AI Turns Weather Data Into Health Warnings, But a New Review Finds the Field Is Flying Blind

Artificial intelligence is quietly becoming one of the most powerful tools in environmental health research, capable of sifting through vast, tangled streams of weather data to predict who will fall ill when the atmosphere turns hostile. But a new rapid review published in the journal Air Quality, Atmosphere & Health suggests that while the promise is real, the field is riddled with methodological blind spots that could undermine the very predictions on which lives may one day depend. The review, led by researchers at the University of Southampton and the University of Oxford, synthesised studies published between 2020 and 2025 that applied artificial intelligence to meteorological data in the context of human health, and its findings reveal both a technology racing ahead and a reporting culture struggling to keep up.

The stakes could hardly be higher. Air pollution alone contributes to an estimated 4.2 million premature deaths each year, with the burden falling disproportionately on low- and middle-income countries. Mortality risk rises by 10 to 25 percent for every degree Celsius above regional temperature thresholds, and heat-related deaths among adults aged 65 and over have surged by 167 percent compared with the 1990s, a rise that exceeds what demographic ageing alone can explain. Across Europe, an estimated 368,183 heat-related deaths occurred between 2010 and 2022, 89 percent of them among people aged 65 or older. Children, whose thermoregulation is still immature, and people living with mental health conditions or long-term illness face elevated risks from heat, cold spells, and flooding alike.

What makes meteorological data so difficult for traditional statistics is that the atmosphere behaves as a unified physical-chemical continuum in which one variable drives another. During temperature inversions or atmospheric blocking events, suppressed boundary layer mixing can allow particulate matter to accumulate at the surface while ozone levels remain largely unaffected. Inhaled particles then travel through the airways, triggering systemic inflammation, oxidative stress from free radicals, and cardiac and neurological damage as toxins enter the bloodstream. These relationships are non-linear, spatio-temporally complex, and conditional on interactions between variables, which means the very signals that drive weather-related health risks can be excluded from conventional models such as generalised additive models that have historically defined climate-health research.

This is precisely where machine learning enters the picture. AI methods can automatically handle non-linear relationships and spatio-temporal structures without manual model specification, and hybrid models combining physical and data-driven approaches have demonstrated improved forecasting accuracy, better detection of extreme weather events, and reduced systematic biases compared with traditional numerical weather prediction. AI-assisted systems can also improve downscaling, resolving fine-scale atmospheric structures by balancing multi-scale constraints. In principle, this positions AI to map how specific cascading climate states trigger downstream morbidity and mortality through a domino effect, offering a granularity that classical statistics cannot match.

To understand how this potential is being realised, the Southampton team conducted a rapid review following PRISMA-RR guidelines, searching PubMed, Web of Science, and Scopus for peer-reviewed studies published between January 2020 and October 2025. Two reviewers independently screened 876 identified records, ultimately including twelve studies that applied AI tools to meteorological data in human health research. The team deliberately imposed no prespecified definition of artificial intelligence, including any article in which the term was explicitly stated, a decision that itself reflects how heterogeneously the label is applied across the literature.

The twelve studies were strikingly diverse. Eight were observational, with the remainder experimental, longitudinal, or retrospective cohort designs. Sample sizes ranged from just 12 participants to more than 10 million, and data collection periods varied from 20 minutes to 11 years. Six studies were conducted in high-income countries with temperate climates, while the rest took place in middle- to low-income settings with extreme climates, including Bangladesh, China, Kenya, Mali, and Vietnam, where weather-related health risks are generally more acute. The most common meteorological variables were air quality, including particulate matter, and temperature, each appearing in six studies, followed by humidity and rainfall, pressure and wind, and seasonality. Ground stations and air sensors were used in every study, and seven also incorporated satellite data.

Health outcomes clustered around respiratory disease such as asthma and chronic obstructive pulmonary disease, which appeared in six studies, and bacterial or viral illness such as diarrhoea, which appeared in four. The remainder examined heat-related illness, migraine and headaches, COVID-19, febrile illness including dengue, and cardiovascular disease. On the technical side, Random Forest and other decision-tree methods dominated, appearing in five studies, followed by time-series and clustering analyses, regression, gradient boosting, and multivariate or multilayer approaches. Model performance was most commonly evaluated using sensitivity, though four studies specified no performance metrics at all, and metrics such as area under the curve, specificity, and predictive capability appeared inconsistently across the set.

The headline finding was encouraging: in study after study, incorporating meteorological variables improved the predictive value of the models. Random Forest analyses showed that adding air pollution and seasonality sharpened predictions of increased prevalence and severity of diarrhoea, dyspnoea, COVID-19, asthma, and pneumonia. Time-series and gradient boosting approaches found that temperature, humidity, air pressure, wildfire smoke, and seasonality similarly enhanced predictions of respiratory, cardiovascular, heat-related, febrile, and headache conditions. Regression and unspecified machine learning methods linked rainfall, pollution, and temperature extremes to bronchitis and even vaginal bacterial composition. Yet a clear association between the AI technique used, the meteorological variable integrated, and the health outcome addressed emerged only for the well-established link between pollution and respiratory disease.

Beneath the encouraging results, however, the review uncovered a troubling pattern of underreporting. In seven of the twelve studies, the reasoning behind the choice of AI tool or performance metric was simply absent, leaving the added benefit of machine learning over conventional statistics unclear. Explainable AI appeared in only one study, despite techniques such as Shapley additive explanations, local interpretable model-agnostic explanations, and permutation feature importance being available to unpack black-box models. More worryingly, ethnicity was reported in only one study, and the review’s authors note that many excluded studies failed to record even basic sample sizes, suggesting the true scale of AI-driven climate-health research is far larger than the twelve studies captured. The absence of demographic detail likely degrades model performance and risks exacerbating existing health inequities, particularly for older adults and people with mental health or long-term conditions who are most vulnerable to weather-induced illness but least represented in the training data.

The review’s conclusions are a call to order for a field that could reshape public health preparedness. The authors recommend integrating explainable AI frameworks such as SHAP and LIME to transparently map non-linear interactions between meteorological variables and health outcomes, enforcing the reporting of age, sex, ethnicity, and socioeconomic status to prevent algorithmic bias, and requiring clear rationales for the selection of machine learning tools and performance metrics. They also urge energy-efficient AI designs combined with real-time climate data and physics-based standards, noting that the computational demands of AI carry their own environmental footprint through greenhouse gas emissions and electronic waste. Done well, AI-driven modelling of meteorological data could enable targeted interventions that prevent hospitalisation and premature death as climate change intensifies; done carelessly, it could produce opaque, biased tools that deepen the very inequities they are meant to address. The technology, the review makes clear, is ready. The reporting standards are not, and closing that gap is now the field’s most urgent task.

Subject of Research: The application of artificial intelligence to meteorological data for predicting weather-related health outcomes

Article Title: Application of Artificial Intelligence in analysing meteorological data for health research: A rapid review

Article References: Michels, L. V., Smith, L., Ghannam, S., Gadd, C., & Dambha-Miller, H. (2026). Application of Artificial Intelligence in analysing meteorological data for health research: A rapid review. Air Quality, Atmosphere & Health, 19(9), Article 201. https://doi.org/10.1007/s11869-026-02085-3

Image Credits: AI Generated

DOI: 10.1007/s11869-026-02085-3

Keywords: artificial intelligence, machine learning, meteorological data, climate change, public health, air pollution, Random Forest, predictive modelling, respiratory disease, explainable AI, heat-related mortality, health equity