Building the model: a review of input structures in extreme temperature–health Authors: Domenico D’Ausilio, Marta Ellena, Alfredo Reder, Alessandro Pugliese, Paola Mercogliano Published: 2026 ABSTRACT AbstractBackgroundThe health impacts of extreme temperatures have been extensively studied through epidemiological models. However, limited attention has been paid to the specification of these models, particularly regarding input structure and model selection. Although exposure metrics and statistical techniques have evolved over time, a comprehensive synthesis of the variables included in these models, and the rationale behind their inclusion, is still lacking. This gap limits the comparability of studies and may compromise the robustness of temperature–health evidence.MethodsWe conducted a systematic review of peer-reviewed studies published between 2014 and 2024 that employed quantitative epidemiological models to estimate the association between extreme temperatures and health outcomes. Following PRISMA guidelines, we selected 119 studies through searches conducted in Scopus, PubMed, and Web of Science. Each study was analysed in terms of spatial coverage, modelling framework, and model inputs. Inputs were classified into six functional groups: thermal exposures; environmental covariates (including both non-thermal meteorological variables and air pollutants); temporal controls; socio-demographic factors; health system indicators; and built environment characteristics.ResultsSubstantial heterogeneity was observed in both input selection and model specification. Daily mean temperature was the dominant exposure metric, though rarely justified over alternatives. Environmental covariates were inconsistently included: while relative humidity was frequent, other meteorological modifiers and air pollutants were often omitted without clear rationale. Temporal adjustments were common but heterogeneous. Distributed lag non-linear models were the prevailing framework, varying greatly in lag structure, spline specification, and covariate integration. Socio-economic, health, and infrastructural indicators appeared in less than one third of studies, typically as effect modifiers in meta-regression analyses, highlighting uneven integration of contextual determinants. No consensus currently exists on what constitutes a minimum model specification necessary to ensure reliable and interpretable effect estimates.ConclusionsCurrent temperature and health modelling remains fragmented, with notable variability in input specification and transparency. Strengthening methodological coherence through clearer guidance on input selection is essential. Greater integration of socio-economic and infrastructural variables would further enhance models’ capacity to capture contextual vulnerability. To ensure reliability and policy relevance, future research should develop shared guidelines for input specification, define minimum modelling standards, and promote transparent reporting of analytical decisions.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12940-026-01290-y. Introduction Extreme temperatures, including both high and low ambient temperatures as well as related phenomena such as the Urban Heat Island, pose a significant threat to human health and have become a central concern in both epidemiological and climate research. The intensification of climate change has increased population exposure to harmful thermal environments, not only through more frequent heatwaves and cold spells, but also via sustained deviations from thermoneutral conditions [1–3]. These trends underscore the urgent need for robust epidemiological models that can assess the short- and long-term health impacts of temperature variability and extremes. Understanding these impacts requires conceptual clarity on the definitions of exposure and risk. According to the AR6 WGII Intergovernmental Panel on Climate Change report, the term exposure refers to the presence of individuals, communities, infrastructure, or ecosystems in areas where climate-related hazards occur. Exposure alone does not determine the potential for harm; it must be considered alongside vulnerability, which reflects the susceptibility of individuals or systems to suffer adverse effects [1]. These dimensions of exposure and vulnerability are unequally distributed across populations, as socio-economic and infrastructural disparities modulate both the likelihood of exposure and the capacity to cope with temperature-related risks. The integration of these two factors with the climate hazard defines climatic risk. In epidemiological studies, however, the term exposure typically refers not to the spatial presence or absence of the population at risk, but rather to the intensity of the environmental hazard (e.g., ambient temperature, heat index, air pollution) that can actually affect the exposed population under consideration [4]. For clarity, we adopt this definition, aligning exposure with the hazard component of the IPCC risk framework. Modelling such exposures requires analytical approaches that reflect the distributional and temporal features of the risk phenomena. In most studies, the health outcomes of interest, such as mortality or morbidity (e.g. hospitalisation, emergency department visit), are measured as daily counts, making count-based regression models the standard analytical choice. These models typically rely on the Poisson distribution, which assumes that the variance equals the mean. However, health data often exhibit overdispersion [5, 6], where the variance exceeds the mean, prompting many studies to adopt the quasi-Poisson alternative, which incorporates a dispersion parameter to better account for this variability. A major methodological advancement in this field has been the development and widespread adoption of Distributed Lag Non-Linear Models (DLNMs) [7] This framework allows for the simultaneous modelling of non-linear associations between temperature and health outcomes, as well as delayed effects distributed over time (i.e., lag). Such a modelling approach is particularly relevant given that prolonged exposure to non-optimal temperatures exacerbates a range of health risks, including cardiovascular, respiratory, and heat-related illnesses [8–10]. Heat exposure can cause acute health events such as heatstroke, dehydration, and exacerbation of pre-existing cardiovascular and respiratory conditions [11]. In contrast, cold exposure tends to exert more gradual but longer-lasting effects, increasing the risk of hypothermia and cardiovascular stress [12–14]. Several studies suggest that cold-related mortality often exceeds that from heat, particularly in vulnerable populations [15–18]. Moreover, the effects are unevenly distributed, with older adults, infants, pregnant women, and individuals with chronic illnesses facing heightened risks due to physiological and social vulnerabilities [19–22]. The association between temperature and health has been assessed in numerous reviews, often referred to as the heat-health nexus [23, 24], primarily focusing on its detrimental effects across a wide range of populations and outcomes [25–27]. Nonetheless, while methodological reviews on climate-related health risks do exist, they have largely concentrated on the definition and measurement of the components, according to the IPCC framework [28–30]. Relatively little attention has been devoted to the analytical foundations of epidemiological models themselves, particularly to how these models are specified and to the rationale behind the inclusion of certain covariates [31]. The analysis of these aspects is fundamental as these methodological choices, such as the selection of lag structures, exposure metrics, confounding adjustment strategies, or the inclusion of socio-economic proxies can significantly influence the estimates produced and their comparability across contexts [32–34]. Beyond these analytical aspects, it is also essential to consider the geographical distribution of studies and data sources, as modelling practices and empirical evidence are unevenly represented across world regions. Such disparities, driven by differences in research capacity, data availability, and climatic conditions, can bias the global understanding of the temperature–health relationship. The uneven spatial distribution of data and research capacity across climatic and socio-economic contexts affects not only the quality of input data but also the transferability and generalisability of modelling frameworks across regions [35]. Analysing these geographical patterns helps identify where evidence is concentrated and where major knowledge gaps remain. To address this gap, the present study systematically reviews the epidemiological models employed to analyse the temperature–health relationship, with a specific focus on input variables and their role within the modelling frameworks. Input variables play a crucial role in adjusting for confounding, capturing effect modification, and improving model fit and among them contextual and socio-economic factors are particularly relevant for explaining between-site heterogeneity, reflecting how demographic, infrastructural, and economic disparities shape population vulnerability [36]. In this sense, the inclusion of socio-economic and contextual indicators is not only a matter of confounding control but also a key to understanding how social and spatial inequalities shape climatic vulnerability (e.g. income inequality, urban density, healthcare availability). To the best of our knowledge there is currently no systematic overview of how they are selected, categorised, and integrated across studies. This review contributes to the current state-of-the-art by providing a structured synthesis of current modelling practices related to covariate use in the extreme temperature–health nexus (hereafter referred to as the ET–health nexus). By mapping covariate types, data