cleaned dataset included 140 independent variables (features) and one outcome variable. Feature engineering was performed to identify variables potentially associated with cold-related injuries. To simplify classification tasks, two binary outcomes were analyzed: (i) respiratory infections and symptoms, and (ii) cold injuries.

Statistical analysis

We first applied conventional statistical tests, including Chi-square and independent sample t-tests, to examine differences in demographic and household characteristics among respondents. However, the dataset’s high dimensionality (140 features across multiple subgroups, pregnant women, children, older adults, and other household members) limited the ability of traditional methods to capture nonlinear interactions and higher-order associations. To overcome these limitations, we applied supervised ML methods to identify key predictors of cold-related illnesses and energy-related health risks.

Machine learning models

To predict respiratory infections and cold injuries, and identify the most influential predictors, we used six ML algorithms: Decision Tree (DT), Random Forest (RF), XGBoost, AdaBoost, Support Vector Machine (SVM), and Multi-layer Perceptron (MLP)^9,10. ML was chosen over conventional statistical models because it can efficiently handle large, complex datasets with intricate inter-variable relationships¹¹. These models were selected to represent a diverse set of approaches: tree-based methods (DT, RF, XGBoost, AdaBoost) for their interpretability and ability to capture nonlinear relationships, and non-tree-based methods (SVM, MLP) for their capacity to model complex decision boundaries in high-dimensional spaces.

Ensemble methods, particularly RF, AdaBoost, and XGBoost, offer additional mechanisms that enhance robustness to correlated predictors and reduce overfitting. In RF, feature subsampling reduces inter-tree correlation by considering only a random subset of predictors at each split. AdaBoost and XGBoost apply iterative reweighting and boosting, sequentially combining weak learners to focus on harder-to-classify instances and smooth model variance. While DT, SVM, and MLP do not incorporate these specific mechanisms, they are widely used benchmarks that provide complementary perspectives on model performance and predictive patterns.

Model performance was assessed using multiple evaluation metrics to account for class imbalance within each outcome variable (common cold and cold injuries). While accuracy was calculated, it can be misleading for imbalanced datasets; thus, precision, recall, F1-score, and confusion matrices were also reported. Accuracy represents the proportion of correctly classified instances; precision indicates the proportion of true positives among predicted positives; recall (or sensitivity) measures true positives among all actual positives; and the F1-score, the harmonic mean of precision and recall, balances both metrics. All metrics were reported as percentages, with higher values indicating stronger performance.

Feature importance was computed using the mean decrease in impurity (MDI), allowing identification of variables that contributed most strongly to outcome prediction.

Feature selection

Feature selection focused on demographic, socioeconomic, environmental, and behavioral variables (Tables S1, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14, S15, S16, S17). The questionnaire gathered data on respondents’ living and health conditions, including diseases such as respiratory infections and symptoms, cold injuries, and heart attacks (Tables S1, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14, S15, S16). Importance values ranged from 0 to 1, with higher values indicating greater predictive significance. Features exceeding a threshold of 0.02 were retained. To enhance interpretability and biological plausibility, variables lacking environmental or physiological relevance were excluded. The iterative selection process improved model performance while minimizing redundancy and maintaining generalizability.

Model validation

Rigorous model validation was conducted using hyperparameter optimization and fivefold cross-validation. A grid search was employed to systematically explore hyperparameter combinations and identify the best-performing configurations, as measured primarily by the F1-score⁹.

During fivefold cross-validation, the dataset was divided into five subsets, with four used for training and one for validation, repeated five times to ensure stability. This procedure mitigated overfitting and yielded a more reliable estimate of generalization performance.

Despite a small number of positive samples for older adults (increasing overfitting risk), the larger sample size for household-level respiratory infections reduced this risk. Consistent high accuracy, precision, recall, and F1 scores across models and datasets (Tables 1, 2, 3), along with low standard deviations across folds, demonstrate the robustness, reliability, and generalizability of the final models. These results confirm that the ML models accurately predicted seasonal respiratory infections, symptoms, and cold injuries with high sensitivity and specificity across diverse household conditions.

Table 1.

Classification performance of various ML algorithms on seasonal respiratory infections and symptoms at the household level (all population).

Method	Accuracy		Precision		Recall		F1-score
	Mean	Standard error	Mean	Standard error	Mean	Standard error	Mean	Standard error
Decision tree	69.453	0.012	76.892	0.010	75.651	0.010	76.261	0.009
Random forest	73.519	0.005	76.456	0.004	85.523	0.007	80.725	0.004
XGBoost	74.167	0.006	76.497	0.006	86.922	0.010	81.353	0.005
AdaBoost	72.523	0.007	74.944	0.005	86.590	0.007	80.342	0.005
SVM	73.346	0.005	75.576	0.004	87.058	0.008	80.900	0.004
Neural network	70.185	0.006	76.439	0.010	78.524	0.025	77.274	0.009

Accuracy: Accuracy measures a classification model’s performance by calculating the ratio of correctly predicted instances to total cases. It provides an overall assessment of the model’s ability to classify data correctly.

Precision: Precision is a metric that evaluates how well a classification model identifies positive instances. It represents the ratio of correctly predicted positive cases to the total number of cases predicted as positive. Precision is essential when minimizing false positives is critical.

Recall: Recall measures how effectively a model captures all the relevant positive cases and is critical when the goal is to minimize the number of missed positive instances.

F1-score: The F1-score is a machine learning metric that assesses a classification model’s performance, especially in cases with imbalanced datasets, by combining precision and recall into a single measure.

The mean values represent the average performance of the model across the fivefold cross-validation, reported as percentages. The standard error (SE) indicates the variability of the performance measures across the different folds, with lower SE values indicating more consistent model performance. For more details, refer to the Methods section.

Table 2.

Classification performance of various ML algorithms on respiratory infections and symptoms among the elderly (> 60 years old) population.

Method	Accuracy		Precision		Recall		F1-score
	Mean	Standard error	Mean	Standard error	Mean	Standard error	Mean	Standard error
Decision tree	71.196	0.015	82.624	0.004	81.142	0.022	81.803	0.012
Random forest	80.478	0.004	81.827	0.001	97.254	0.005	88.873	0.003
XGBoost	81.053	0.002	82.789	0.002	96.421	0.003	89.085	0.001

Table 3.

Classification performance of various ML algorithms on cold injuries among the elderly (> 60 years old) population.

Method	Accuracy		Precision		Recall		F1-score
	Mean	Standard error	Mean	Standard error	Mean	Standard error	Mean	Standard error
Decision tree	69.282	0.013	30.177	0.028	26.551	0.030	27.960	0.026
Random forest	76.555	0.010	44.802	0.123	9.681	0.024	15.683	0.037
XGBoost	77.895	0.005	52.886	0.036	18.564	0.022	27.351	0.028
AdaBoost	76.746	0.005	47.656	0.023	26.596	0.025	33.876	0.025
Neural network	69.665	0.018	29.444	0.037	20.683	0.022	23.617	0.022

Results

Respiratory infections and symptoms were prevalent (75.2%), especially in larger households and among adults aged 30–44 (78%) or 60 + (80%) (Table S11, Figs. S1, S2). Cold injuries affected 3.76% of participants overall, but 10% of older adults (Tables S7, S8). Households with children or older adults had higher infection rates (Figs. S5, S6, S7). Six oblasts experienced an increase in the number of households since the start of the invasion (11.4–18.6%), while others saw a decline in household numbers (10–100%) (Table S3). Nearly 50% of participants lost their heating equipment during the invasion (Table S8), and 29% reported a lack of necessary medicine (Table S6). Most respondents from Russian-occupied areas moved after the invasion and tended to have higher cold injury rates compared to non-occupied areas (Fig. S4).