sources, modelling strategies (e.g., functional forms, interaction terms), and adjustment approaches, we aim to enhance transparency and comparability across studies, while also identifying recurring blind spots and inconsistencies that may compromise the robustness of epidemiological evidence on the temperature–health relationship. This paper is structured as follows: Section "Materials and Methods" outlines the methodological framework adopted for the systematic review, detailing the search strategy, eligibility criteria, and data extraction process. Section "Results" presents the results, classifying studies based on geographic scope and modelling approaches, with particular attention to the covariates integrated into statistical models. Section "Discussion" discusses the key findings, methodological challenges, and future research directions, while Section "Conclusion" summarises the main conclusions and implications for research and policy. Materials and methods Literature selection and screening process The literature selection process involved the identification of peer-reviewed studies published in English between January 2014 and August 2024 that investigated the ET-health nexus using epidemiological modelling. The chosen timeframe minimises the inclusion of outdated methodologies and ensures alignment with current analytical standards in climate-health epidemiology. The proposed review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [37]. Three databases, i.e., Scopus, PubMed, and Web of Science (WoS), were searched, given their comprehensive coverage of peer-reviewed scientific literature. To ensure a comprehensive inclusion of relevant studies, search terms were grouped into five thematic categories following a hierarchical approach. At the top of the hierarchy, the (i) climate category acts as an overarching domain encompassing the environmental context. This is followed by more specific dimensions reflecting (ii) the type of climatic hazard (extreme temperatures), (iii) its health implications (health), and (iv–v) the social mechanisms through which risk is differentially distributed (vulnerability and inequality) (see Appendix A for full keyword structure). In this review, we use the term extreme temperatures in a broad sense to indicate non optimal thermal conditions, including studies that focus exclusively on heat or on cold or both. The health category was included to ensure that the search captured studies that quantified temperature related associations through epidemiological models, regardless of the specific health endpoint considered. Socioeconomic and vulnerability related keywords were included to improve retrieval of studies on population level risk rather than as formal inclusion criteria, and studies were not excluded based on whether such variables were used as model inputs. Keyword selection was guided by a combination of established frameworks in the academic literature [19, 38], definitions proposed by major institutional sources [1, 39] and inductive insights from a preliminary mapping of the field. This triangulated approach ensured that the terminology captured both the conceptual architecture of the climate–health nexus and the empirical diversity of existing studies. The selection process is summarised in Fig. 1. Study selection workflow The initial search yielded 1,972 records: 955 from Scopus, 662 from WoS, and 355 from PubMed. After applying inclusion criteria—peer-reviewed articles, written in English, and published between 2014 and 2024—the dataset was reduced to 1,295 articles. Removal of 471 duplicates resulted in 822 unique records, which underwent title screening to retain studies explicitly addressing the ET-health nexus. with the objective of identifying studies that employed a quantitative epidemiological modelling approach. Of these, 12 were excluded due to insufficient alignment with the research objectives, yielding a final sample of 72 studies. Additional 20 studies were identified through snowballing techniques, and 27 Multi-Country Multi-City (MCC) Collaborative Research Network studies were also incorporated to ensure comprehensive coverage of large-scale comparative frameworks. These MCC studies, which employ harmonised analytical protocols across multiple national contexts, were not fully captured by the initial search due to their specific publication structure and consortium-based design. The final corpus therefore comprised 119 articles. Risk of bias assessment As recommended by the PRISMA 2020 statement [37], risk of bias assessments are a critical component of systematic reviews. Traditional risk of bias assessment tools are primarily designed for clinical or observational studies with well-defined exposure–outcome structures. However, this review focuses on the input dimension of epidemiological models applied to the ET-health nexus—specifically, how exposures, confounders, and temporal trends are defined, measured, and integrated into the model structure. As such, evaluating the methodological soundness of these modelling inputs is essential to determine whether the statistical foundations of each study are adequate to support robust and interpretable findings. By systematically examining how inputs are selected, specified, and adjusted for bias, we are able to distinguish between studies that demonstrate high methodological rigour and those that introduce structural limitations. This distinction provides a critical foundation for interpreting variation in modelling practices and enhances the comparability and transparency of the synthesis. To this end, and following [28], we adapted the Office of Health Assessment and Translation (OHAT) Risk of Bias Rating Tool for Human and Animal Studies (National Toxicology Program, 2015) The customized tool includes ten questions (Appendix B), each rated using a four-point scale: “definitely low” (+ +), “probably low” (+), “probably high” (–), and “definitely high” (– –) risk of bias [40]. Responses were then averaged to assign each study to one of four qualitative categories: “definitely high quality”, “probably high quality”, “probably low quality”, or “definitely low quality”. Definition and classification of model inputs In this review, we define model inputs as all variables explicitly included on the right-hand side of the statistical models used to estimate associations between extreme temperatures and health outcomes. To ensure conceptual clarity and comparability across studies, we classify these inputs based on their functional role within the model structure. The first category includes exposure variables, such as daily air temperature, apparent temperature, or other derived heat stress variables and indicators. These are the central independent variables of interest, representing the climatic component of the temperature–health relationship. The second category comprises inputs introduced to adjust for potential sources of bias or to improve the statistical specification of the model. These include environmental confounders like humidity and air pollution, temporal structures such as long-term trends, seasonality, or short-term effects, and occasional controls for age, sex, or socio-economic status. Although their epidemiological status may vary, they serve a shared purpose of refining the attribution of health outcomes to temperature-related exposures. This classification is based on the functional role of variables within the statistical estimation process, rather than on strict causal hierarchies. It reflects a pragmatic approach widely adopted in environmental epidemiology, where improving model precision, accounting for contextual factors, and reducing residual confounding are often prioritised over establishing direct causal pathways. By consistently applying this typology, the review enables a structured synthesis of modelling strategies and supports a clearer interpretation of how input variability shapes epidemiological evidence on the ET–health nexus. Results This section presents the main findings of the systematic review. For the sake of clarity, the results are structured into four key analytical dimensions. First, we report the results of the risk of bias assessment for the included studies (in Section "Risk of bias"), providing an overview of their overall methodological quality. Second, we examine the geographical distribution and spatial resolution of all the studies, highlighting regional concentration patterns and differences in the unit of analysis (in Section "Geographical scope and spatial resolution of modelling frameworks"). Third, we summarise the statistical modelling frameworks employed to estimate the health impacts of extreme temperatures, with particular attention to the use of distributed-lag structures, distributional assumptions, and model types (in Section "Modelling approaches for assessing the impact of ET-health nexus outcomes"). Finally, we classify the model inputs into six functional categories: (i) thermal exposures, (ii) environmental and pollution-related covariates, (iii) temporal controls, (iv) socio-demographic variables, (v) health system indicators, and (vi) built environment factors (therefore providing a structured synthesis of their inclusion and operationalisation across studies) (in Section "Classification and impact of model input"). To support this analysis, Table 1 provides a consolidated overview of the 119 studies included in the review. For each study, we report the type of exposure assessed, the health outcome analysed, the study location and spatial resolution, the period covered, and the modelling approach used, including the assumed statistical distribution. This reference table is used throughout the section to contextualise examples and identify broader trends in the modelling literature. Table 1Characteristics of selected studies in the reviewIDExposureHealth outcomeLocationSpatial ResolutionStudy PeriodApplied ModelDistribution[41]HeatMortality4 cities, ChinaCity2004—2012GAMPoisson[42]HeatMortalityPudong, ChinaCity2008—2013GAMQuasi-Poisson[43]Heat; ColdMortality3 counties, ChinaCounty2008—2012GLM—DLNMQuasi-Poisson[44]HeatMortalityGeorgia, USCounty1995—2004GLM with mixed effectPoisson[45]Heat; ColdMortality306 locations in 12 countriesLocations1988—2011GLM—DLNM—MAQuasi-Poisson[46]HeatMortality66 communities, ChinaCommunity2006—2011GLM—DLNM—MA- MRPoisson[47]Heat; ColdMortalityFukuoka, JapanCity1973—2012GLM—DLNM—MAPoisson[48]Heat; ColdMortalityMadrid, SpainCity2001—2009GLMPoisson[49]Heat; ColdMortality384 locations in 13 countriesLocations1985—2012GLM—DLNM—MA—MRQuasi-Poisson[50]HeatMortality272 locations in 7 countriesLocations1985—2012GLM—DLNM—MAQuasi-Poisson[51]HeatMortalityAdministrative units, South-KoreaAdministrative units1993—2012GAMPoisson[52]HeatMortalityMaricopa County, USCounty2005—2010GLMPoisson[53]Temperature variabilityMortality372 locations in 12 countriesLocations1972—2012GLM—DLNM—MAQuasi-Poisson[54]HeatMortality305 locations in 9 countriesLocations1985—2012GLM—DLNM—MA—MRQuasi-Poisson[55]HeatMorbidity (HS)*25 cities, VietnamProvince2005—2015GLM—DLM—MA—MRPoisson[56]Heat; ColdMortalityJinan, ChinaCity2011—2014GLMPoisson[57]HeatMorbidity (HS)*Western Australia, AustraliaCity2006–2015GLMPoisson[58]HeatMorbidity (HS)*Chongqing, ChinaCity2009–2013GLMZero inflated Poisson[59]HeatMortalityJinan, ChinaCity2007–2013GAMQuasi-Poisson[60]Temperature variabilityMortalityHubei, ChinaCounty2009–2012GLM—DLNM—MAQuasi-Poisson[61]Heat; ColdMortality278 locations in 12 countriesLocations1972—2012GLM—DLNM—MAGaussian[62]HeatMortality400 locations in 18 countries/regionsLocations1985—2013GLM—DLNM—MA—MRQuasi-Poisson[63]Heat; ColdMorbidity (HS)*Seoul, South KoreaCity2007–2014GLMPoisson (assumed)[64]HeatExposure50 metropolitan areas, USMetropolitan area2004–2015Fixed effect regressionLinear[65]Heat; ColdMortality52 districts, South AfricaDistrict1997—2013GLM—DLNM—MAQuasi-Poisson[66]Heat; ColdMortality308 locations in 10 countriesLocations1972—2013GLM—DLNM—MA—MRQuasi-Poisson[67]ColdMortality31 cities, ChinaCity2007—2013GLM—DLNM—MAQuasi-Poisson[68]HeatMorbidity (ABO)*Seoul, South KoreaCity2004—2012Logistic regressionBinomial[69]HeatMorbidity (ED)*Perth, AustraliaCity2006—2015GLMPoisson[70]Heat; ColdMortality89 counties in Zejiang, ChinaCounty2009—2015GLM—DLNM—MA—MRQuasi-Poisson[71]HeatMorbidity (OV)*Cangan, ChinaCity2010—2012Conditional regressionPoisson[72]Heat; ColdMorbidity (HS)*Hefei, ChinaCity2005—2014GLM—DLNMQuasi-Poisson[12]Heat; ColdMortality340 cities in 22 countriesCity1985–2014GLM—DLNM—MRQuasi-Poisson[73]ColdMortality10 provinces, SpainProvince1983—2003GLM—DLNM—MAPoisson[74]HeatMortality; Morbidity (HS)*; Morbidity (ED)*TaiwanCountry2000—2014GAMPoisson[75]Temperature variabilityMortality89 districts in Zejiang, ChinaDistrict2009—2015GLM—DLNM—MAQuasi-Poisson[76]Heat; ColdMortality7 cities, South KoreaCity1998—2013Conditional regression—DLNM—MRPoisson[77]HeatMortality31 cities, ChinaCity2007—2013GLM—DLNM—MAQuasi-Poisson[78]Temperature variabilityMorbidity (HS)*1814 cities, BrazilCity2000—2015Conditional logistic regression—MA—MRBinomial[79]HeatMorbidity (ED)*8 communities in Queensland, AustraliaCommunity2013—2015GAM—DLNM—MAQuasi-Poisson[80]Heat; ColdMortality50 provinces, SpainProvince1993—2013GLM—DLNM—MA—MRQuasi-Poisson[81]HeatMortality445 locations in 24 countriesLocations1972—2015GLM—DLNM—MA—MRQuasi-Poisson[82]Heat; ColdMortality341 locations in 12 countriesLocations1986—2013Conditional regression—DLNM—MA—MRQuasi-Poisson[23]HeatMortalityTurin, ItalyCity1982—2018GLM—DLNMQuasi-Poisson[83]HeatMorbidity (HS)*1814 cities, BrazilCity2000—2015GLM—DLM—MA—MRQuasi-Poisson[84]HeatMorbidity (ED)*18 sites, ChinaGeo-meteorological regions2014—2017GAM—MA—MRQuasi-Poisson[85]Heat; ColdMortality289 small areas in Hong Kong, ChinaTertiary Planned Units (TPUs)*2007—2015GLM—DLNMQuasi-Poisson[86]Heat; ColdMortality12 cities, GermanyCity1993—2015GLM—DLNM—MRQuasi-Poisson[87]Heat; ColdMortality31 cities, ChinaCity2007—2013GLM—DLNM—MA—MRQuasi-Poisson[88]HeatMortality4 cities, SpainCity1990—2014GAM—DLNMQuasi-Poisson[89]Heat; ColdMortalityUrmia, IranCity2010—2015GLM—DLNMQuasi-Poisson[90]Heat; ColdMortalityHong Kong, ChinaCity2006—2016GLM—DLNMQuasi-Poisson[91]HeatMortality311 locations in 4 countriesLocations1972—2009GLM—DLNM—MRQuasi-Poisson[92]Heat; ColdMorbidity (HS)*; MortalityBrisbane, AustraliaCity2005—2013Conditional logistic regressionBinomial[93]Heat; ColdMortalitySofia, BulgariaCity2000—2017GLM—DLNMQuasi-Poisson[94]HeatMortalityHong Kong, ChinaTertiary Planned Units (TPUs)*2015—2017GAM—Multi-scale geographically weighted regression (MGWR)Gaussian (assumed by the MGWR framework, not explicitly stated)[95]Heat; ColdMortalityQueensland, AustraliaCity1997—2013Conditional regression—DLNMQuasi-Poisson[96]HeatMortalityBendafassi, SenegalCity1973—2012GAM—DLNMPoisson[97]Heat; ColdMortalityMulti-country, AsiaCountry2010—2016GLM—DLNM—MA—MRQuasi-Poisson[98]ColdMorbidity (HS)*Hefei, ChinaCity2013—2016GAM—DLNMPoisson[99]HeatMorbidity (ED)*Adelaide, AustraliaCity2014–2017GLM – DLNMQuasi-Poisson[100]Heat; ColdMortality750 locations in 43 countriesLocations2000—2019GLM—DLNM—MRQuasi-Poisson[101]HeatMortality11 cities in 4 countriesCity2000—2012GLM—DLNM—MAQuasi-Poisson[102]HeatMortality732 locations in 43 countriesLocations1991—2018GLM—DLNM—MRQuasi-Poisson[103]Heat; ColdMortality658 locations in 43 countriesLocations1984—2016GLM—DLNM—MA—MRQuasi-Poisson[104]Heat; ColdMortality719 locations in 34 countriesLocations1969—2016GLM—DLNM—MA—MRQuasi-Poisson[105]Heat; ColdMorbidity (IR)*30 cities, ChinaCity2016–2019GLM – DLNM – MA – MRQuasi-Poisson[106]HeatMortality452 locations in 24 countriesCity2000–2016GLM – DLNM – MRQuasi-Poisson[107]Heat; ColdMorbidity (ABO)*16 counties, ChinaCounty2014—2018Cox regression—MA—MRN/A (semi-parametric)[15]Heat; ColdMortalityTurin, ItalyCity1982—2018GLM—DLNMQuasi-Poisson[108]ColdMortalityJinan, ChinaCity2011—2020Conditional logistic regressionBinomial[109]HeatMortalitySeoul, South KoreaCity1999–2018GAMPoisson[110]ColdMortality10 provinces, SpainMunicipalities2000—2013ARIMA + Linear Mixed ModelGaussian[111]HeatMorbidity (HS)*Hong Kong, ChinaCity2010—2019GAM—DLNMQuasi-Poisson[112]HeatMortalitySE England and Aberdeenshire, United KingdomCity1981—2018GLM—DLNMQuasi-Poisson[113]Heat; ColdMortality171 countries, multi-CountryCountry2010—2019Two-way fixed-effects—dynamic panel GMM modelGaussian[114]HeatMorbidity (HS)*443 postal areas in Queensland, AustraliaPostal area1995—2016Conditional regression—DLNMQuasi-Poisson[115]Heat; ColdMortality364 communities, ChinaCommunity (County or district of a city)2013–2017GAM with Difference-in-Differences (DID) frameworkQuasi-Poisson[116]Heat; ColdMortalityPune, IndiaCity2004—2012Conditional regression—DLNMQuasi-Poisson[117]Heat; ColdMortality5 cities, RomaniaCity1999—2016GLM—DLNMQuasi-Poisson[118]Temperature variabilityMortality717 locations in 36 countriesLocation1972—2018GLM—DLNM—MA—MRQuasi-Poisson[119]Temperature variabilityMortality750 locations in 43 countriesLocation1979—2019GLM—DLNM—MRQuasi-Poisson[120]HeatMortalityCountry, ChinaCountry2013–2019Conditional logistic regression – DLNMBinomial[121]HeatMorbidity (HS)*5572 municipalities, BrazilMunicipality2008–2018GLM – DLNM – MAQuasi-Poisson[122]HeatMorbidity (HS)*23 sites, ChinaSite2014—2019Conditional regression – DLNM – MA – MRQuasi-Poisson[123]Heat; ColdMortality48 provinces, SpainProvince1980–2018GLM—DLNM—MA—MRQuasi-Poisson[124]HeatMorbidity (HS)*Madrid, SpainCensus-section of residence2015–2018Conditional logistic regression – DLNMBinomial[125]Heat; ColdMorbidity (HS)*Country, South KoreaCountry2002–2019Conditional regression—DLNMQuasi-Poisson[126]Heat; ColdMortalityShenzen, ChinaMunicipality2013—2018GLM—DLNMQuasi-Poisson[127]Heat; ColdMorbidity (ED)*58 cunties in California, USCounty2005–2019Conditional logistic regression – DLNMBinomial[128]Heat; ColdMortality17 cities, TaiwanAdministrative unit2005–2018GLM – DLNM—MAQuasi-Poisson[129]Heat; ColdMorbidity (HS)*31 provinces, ChinaProvince2014–2020Difference-in-Differences (DID)Gaussian[130]Heat; ColdMortality325 cities, Latin AmericaCity2002–2015Conditional regression – DLNM – MRPoisson[131]Heat; ColdMorbidity (HS)*Country, IsraelCountry2004–2018Conditional logistic regression – DLNMBinomial[132]HeatMorbidity (HS)*9917 zip codes, USMetropolitan area2000–2017GLM – DLNM – MAQuasi-Poisson[133]HeatMortality24 districts, ChinaDistrict2014–2019Conditional regression – DLNM – MAPoisson[134]HeatMorbidity (HS)*1816 municipalities, BrazilMunicipality2000–2015Conditional logistic regression—DLNM—MRBinomial[135]Heat; ColdMortality6 regions, MalaysiaAdministrative unit2014–2018Conditional regression – DLNM—MA—MRQuasi-Poisson[136]Heat; ColdMorbidity (HS)*Chengdu, ChinaCity2015–2020GLM – DLNMQuasi-Poisson[137]Temperature variabilityMorbidity (HS)*Hong Kong, ChinaCity1999–2019GLM – DLNMQuasi-Poisson[138]HeatMortalityNorthern Australia, AustraliaCommunity (County or district of a city)1980–2019Conditional regression – DLNMQuasi-Poisson[139]HeatMortality482 locations in 24 countriesLocations2000—2018GLM—DLNM—MAQuasi-Poisson[140]Heat; ColdMortality567 locations in 27 countriesLocations1979—2019Conditional regression—DLNM—MA—MRQuasi-Poisson[16]Heat; ColdMortality854 locations in 30 countriesLocations2000—2019GLM—DLNM—MRQuasi-Poisson[141]HeatMortality620 locations in 36 countriesLocations1995—2019GLM—DLNM—MAQuasi-Poisson[142]Heat; ColdMorbidity (HS)*12 Metropolitan Statistical Areas in Texas, USMetropolitan Statistical Areas (MPA)*2004—2013GAM—MAQuasi-Poisson[143]HeatMorbidity (ED)*18 districts in Hong Kong, ChinaDistrict2010–2019GAM – DLNMQuasi-Poisson[144]HeatMortality30 cities, ChinaCity2010–2016GAM – DLNMQuasi-Poisson[145]HeatMorbidity (HS)*Hong Kong, ChinaMetropolitan area2000–2019GAM – DLNMQuasi-Poisson[146]Heat; ColdMorbidity (HS)*13 prefectures in Jiangsu, ChinaPrefecture2010–2017GLM – DLNMQuasi-Poisson[147]Heat; ColdMortalityZibo