To evaluate the significance of independent features, other than the involvement of older adults and children in the household, in predicting respiratory infections and symptoms at the household level, considering all population groups, we experimented with predicting this target without using those two features as inputs. Table 1 presents the classification performance of different ML algorithms for this experiment. The evaluation metrics include accuracy, precision, recall, and F1-score, with results averaged over a fivefold cross-validation. AdaBoost outperformed all models, while XGBoost had the highest F1 score (Table 1). The study highlights the importance of feature selection and the impact of specific features on predictive performance. Figure 1 highlights the role of each input feature in predictive models, revealing the connections between these features and the occurrence of respiratory infections and symptoms in households. Results consistently indicated that age and household size were the most significant features in predicting seasonal respiratory infections and symptoms, as per DT, RF, XGBoost, and Adaboost models. However, the XGBoost model prioritized “household size” as the most crucial factor in predicting seasonal respiratory infections and symptoms, surpassing “age”. ML algorithms differed in key predictors, revealing varied patterns in household respiratory infection and symptom risks (Fig. 1). The number of true positives (correctly predicted occurrences of seasonal respiratory infections and symptoms) was high across all models, while false positives and true negatives were very close (Fig. S8).

Fig. 1.

The importance of each input feature in respiratory infections and symptoms prediction at the household level (all population) using various ML algorithms. (A) DT, (B) RF, (C) XGBoost, (D) Adaboost (feature name on the left side).

The algorithms, including RF and XGBoost, showed high accuracy and F1-score in predicting seasonal respiratory infections and symptoms among older adults (Table 2). Financial stability significantly influenced the likelihood of experiencing seasonal respiratory infections and symptoms, with concerns about the currency exchange rate, pension size, and low-income levels emerging as impactful features. In addition, "Anxiety (in older adults)" was identified as another influential feature shared between RF and XGBoost, highlighting its role in predicting seasonal respiratory infections and symptoms. DT and RF models highlighted the significance of energy-saving behaviors, such as showering without hot water and staying warm with clothes and blankets, in predicting seasonal respiratory infections and symptoms among older adults (Fig. 2). The models exhibit high sensitivity to seasonal respiratory infections and symptoms, with a high number of true positives and a relatively low number of true negatives (Fig. S9). Out of the 1045 households including older adults, only 207 had not experienced respiratory infections and symptoms (19.8%, Fig. S9).

Fig. 2.

The most important and plausible features in cold injury prediction in older adults (> 60 years) using various ML algorithms. (A) DT, (B) RF, (C) XGBoost.

XGBoost demonstrated superior accuracy and precision in predicting cold injuries among older adults, while AdaBoost achieved the highest F1 score (Table 3). Across DT, RF, and Adaboost models, “Age” emerged as the most crucial feature for predicting cold injuries in older adults, indicating its significant impact on the outcome. Additionally, the models highlighted the importance of “Oblast” (region) and “Household size” in predicting cold injuries. In contrast, XGBoost ranked "Anxiety (in older adults)" as the most influential feature, followed by "Pneumonia (in older adults)" and "Damp," indicating its unique ranking compared to other models (Fig. 3A–D). The study suggests high specificity in predicting cold injuries, consistent with high true negatives and low true positives in all models (Fig. S10). Only 10% of older adults experienced cold injuries. The presence of children or the elderly increased infections, driving 90.4% model accuracy (Table S18).

Fig. 3.

The most important and plausible features in respiratory infections and symptoms prediction in older adults (> 60 years) using various ML algorithms. (A) DT, (B) RF, (C) XGBoost, (D) Adaboost (feature name on the left side).

Discussion

In this study of Ukrainians surveyed during the Russian invasion, over 80% of respondents reported respiratory infections and symptoms, and 5% reported cold injuries in their households. The prevalence of cold injuries was 10% among the older adult population (> 60 years). Nearly 50% of respondents reported that their heating equipment had been destroyed during the crisis, and 13.7% faced difficulties obtaining wood or fuel for their homes due to severe winter conditions caused by the extensive damage to energy infrastructure from Russian attacks.

In war-frontline oblasts such as Cherkasy and Ivano-Frankivsk, respiratory infections and symptoms were positively associated with household size, likely due to mass refugee movements and poorer living conditions with limited access to medical care and necessities. Among participants in this sample, six Oblasts (Cherkasy, Ivano-Frankivsk, Poltava, Ternopil, Zakarpattia, Zhytomyr) saw an increase in households ranging from 11.4 to 18.6% during the invasion, while another six Oblasts (Donetsk, Kharkiv, Kherson, Kyiv, Luhansk, and Zaporizhia) decreased in total households ranging from 10 to 100%.