city, ChinaCity2015–2019GLM – DLNMQuasi-Poisson[148]HeatMortality1287 municipalities in Paris, FranceAggregated units for statistical information (IRIS)*2000–2020GLM – DLNMQuasi-Poisson[149]HeatMorbidity (HS)*23 study sites, ChinaSite2014–2019Conditional regression – DLNM—MAQuasi-Poisson[150]HeatMorbidity (ED)*Curitiba, BrazilUrgency and Emergency Units (UPA)*2017–2021GLM – DLNM—MRQuasi-Poisson[151]ColdMortality1960 locations in 59 countriesLocations2000—2019GLM—DLNM—MRQuasi-Poisson[152]Heat; ColdMortality757 locations in 47 countriesLocations1979—2020GLM—DLNM—MAQuasi-Poisson[153]HeatMortality750 locations in 43 countriesLocations1972—2018GLM—DLNM—MRQuasi-Poisson[154]Heat; ColdMortality522 locations in 26 countriesLocations1979—2019Conditional regression—DLNM—MA—MRQuasi-Poisson[155]Heat; ColdMortality1847 locations in 32 countriesLocations1969—2019GLM—DLNM—MRQuasi-Poisson*MA Meta-analysis, MR Meta-regression, ED Emergency department visits (or attendances/presentations), HS Hospitalisation, OV Outpatient visit, ABO Adverse birth outcome, IR Influenza rate Characteristics of selected studies in the review *MA Meta-analysis, MR Meta-regression, ED Emergency department visits (or attendances/presentations), HS Hospitalisation, OV Outpatient visit, ABO Adverse birth outcome, IR Influenza rate To facilitate tabular presentation and improve the overall clarity of the manuscript, studies included in the review are hereafter referred to by their unique ID numbers, as reported in Table 1. Risk of bias Among the 119 studies included in the review, 51.26% (n = 61) were classified as probably high quality, while the remaining 48.74% (n = 58) were classified as definitely high quality. Studies classified as probably high quality generally exhibited sound methodological practices but presented minor limitations, such as partial reporting of covariate selection criteria or incomplete sensitivity analyses. Conversely, studies assessed as definitely high quality demonstrated full transparency in model input specification, rigorous confounder adjustment, and extensive robustness checks. No studies were rated as probably low or definitely low quality, indicating a generally high methodological standard across the reviewed literature. The detailed risk of bias assessment for each individual study is reported in Appendix B. Geographical scope and spatial resolution of modelling frameworks The geographical coverage of the reviewed literature is a key descriptive dimension, as it reflects where empirical evidence on the temperature–health relationship is being generated. The majority of studies are concentrated in Asia, which accounts for 45.38% of the included articles (n = 54). China alone concentrates 31.93% of the studies (n = 38), followed by Australia (6.72%, n = 8), the United States (5.88%, n = 7), and Spain (5.88%, n = 7). Europe accounts for 11.76% of the studies (n = 14), while North and South America each account for 5.04% (n = 6) and only 1.68% (n = 2) were conducted in African countries. Moreover, 26.89% of the studies (n = 32) adopt a multi-country design, of which 25.21% (n = 30) mostly belonging to the MCC framework, collectively providing a global perspective that spans multiple continents and climatic zones. This information is summarised in Fig. 2. Global distribution of research studies The spatial resolution of the modelling frameworks varied considerably across the reviewed studies. Cities were by far the most commonly used unit of analysis, appearing in 38.66% of the studies (n = 46). These included both single-city analyses (e.g., Zhejiang [75], Queensland [79], Madrid [124]) and multi-city designs, ranging from small-scale comparisons such as 5 cities in Romania [117] or 12 cities in Germany [86] to large-scale investigations such as Xu et al. [92] that analysed 1,814 cities in Brazil. The second most frequent spatial units was represented by locations in the MCC studies (21.85%; n = 26), which include different administrative levels such as cities, provinces or metropolitan areas within harmonised multi-country frameworks (e.g., [54, 81]). Counties and countries followed, being used in 5.88% (n = 7) and 5.04% (n = 6) of the studies, respectively. Counties featured in studies from the United States [44, 52] and China [43, 60, 70, 107], while national-level analyses were conducted in Taiwan [74], Israel [131], China [120] and South Korea [125], as well as in broader regional [97] and global [113] frameworks. Metropolitan regions were used in 3.36% of the studies (n = 4), in the United States [64, 132, 142]) and China [145]. Municipalities were analysed in Brazil [121, 134]), China [126], and Spain [110]. Districts featured in 3.36% of the studies (n = 4) from Africa [65] and China [75, 134, 143]. Communities were also the unit of analysis in 3.36% of the studies [46, 79, 115, 138]; the term referred to sub-provincial divisions in China and to remote or rural settlements in Australia. Moreover, 2.52% of the studies used provinces as the spatial unit (n = 3) used provinces as the spatial unit, notably in Spain [73, 123] and China [129]. A total of 7.56% of the studies (n = 9) employed more granular spatial frameworks. In Hong Kong, these included Tertiary Planning Units (TPUs: [85, 94]); in France, IRIS units (Ilots Regroupés pour l’Information Statistique; [126]); and in Madrid, census sections [124]. Curitiba (Brazil) used UPAs (Unidades de Pronto Atendimento: [150]), while prefecture-level divisions were used in China [146]. Modelling approaches for assessing the impact of ET-health nexus outcomes The reviewed literature employed a diverse range of statistical approaches to estimate the health impacts of extreme temperatures, reflecting both the distributional properties of health outcome data and the complexity of exposure–response relationships. Poisson regression was used in 18.49% of the studies (n = 22), while 68.08% (n = 81) relied on quasi-Poisson models to address overdispersion, a common issue in time-series analyses of mortality and morbidity. Most analyses were implemented within Generalised Linear Models (GLMs; 59.66%, n = 71) and Generalised Additive Models (GAMs; 16.81%, n = 20), while conditional regression (9.24%, n = 11) and conditional logistic regression models (6.72%, n = 8) were used primarily in case-crossover designs. Within these frameworks, Distributed Lag Non-Linear Models (DLNMs) were applied in 74.79% of the studies (n = 89) to characterise delayed and non-linear associations between temperature and health outcomes. Since DLNMs are embedded within broader statistical frameworks rather than representing independent models, their use overlaps with the other modelling categories. A structured overview of these modelling strategies is provided in Table 2. Table 2Description and relative representation of the applied epidemiological modelsMODELDESCRIPTION% OF PAPERSREFERENCESGeneralised Linear Model (GLM)An extension of linear models to accommodate non-Gaussian response variables using a link function58.82[12, 15, 16, 23, 43–50, 52–58, 60–63, 65–67, 69, 70, 72, 73, 75, 77, 80, 81, 83, 86, 87, 89–91, 93, 99–106, 112, 117–119, 121, 123, 126, 128, 132, 136, 137, 139, 142, 146–148, 150–153, 155]Generalised Additive Model (GAM)A flexible model that combines linear and non-linear components to capture complex relationships between variables16.81[41, 42, 51, 59, 74, 79, 84, 85, 88, 94, 96–98, 109, 111, 115, 142–145]Distributed Lag Non-linear Model (DLNM)A model designed to assess delayed and non-linear effects of exposures on outcomes74.79[12, 15, 16, 23, 43, 45–47, 49, 50, 53, 54, 60–62, 65–67, 70, 72, 73, 75–77, 79–82, 85–91, 93, 95–106, 111, 112, 114, 116–128, 130–141, 143–155]Meta-analysisA statistical approach that combines results from multiple studies to derive a pooled estimate of the overall effect39.5[45–47, 49, 50, 53–55, 60–62, 65–67, 70, 73, 75, 77–84, 87, 97, 101, 103–105, 107, 118, 121–123, 128, 132, 133, 135, 139–142, 149, 152, 154]Meta-regressionAn extension of meta-analysis that examines how study-level characteristics influence the estimated effects across studies33.61[12, 16, 46, 49, 54, 55, 62, 66, 70, 76, 78, 80–84, 86, 87, 91, 97, 100, 102–107, 118, 119, 121–123, 130, 134, 140, 150, 151, 153–155]Distributed Lag Model (DLM)A model that estimates the lagged effects of one variable on another over time1.68[55, 83]Fixed effect modelA model controlling for unobserved time-invariant heterogeneity within specific units1.68[64, 113]Logistic regression modelA model for binary outcomes that estimates the probability of an event using a logistic function0.84[68]Conditional regression modelA regression approach conditioned on a set of covariates or specific data structures10.08[71, 76, 82, 95, 114, 116, 122, 125, 130, 133, 135, 154]Conditional logistic regression modelA variation of logistic regression tailored for paired data or case–control studies6.72[78, 92, 108, 120, 124, 127, 131, 134]Multiscale Geographically Weighted Regression (MGWR)An extension of GWR allowing for spatial relationships at varying scales0.84[94]Cox regression modelA survival model analysing time-to-event data, accounting for censoring0.84[107]Generalised Mixed ModelModels incorporating both fixed and random effects, suited for hierarchical or grouped data0.84[113]Differences-in-DifferencesA technique evaluating the impact of an intervention by comparing temporal changes between groups1.68[115, 129]General Method of Moments (GMM)A method for estimating model parameters by leveraging sample moments as constraints0.84[113]ARIMAA time series model capturing autoregressive and moving average components to describe temporal dependencies and trends0.84[110]Linear mixed modelExtends linear regression by including both fixed and random effects to account for correlated or nested observations0.84[110] Description and relative representation of the applied epidemiological models In multi-site and multi-region settings, synthesis techniques were frequently applied to derive generalisable effect estimates and assess between-site heterogeneity. Overall, 39.5% of the studies (n = 47) performed a meta-analysis and 33.61% (n = 40) conducted a meta-regression to pool or explain variability in location-specific associations. Notably, 18.49% of the studies (n = 22) implemented both methods within a multi-stage analytical design, combining estimation and effect modification analysis. Classification and impact of model input The inclusion of confounders in epidemiological models is essential, as they account for factors that can influence both the climate hazard and the health outcome [156]. Misspecification of confounders may result in spurious associations or in over- or underestimation of the true effect, thereby reducing the validity of epidemiological estimates [157]. To facilitate analysis, we classified the confounders reported across the 119 studies into several categories, providing a systematic framework for comparing and understanding their role within the ET–health nexus. Exposure variables: thermal variables Across the reviewed studies, we identified 12 types of temperature-related variables used to model the health effects of extreme thermal conditions, as summarised in Table 3. Table 3Summary of thermal variablesVARIABLEDESCRIPTION% OF PAPERSREFERENCESDaily maximum air temperatureHighest air temperature recorded in a 24-h period15.13[41, 42, 46, 51, 52, 58–60, 65, 74, 77, 96, 112, 117, 124, 127, 134, 149]Daily minimum air temperatureLowest temperature measured over a 24-h cycle9.24[59, 60, 65, 66, 73, 96, 98, 110, 117, 127, 132]Daily mean air temperatureAverage of the daily maximum and minimum temperatures, or 24-h hourly mean73.11[12, 15, 16, 23, 43, 45, 47–50, 53, 54, 56, 57, 59, 61–63, 67, 70–72, 75, 76, 78–84, 86–88, 90–93, 95, 98, 100–108, 111, 114–116, 118–120, 122, 123, 125, 128–155]Daily maximum apparent temperatureHighest perceived temperature during the day, typically based on dailyand humidity2.52[44, 98, 127]Daily minimum apparent temperatureLowest perceived temperature during the day, usually based onand humidity2.52[44, 98, 127]Daily mean apparent temperaturePerceived temperature calculated usingand relative humidity5.04[55, 64, 68, 74, 97, 109]Wet-Bulb Globe TemperatureComposite index combining temperature, humidity, wind speed, and solar radiation; used to assess occupational and outdoor heat risk0.84[74]Universal Thermal Climate IndexPhysiological index of thermal stress based on a multi-node human thermoregulation model; includes temperature, humidity, wind, and radiation1.68[113, 126]Excess Heat FactorMetric combining short- and long-term temperature anomalies to identify heatwave events contextually2.52[69, 85, 99]Mean Land Surface TemperatureTemperature of the Earth’s surface derived from satellite thermal imagery1.68[94, 121]Physiological Equivalent TemperatureIndex estimating perceived temperature based on meteorological conditions and human physiology; includes wind, humidity, radiation0.84[89] Summary of thermal variables The most frequently employed metric was daily mean air temperature (), used in 73.11% of the studies (n = 87) followed by daily maximum temperature (; 15.13%, n = 18) and daily minimum temperature (; 9.24%, n = 11). The second most commonly used thermal exposure was apparent temperature, applied in 10.08% of the studies (n = 12). Among these, 5.04% (n = 6) used daily mean apparent temperature (), while 2.52% (n = 3) focused on daily maximum apparent temperature (), and another 2.52% (n = 3) on daily minimum apparent temperature (). Other thermal metrics included the Universal Thermal Climate Index (; 1.68%, n = 2), Wet-Bulb Globe Temperature (; 0.84%, n = 1), average Land Surface Temperature (; 1.68%, n = 2), the Excess Heat Factor (; 2.52%, n = 3), and Physiological Equivalent Temperature (; 0.84%, n = 1). A total of 9.24% of the studies (n = 11) employed more than one temperature-related variable within the same model. These combinations frequently included , , and , often alongside perceived temperature indices such as , or . Multi-metric approaches were adopted to jointly assess average and extreme exposures, or to compare the performance of different thermal indicators. Environmental variables: non-thermal and air pollution In addition to temperature, several studies included meteorological variables to account for broader environmental conditions. Relative humidity was the most commonly used, appearing in 48.74% of the studies (n = 58). It was included either as a standalone covariate—often modelled using splines—or as a component of composite indices such as apparent temperature and the heat index. Humidity was modelled in various ways, including non-linearly [97], percentile thresholds [146], and standard deviations [129]. Air pressure was included in 11.76% of the studies (n = 14), typically in models addressing cold-related effects or in high-altitude regions. It was usually introduced as a continuous or spline-transformed variable [84, 85]. Wind speed was included in 8.4% of studies (n = 10), often modelled as a continuous variable or spline, and sometimes incorporated into thermal indices such as apparent temperature [113]. Rainfall was included in 10.08% of the studies (n = 12), especially in studies from tropical or monsoon regions. It was modelled either linearly [55, 71]), as a standard deviation [129], using root-squared terms [143], or with splines [111, 137]. Sunshine duration was included in three studies [71, 72, 136], primarily as an auxiliary variable or adjustment factor. Cloudiness was used in one study [89], where it contributed to the computation of the physiological equivalent temperature (PET). Air pollution was featured as an adjustment covariate in 48.74% of the studies (n = 58). These studies incorporated concentrations of specific atmospheric pollutants known to influence health outcomes. The most frequently included pollutants were particulate matter with aerodynamic diameter ≤ 2.5 µm (; 17.65%, n = 21), ozone ( 14.29%, n = 17), nitrogen dioxide (; 13.45%, n = 16), particulate matter with aerodynamic diameter ≤ 10 µm (; 10.08%, n = 12), sulfur dioxide (; 6.72%, n = 8), and carbon monoxide (; 1.68%, n = 2). Pollutants were typically modelled as continuous covariates, either linearly or using spline transformations to allow for non-linear associations. In a smaller number of studies, multi-pollutant models were employed to explore collinearity or potential interactions between pollutants, such as × . For example, Cheng et al. [147] used hierarchical and joint effect models to estimate pollutant–temperature interaction terms. 2.52% of the studies (n = 3) used composite indices like the Air Pollution Index (), while 1.68% of the studies (n = 2) used the Air Quality Index () to summarise exposure to multiple pollutants. One study [48] additionally included air pollen data and metrics of environmental noise, such as equivalent continuous sound levels measured over 24 h, including daytime and nighttime periods. A detailed summary of all meteorological and environmental variables, including definitions and frequency of use, is provided in Table 4. Table 4Summary of meteorological and air pollution variablesVARIABLEDESCRIPTION% OF PAPERSREFERENCESRelative humidityRatio of actual water vapour in the air to the maximum possible at a given temperature48.74[41, 42, 45, 46, 48, 51, 55, 56, 58–60, 63, 66, 68, 71–73, 75–79, 81, 84, 85, 87, 89, 90, 92, 95, 97, 98, 104, 105, 108, 111, 112, 114, 115, 120–123, 125, 128, 129, 132–137, 143, 145–147, 149, 150]Air pressureForce exerted by the atmosphere at a given point; associated with weather variability and physiological responses11.76[48, 51, 59, 63, 66, 71, 72, 77, 84, 85, 87, 98, 134, 136]Sunshine hoursTotal duration of direct sunlight within a 24-h period2.52[71, 72, 136]Wind speedRate of air movement, affecting evaporative cooling and wind chill8.4[42, 71, 72, 78, 89, 111, 125, 128, 145, 146]RainfallTotal precipitation within a day; affects humidity, behaviour, and exposure patterns10.08[42, 55, 71, 72, 98, 111, 120, 128, 136, 137, 143, 146]CloudinessProportion of sky covered by clouds; used to approximate solar radiation and surface heat load0.84[89]Daily concentration of fine inhalable particles with diameter ≤ 2.517.65[48, 69, 74, 75, 84, 85, 90, 98, 104, 105, 111, 121, 126, 128, 133, 136, 139, 141, 143–145]Daily concentration of ground-level ozone, a secondary pollutant formed by photochemical reactions14.29[48, 51, 68, 69, 75, 79, 84, 85, 90, 98, 111, 121, 128, 133, 139, 141, 150]Daily concentration of nitrogen dioxide, a traffic-related gas13.45[41, 48, 68, 69, 79, 85, 89, 90, 92, 98, 111, 126, 133, 139, 141, 144]Daily concentration of inhalable particles with diameter ≤ 1010.08[41, 48, 51, 68, 69, 79, 89, 92, 115, 139, 141, 150]Daily concentration of sulphur dioxide, primarily from fossil fuel combustion6.72[41, 69, 79, 85, 89, 90, 98, 111]Daily concentration of carbon monoxide, a colourless and toxic gas interfering with oxygen transport in the blood1.68[69, 89]Air Pollution IndexAggregated index of air quality combining multiple pollutant concentrations into a single score2.52[71, 77, 87]Air Quality indexStandardised index of overall air quality, based on regulatory thresholds for key pollutants1.68[66, 147]Air pollenAirborne biological particles triggering allergic responses0.84[48]Equivalent Continuous Sound LevelEnergy-averaged sound pressure level over a time period; general measure of environmental noise0.84[48]Daytime Equivalent Continuous Sound LevelAverage noise level during daytime (typically 7:00–22:00)0.84[48]Nighttime Equivalent Continuous Sound LevelAverage noise level during night hours (typically 22:00–7:00)0.84[48] Summary of meteorological and air pollution variables Temporal adjustments: short-term, seasonality, and long-term trends Temporal covariates were included in the reviewed studies