Considering the pre-Russian invasion^12,13, our findings suggested that 75.2% of participants were infected by respiratory infections and symptoms, meaning infections were most common among people aged 30–44 (78%) who worked outside of the home and also had a child or older adults in the household. This prevalence is substantially higher than typical baseline or wartime estimates, likely reflecting Ukraine’s cold-climate conditions and the severe disruption of heating and energy infrastructure during the invasion¹⁴. Household size was a key factor in seasonal respiratory infections and symptoms; overcrowded homes may result from hosting other families, potentially increasing the risk of infection transmission among residents. Different ML algorithms predicted seasonal respiratory infections and symptom occurrences in households differently, with all models generally predicting positive associations between these household characteristics and infections, especially for older adults.

Our findings suggest that concerns about currency exchange rates, pension size, and low income among older adults may have hindered the effectiveness of cold prevention measures, potentially increasing their risk of respiratory infections and symptoms. As suggested by the World Bank’s report¹⁵ and other studies¹⁶. Many direct and indirect associations can be made here: having a high pension means having access to money, insurance, and other resources to alleviate illnesses. It is also indicative of having greater stability¹⁵. Factors like food storage, furnace, and electric heater possession can reduce anxiety while showering without hot water¹⁷. Not wearing warm clothes and staying without blankets can increase respiratory infections and symptoms.

Our results showed a high incidence of cold injuries among older individuals from Donetsk, Kherson, Luhansk, and Zaporizhia during the Russian invasion. This may be attributable to failure to leave their homes despite the compromised energy grid, as reported elsewhere^18,19. It may also reflect relocation to avoid attacks during the cold, snowy winter season; prior literature demonstrates that older adults are particularly at risk of cold injuries due to decreased body temperature regulation and reduced heat energy production²⁰.

Our findings underscore the crucial role of housing conditions in shaping health outcomes during conflicts¹. Patients with cold injuries were more likely to live in poorly heated homes, particularly in oblasts near conflict zones. Published literature suggests that warmer indoor environments may reduce the risk of cold injuries and lower the risk of respiratory infections²¹. Another study suggests that cold and damp conditions promote the spread of viruses and bacteria, thereby increasing the risk of pneumonia²². Additionally, anxiety and poor living conditions may further exacerbate cold injuries in older adults^23,24. These results underscore the importance of improving access to housing and heating to enhance health resilience in conflict-affected regions.

Consistent with Hryhorczuk et al.²⁵ findings during the invasion in Ukraine, our findings highlight the impact of damaged heating infrastructure on cold injuries. The destruction of Ukraine’s fireplaces, heating systems, and water supply likely contributed to an increase in cold-related injuries. Our study reinforces the strong link between heating access and cold injury risk, suggesting that a functioning fireplace may be a crucial protective factor against freezing temperatures in affected households²⁵.

Using ML approaches enabled us to significantly improve the performance of AdaBoost and XGBoost, which substantially enhanced risk prediction for cold injuries and respiratory infections, allowing for more precise prevention strategies for vulnerable populations. XGBoost’s ability to handle structured medical data establishes it as a powerful tool for epidemiological modeling, while its feature importance rankings highlight key risk factors, such as temperature exposure and comorbidities. These insights optimize resource allocation and intervention efforts.

The study also leveraged iterative learning and built-in regularization methods to strike an optimal balance between bias and variance, a crucial consideration in real-world scenarios with imperfect data. The analysis confirms the superior performance of boosting techniques in the specific context by systematically evaluating multiple algorithms. It provides valuable benchmarks for selecting and fine-tuning models for similar problems. This comparative assessment underscores the significance of algorithmic selection and parameter optimization, offering insights to inform future research and practical applications in environments characterized by challenging data conditions.

The models generally predicted respiratory infections and symptoms with high true positive rates but also showed increased false positives due to inherent uncertainties. The high-performance metrics in older adults were significantly influenced by the high prevalence of seasonal respiratory infections and symptoms, which affected over 80% of individuals. The models showed high specificity in predicting cold injuries, with high true negatives and low true positives, suggesting a higher likelihood of negative outcomes in older adults. The study highlights the difficulty in predicting cold injuries in older adults, as evidenced by poor performance metrics and low recall scores across all models.