to account for short-term patterns, seasonal cycles, and long-term trends. These variables were used to adjust for time-structured influences that may confound the association between extreme temperatures and health outcomes. Table 5 provides a summary of all temporal variables used. Table 5Summary of studies accounting for temporal structureVARIABLEDESCRIPTION% OF PAPERSREFERENCESShort-termRefers to fluctuations occurring over brief periods, typically days or weeks, driven by immediate or transient factors88.24[12, 15, 23, 41–43, 45–51, 53–64, 66, 69–88, 90, 92, 93, 95–106, 108, 112, 114–116] [16, 117–128, 130–155]SeasonalityRepresents recurring cyclical patterns within a year, often influenced by climatic, ecological, or behavioural changes73.95[12, 15, 16, 23, 41–43, 45, 47–51, 53–57, 61, 62, 64–69, 71, 73, 76–79, 81–83, 85, 86, 91–93, 95, 96, 100–106, 108–110, 114, 116–122, 124–127, 130–141, 143–150, 152–155]Long-termCaptures gradual and sustained changes over extended periods, reflecting overarching social, environmental, or demographic shifts84.86[12, 15, 16, 23, 41–45, 47–51, 53–57, 59–63, 65–68, 70–73, 76–88, 91–93, 95–106, 108, 109, 112, 114–127, 130, 131, 133–149, 151–155] Summary of studies accounting for temporal structure Short-term indicators appeared in 88.24% of the studies (n = 105), including the day of the week (DOW), used in 76.47% of the studies (n = 91), and public holidays, used in 30.25% of the studies (n = 36). DOW was modelled as a categorical variable with seven levels. Public holidays were typically included as binary indicators, and in some cases (e.g., [57, 117]) both national and local holidays were considered separately. Week-of-the-year was used in 4.2% of the studies (n = 5) to account for periodicity not captured by other short-term indicators, modelled either categorically (1–52) or via spline-based representations. Seasonality was addressed in 73.95% of the studies (n = 88) using categorical indicators (e.g., season or month) or spline-based smoothing functions. For instance, Forceville et al. [148] and Cheng et al. [147] applied natural cubic splines on time with 4–7 degrees of freedom (df). Achebak et al. [123] implemented an interaction between day-of-season and year, and Corvetto et al. [150] used a 7 df spline selected via quasi-Akaike Information Criterion (qAIC). Long-term trends were accounted for in 84.87% of the studies (n = 101). Some studies modelled year as a categorical factor [106], while others used smoother functions with fewer degrees of freedom applied only to the year variable [44], capturing long term variation without explicitly adjusting for seasonality. In some multi-year studies, time was segmented into moving windows, such as in Kim et al. [125], which used 10-year sliding periods. In many studies, seasonal and long term trends were modelled jointly through flexible temporal structures. One common approach involved specifying an interaction between a natural cubic spline with four degrees of freedom for the day of the year and yearly indicator variables [91, 101]. Another frequently used strategy relied on a natural cubic spline applied to the time variable with eight degrees of freedom per year (for example [66, 100]), which allows both seasonal and multi-year patterns to be captured within a single smooth function. In addition, in studies adopting a case-crossover design, temporal variation was inherently incorporated through the time-stratified design or stratum definition, rather than by including explicit temporal covariates or spline terms (e.g., [133, 149]). Demographic, social, economic and health controls Demographic, social, and economic variables were involved in 36.14% of the studies (n = 43). Although this proportion may appear relatively low, it should be borne in mind that socioeconomic dimensions are often represented within the outcome definition or the stratification structure rather than among the model inputs. These variables were generally operationalised at the area level using census-derived data. In GLM-based models, they were incorporated as categorical or continuous predictors. For example, López-Bueno et al. [110] included the percentage of elderly residents, a rurality index, and unemployment rates in a mixed-effects model. Du et al. [113] used GDP per capita and urban population share as fixed effects in a panel regression. In several studies, population size was included as an offset term to standardise outcome counts into rates [57, 126, 142]. Some studies also combined multiple socioeconomic indicators in composite analytical frameworks; for instance, Masselot et al. [16] applied a Partial Least Squares (PLS) approach incorporating several socioeconomic variables to capture latent dimensions of vulnerability. Beyond their direct inclusion in regression models, the majority of studies included these indicators in multi-stage analyses as meta-predictors for meta-analysis, most often represented by GDP per capita (e.g., [140, 151]) to identify contextual effect modifiers, assess population vulnerability, and enhance the generalisability of pooled estimates. Health-related indicators appeared in 15.12% of the studies (n = 18) and were included to represent underlying vulnerability or healthcare system characteristics. In Maier et al. [44], a binary indicator for SARS outbreaks was used as a covariate. Ren et al. [107] included hospitalisation rates, average length of stay, and outpatient visits as continuous variables. Du et al. [113] used the WHO IHR Core Capacity Score, and Zeng et al. [129] included hospital bed density and staff ratios as covariates and interaction terms. In DLNM-based designs, health variables were sometimes modelled as time-varying confounders. Martínez-Solanas et al. [80] included daily influenza hospitalisations, and Sharma et al. [128] incorporated pneumonia and influenza cases using splines. Also in this case, the majority of studies included health-related indicators as meta-predictors in multi-site analyses, typically to represent differences in healthcare capacity or population health status across locations proxied through indicators such as number of hospital beds per 10,000 inhabitants (e.g., [77]) or number of hospitals for 10,000 inhabitants (e.g., [78]). Table 6 summarises the use of these variables. Table 6Summary of demographic, social, economic and health controlsVARIABLEDESCRIPTION% OF PAPERSREFERENCESDemographic and socio-economicVariables describing the population structure and socioeconomic context, such as age, sex, education, income, employment, urbanisation, and population size (including when used as an offset)36.13[12, 16, 44, 45, 51, 52, 55, 57, 64, 68–70, 76, 80, 81, 83, 84, 87, 94, 100, 102–105, 107, 110, 113, 115, 119, 122, 123, 126, 129, 130, 132, 134, 135, 140, 142, 151, 153–155]HealthVariables related to population health status and healthcare system capacity, including comorbidities, mortality rates, hospital admissions, medical staff availability, and number of hospital beds15.12[12, 16, 44, 70, 74, 76, 77, 87, 107, 113, 123, 128–130, 135, 151, 153, 154] Summary of demographic, social, economic and health controls Built-environment controls Built environment indicators were included in 15.12% of the studies (n = 18) to account for spatial heterogeneity in exposure intensity and adaptive capacity. These variables captured structural characteristics such as building density, housing conditions, thermal inertia, and levels of urbanisation. They were incorporated either as covariates in regression models or as effect modifiers in spatially resolved designs. For example, Song et al. [94] used urban density, vegetation cover, and housing quality in a multiscale geographically weighted regression to assess heat-related mortality across South Korean cities. Similarly, López-Bueno et al. [110] included insulation quality and urban–rural classification in a mixed-effects model examining cold-related mortality in Spain. Beyond their inclusion in local models, the majority of multi-site studies also used built environment indicators as meta-predictors in meta-regression analyses (e.g., [123, 151]), typically to explain spatial variation in exposure–response functions associated with differences in urban structure or green space availability. Table 7 reports the studies involving these variables. Table 7Summary of built environment controlsVARIABLEDESCRIPTION% OF PAPERSREFERENCESBuilt environmentVariables capturing physical and infrastructural features of the urban or residential setting, such as housing quality, thermal insulation, land use, green space (e.g., NDVI), building density, and access to transport or public services15.12[16, 44, 52, 57, 68, 80, 83, 87, 91, 94, 110, 115, 123, 130, 135, 151, 153, 154] Summary of built environment controls Discussion Key input variables and model selected This review identifies three main categories of variables that are essential for ensuring the reliability of epidemiological models assessing temperature-related health effects: thermal variables, environmental and air pollution covariates, and temporal adjustments. Temperature represents the core climatic determinant in epidemiological models because of its direct link to physiological and thermoregulatory stress [158–160]. Among temperature-based indicators, mean air temperature is the predominant exposure metric, as it captures the integrated effect of daily thermal conditions. Although a limited number of studies have adopted maximum or minimum temperature measures (e.g., [96, 127]), these alternatives remain relatively uncommon. The selection of the temperature metric is inherently context-dependent, reflecting population characteristics, the health outcomes investigated, and local climatic regimes [161, 162]. Nevertheless, because mean, maximum, and minimum temperatures are typically highly correlated, differences in estimated exposure–response functions are generally modest [163]. Hence, the preference for mean air temperature lies in its robustness and its ability to represent overall thermal exposure, thereby facilitating comparability across heterogeneous climatic and demographic settings. Only a minority of studies employ composite indices such as AT or WBGT, which incorporate additional environmental variables including humidity, radiation, and wind. While these indices provide a more physiologically grounded representation of thermal exposure, their use is constrained by the need for more complex data inputs and calculations, limiting their feasibility in large-scale or multi-site studies. Notably, only one study included in our review [74] directly compared mean, maximum, minimum temperature, AT, and WBGT, finding that WBGT yielded more consistent associations with health outcomes during heat events. However, a recent multi-country analysis by [164] extended this comparison to eight heat-stress metrics across 604 cities in 39 countries using two-stage DLNMs, showing that the optimal metric varies geographically but that differences in heat-attributable mortality are generally modest. This supports the continued use of mean air temperature as a robust and comparable exposure metric in multi-site epidemiological studies. A final consideration is that the use of composite indices may discourage the separate inclusion of individual meteorological covariates, such as relative humidity, which are already embedded within the index. While this approach simplifies model specification and reduces concerns about collinearity, it may obscure the individual contributions of specific environmental factors, thereby limiting both the interpretability of results and their relevance for targeted policy interventions [165]. Environmental covariates are essential components of epidemiological models assessing temperature-related health effects, as they influence both the direct physiological response to thermal exposure and the broader environmental context in which vulnerability arises. Restricting models to temperature alone may compromise the validity of risk estimates by overlooking co-exposures that can either amplify or buffer the health impacts of heat [166, 167]. Non-thermal meteorological variables and air pollutants interact with temperature in complex ways, modulating thermoregulation, cardiorespiratory stress, and overall health risk [168–170]. Their inclusion enables a more accurate representation of exposure conditions, supports the identification of modifying effects, and enhances both the ecological validity and causal interpretability of model outputs [171]. Interestingly, meteorological modifiers are more frequently incorporated into models than air pollution covariates. This tendency likely reflects a combination of practical and methodological considerations. Relative humidity, for instance, is widely recognised as a critical modifier of thermal stress and is routinely included even in parsimonious modelling designs. In contrast, pollutants such as and ozone are often excluded, particularly in multi-area studies, primarily due to data limitations (e.g., [12, 64, 106]). Furthermore, some models deliberately omit additional environmental exposures to isolate the specific effects of heat, thereby reducing concerns over confounding and facilitating cross-study comparability. As a consequence, although both meteorological and pollution covariates are theoretically important, modelling practice tends to prioritise the former—especially humidity—while acknowledging the latter as a limitation where data are lacking. It is important to emphasise, however, that a substantial proportion of studies omit both types of covariates entirely, raising significant concerns regarding omitted variable bias. This is particularly problematic in regions with high pollutant burdens, where failure to account for air quality may lead to substantial distortions in estimating the true effects of thermal exposure (e.g., Po valley in Italy) [23, 24]. The inclusion of temporal covariates is essential in time-series analyses to isolate the specific effects of thermal exposure from broader temporal dynamics in health outcomes [172]. In contrast, within case-crossover designs these factors play a secondary role, as temporal matching is intrinsic to the framework, thereby reducing the need for additional temporal covariates. Mortality and morbidity data typically exhibit strong temporal patterns that, if unaccounted for, may give rise to spurious associations. As highlighted by Bhaskaran et al. [173], raw health outcomes are influenced by seasonal cycles, long-term trends, and short-term behavioural fluctuations, all of which can either obscure or exaggerate the observed effects of temperature. For instance, if extreme heat events predominantly occur during specific seasons, failure to adjust for seasonality may result in misattributed impacts [174]. A further methodological consideration concerns the consequences of the omission of a certain variable. Evidence across the reviewed studies suggests that not all non-thermal modifiers contribute equally to potential distortions in temperature–health estimates. Variables such as wind speed or solar radiation, as well as the choice among alternative temperature metrics (mean, maximum, minimum), tend to have only marginal influence on the estimated associations. Their omission typically affects model precision rather than the underlying exposure–response function. This is consistent with findings from several studies in which air pollutants were included only in sensitivity analyses and produced negligible changes in the estimated temperature effects (e.g., [49, 140]) or were retained solely for robustness checks without substantially altering the main results (e.g., [101]). In contrast, humidity tends to play a more influential role, although a few studies reported negligible changes when it was included only as part of sensitivity analyses (e.g., [119, 153]). Recognising this differentiation helps clarify which covariates primarily act as precision enhancers and which are essential for isolating the independent health impacts of temperature. Across the reviewed literature, temporal adjustment is widely applied and relatively standardised. Short-term fluctuations are commonly accounted for through day-of-week indicators (e.g. [55, 97, 150]), often supplemented by public holiday variables (e.g. [127, 147]), reflecting a broad consensus on the need to control for daily behavioural and healthcare-related variations. However, short-term indicators do not contribute equally to model specification. In our framework emerges that DOW is the most consistent influential factor, capturing stable weekly patterns in healthcare utilisation and population behaviour. In contrast, public holidays cover a very limited number of days per year and are used far less frequently; consequently, their contribution tends to be modest and may be negligible in many contexts. This asymmetry suggests that, while short-term adjustments are routinely included, only a subset plays a meaningful role in shaping model dynamics. Long-term trends are typically modelled using natural cubic splines [98, 106, 126] capturing structural changes such as demographic shifts or improvements in healthcare systems [175]. In contrast, approaches to seasonality adjustment display greater heterogeneity, ranging from harmonic functions to splines or categorical month indicators, each aiming to represent intra-annual variations in climate, activity patterns, and susceptibility [176]. Many studies, particularly within the MCC Collaborative Network, streamline this process by employing a single natural cubic spline for both long-term and seasonal components (e.g., [102, 103]) or by including interaction terms between season and year (e.g., [91, 101]). Despite this heterogeneity, all approaches converge on the need to account for both seasonal and long-term structures, which remain essential for separating temperature effects from broader temporal dynamics. The diversity of implementations therefore reflects methodological flexibility rather than conceptual disagreement on the necessity of these components. Studies employing DLNMs ([86, 93, 146]) tend to adopt a more systematic and comprehensive approach to temporal adjustment, routinely incorporating short-term, seasonal, and long-term dynamics within an integrated modelling framework. Among the modelling approaches employed in the reviewed literature, DLNMs are consistently the most widely adopted for assessing temperature-related health effects. Its popularity derives from its ability to simultaneously capture non-linear exposure–response relationships and lagged effects, while maintaining flexibility across a range of modelling frameworks [7]. The DLNM is routinely implemented within GLM, GAM, conditional Poisson models, and conditional logistic regression, enabling its application in both time-series and case-crossover studies [76, 87, 126, 143]. Although dominant, the DLNM is not the not the only approach observed in the reviewed studies. Some studies utilise statistical models without explicit distributed lag structures [92, 115, 142], typically in contexts involving shorter exposure windows or simpler analytical settings. A limited number of studies employ alternative strategies, such as difference-in-differences designs, particularly when evaluating the effects of discrete heatwave events or specific policy interventions [129]; however, these approaches remain marginal and highly context-dependent. Overall, the DLNM has become the methodological standard in the field, offering a robust and adaptable framework capable of accurately reflecting the complex, non-linear, and delayed relationships that characterise the ET-health nexus. Integration of socioeconomic, demographic, and environmental covariates Socioeconomic, demographic, health, and built environment variables