To the authors’ knowledge, this study is the first to investigate respiratory infections, symptoms, and cold injuries among non-military personnel in a war setting, using comprehensive questionnaires to gather detailed information. The study highlights the use of ML models for disease forecasting, enhancing interpretability, and facilitating the processing of large quantities of features^26,27. Recognizing cold-related injuries, illnesses, and symptoms is crucial for public health actions. Identifying at-risk population subgroups and their associated factors can help prevent or mitigate most adverse health impacts²⁸.

Although we strove to obtain a representative sample by age and sex, we encountered underrepresentation, which introduced a risk of sampling bias due to the smaller number of rural participants, potentially leading to an underestimation of local cold injuries. Reliance upon self-reported health data may introduce bias, potentially affecting the generalizability of the results. Medical diagnoses of cold injury were self-reported and could not be confirmed through medical records, which may have led to error, and indeed, the model had low sensitivity. Our analysis focused on the observed health outcomes during the war. Still, without a baseline comparison, it is impossible to establish a clear causal link between the war and the infections. Other unmeasured factors could also have contributed to the observed cold-related health concerns.

While the models performed reasonably well in terms of overall accuracy and F1-score for the prediction of respiratory infections in both the household level (Table 1) and the older adults (Table 2), the significant class imbalance in the dataset likely contributed to the low F1-score for the prediction of cold injuries in the older adults (Table 3). The model’s tendency to predict the majority class more often resulted in poor performance for the minority class.

Although fivefold cross-validation was used to evaluate model performance, the relatively small sample size in predictions done on the older adults, combined with a relatively high number of input features, increased the risk of overfitting. In contrast, predicting respiratory infections at the household level (the entire household population) with a larger sample size and fewer input features resulted in a lower risk of overfitting.

Future studies with more balanced and representative data will be crucial for validating and refining these findings, particularly to ensure their applicability to broader, more diverse populations.

Conclusion

Here, we identified risk factors that elevated the chances of respiratory infections, symptoms, and cold injuries within households among Ukrainians during the Russian invasion. Limited efforts have been made to raise awareness about risk factors, adaptation, and resilience among war-impacted populations, particularly families and social groups. Few studies have considered the public health impacts of energy infrastructure destruction during harsh winters and military invasions. This study highlights that all age groups are at risk, with household size, concerns about currency exchange, pension size, and marital status being key factors that predict respiratory infections and symptoms. For cold injuries, important features were households with residents over 60 and the presence of heating equipment. These findings can lead to the development of effective interventions and adaptation and mitigation strategies, such as those addressing cold injuries and respiratory infections, during an energy crisis in a harsh winter. Deliberate attacks on energy infrastructures during the winter season may impose further trauma, increase disability, and result in mass suffering during the current Russian invasion of Ukraine, as well as during other wars set in cold or seasonally variable climates.

Author contributions

Ubydul Haque, Shanshan Wang, and Emily S. Barrett designed the research; Emily S. Barrett arranged funding; Kimia Marvi and Ubydul Haque analyzed the data; Kimia Marvi and Ubydul Haque drafted the manuscript; Juan Espinoza created figures S1–S4, and Table S17. Iftikhar U Sikder, Nancy Fiedler, Julia Pavlova, and Irina Holovanova All authors contributed to the writing and editing and approved the final version of the manuscript.

Funding

This study was supported by the National Institute of Environmental Health Sciences Center (grant P30ES005022).

Data availability

Supporting data for this manuscript are included in the article and supplementary files. Original datasets are available from the corresponding author upon request.

Declarations

Conflict of interest

The authors declare that they have no conflict of interest. Juan Espinoza is a consultant for Sanofi, Dexcom, and Glooko.

Ethical approval

All participants provided informed consent online. The study was approved by the Human Research Ethics Committees of Poltava State Medical University, Ukraine (Protocol No. 212), and the Institutional Review Board of Rutgers University, USA.