are most frequently used to stratify results rather than being incorporated as integral model inputs. When included as covariates, their role remains limited and often context specific, as reflected by the relatively small number of studies that explicitly integrate them into first stage models (e.g., [110, 113]). This is particularly true within the DLNM framework, where such factors are more commonly introduced at later analytical stages, mainly through meta regression, to account for between location heterogeneity in temperature–health relationships (e.g., [16, 105]). These contextual determinants, encompassing indicators of income, education, age composition, healthcare capacity, and urban form, represent key social and infrastructural dimensions of vulnerability. Their integration into model design is essential for moving beyond a purely climatic interpretation of the temperature–health nexus, allowing for a more comprehensive understanding of differential exposure, sensitivity, and adaptive capacity across populations. By linking thermal risk to broader structural conditions, such integration enhances both the interpretability of pooled models and their relevance for health equity and policy translation [177]. Nevertheless, significant barriers persist. Contextual datasets are often fragmented, temporally inconsistent, and compiled at disparate spatial scales, hindering their systematic inclusion within modelling frameworks. In the absence of city level data, national or regional averages are frequently employed as proxies for local socioeconomic or infrastructural conditions. While this practice maintains analytical feasibility, it tends to smooth spatial variability and obscure localised patterns of vulnerability, thereby weakening the capacity of multi-site frameworks to capture genuine heterogeneity. Bridging these gaps will require coordinated efforts to integrate epidemiological, socioeconomic, and environmental information through harmonised, multi domain infrastructures. Such interdisciplinary integration is not merely technical; it is critical to identify the structural drivers of population vulnerability to extreme temperatures and to support the development of equitable and evidence based adaptation strategies. Geographical scope and spatial units Our review highlights a structural divergence in model input specification between studies conducted in high-income or rapidly industrialising countries and those based in lower-income or data-scarce regions. Countries such as China, Australia, and the United States dominate the literature, benefiting from greater access to high-resolution meteorological and health data. Although the MCC Collaborative Network has substantially expanded the empirical base by including several low- and middle-income countries in its multi-country and multi-city time-series analyses, data from these settings remain difficult to access and harmonise. As a result, outside the MCC framework, studies focusing on these regions are still scarce and often limited in scope or data quality. However, data availability alone does not dictate model complexity defined here as the diversity and number of environmental covariates included. Instead, input selection reflects a broader interplay of infrastructural capacity, methodological traditions, and institutional priorities. In Europe, studies tend to adopt standardised input structures, typically including air temperature, relative humidity, and basic temporal controls. While this facilitates cross-study comparability, it may obscure local epidemiological nuances, especially in urban settings where thermal stress results from multiple interacting exposures. Key modifiers such as wind speed, solar radiation, and air pollutants are often omitted, even when epidemiologically relevant. For example, studies from the United Kingdom and Spain frequently exclude these variables despite their likely importance in polluted or densely built environments [48, 73, 112]. East Asian studies display greater heterogeneity in input specification. In cities such as Seoul and Hong Kong, input sets often integrate multiple pollutants and meteorological covariates [68, 90, 109, 111], reflecting enhanced environmental monitoring capacities and methodological openness to complexity. However, this is not universal. Some studies, like Han et al. [56] in Jinan, omit air pollution data due to data gaps, while Du et al. [108]excludes it after sensitivity analyses found negligible effects underscoring the need to distinguish parsimony driven by necessity from that based on deliberate analytic choices. In North America, particularly the United States, models tend to adopt streamlined input structures, often reflecting the practical challenges of harmonising environmental data across large-scale, multi-city studies [127, 132]. Streamlined designs, as seen in Hoehne et al. [64] and Quilty et al. [138], promote comparability and internal validity but may overlook critical modifiers such as air quality or wind dynamics. Australian studies frequently incorporate detailed air pollution metrics [69, 79], yet often omit other meteorological variables like wind or solar radiation. This pattern suggests that beyond infrastructural constraints, national research cultures and policy priorities also shape input selection practices. In low- and middle-income countries, input choices are predominantly shaped by infrastructural limitations. Studies from Africa and Latin America often omit environmental covariates not by design but due to limited access to high-quality data [65, 96]. Nevertheless, where data are available, more complex models emerge, for instance, Zhao et al. [78] includes wind speed and rainfall in its exposure model. In such contexts, parsimony reflects necessity rather than epidemiological appropriateness, potentially underestimating population vulnerabilities. Ultimately, geographical variation in input specification reflects not only the availability of data but also the values and priorities that guide climate-health research, with important implications for the equity and generalisability of the global evidence base. The core gap and future direction Three core gaps emerge from the current literature on the ET-health nexus. The first concerns the substantial heterogeneity in model input structures, which significantly undermines comparability across studies. While DLNMs are widely used and provide a coherent analytical framework, the choice of covariates, especially environmental and socio-demographic ones, varies greatly. As a result, the ET-health nexus is often adjusted for different factors depending on the study context, making cross-study comparisons problematic. Greater standardisation in the inclusion of input variables would help address this issue. Despite this, any attempt to harmonise modelling practices must also account for local specificities, including both structural features of the environment and behavioural patterns of the exposed population. The second and more fundamental issue is the absence of a clearly defined minimum model specification capable of ensuring reliable and replicable estimates. Although many studies using DLNMs report sensitivity analyses such as varying degrees of freedom or selectively including covariates there remains considerable heterogeneity in how models are built. In many cases, this reflects data scarcity. Yet it is not sufficient to assume that estimates from under-resourced regions are valid in the absence of evidence that the chosen specification meets even a minimal adequacy threshold. Without clearer methodological benchmarks, we cannot confidently assess the robustness of results, especially in studies designed to inform public health action. These challenges, including model heterogeneity, limited data availability and difficulties in implementation, may partly explain the recent shift toward more accessible and standardised approaches, such as the use of the IPCC risk assessment framework, which facilitates replication and comparability across diverse settings [24, 178, 179]. The third gap concerns disparities in data availability and accessibility. In high-income settings such as the European Union, open-access hazard data are increasingly accessible thanks to platforms like Copernicus [180], which provide climate variables at multiple spatial and temporal resolutions. In contrast, in many low- and middle-income countries, particularly in the Global South, even basic hazard data sets remain incomplete or unavailable. In addition, health data are often protected by strict privacy regulations or fragmented across different institutions, making integration into modelling frameworks extremely challenging. These data asymmetries limit both the spatial representativeness and methodological replicability of existing studies, raising critical issues for global equity and climate-health research agendas. Towards a core model specification To address the issue of heterogeneity, we propose a minimal set of input variables that could serve as a baseline specification for future short-term time-series epidemiological models, particularly those employing DLNMs, assessing the health impacts of extreme temperatures. This "core input structure" aims to balance epidemiological robustness, feasibility across diverse contexts, and the need for standardisation. The proposed baseline inputs are summarised in Table 8. Table 8Proposed core input structure for DLNM-based models assessing temperature–health associationsCATEGORYSUGGESTED VARIABLERATIONALEThermal exposureDaily mean temperatureWidely available; good proxy for overall thermal stressNon-thermal meteorologicalRelative humidityKey modifier of thermal perception and heat-related health outcomesAir pollution (if available)Major pollutantsImportant independent confounders influencing cardiorespiratory and all-cause health risks. their influence can be context-specific and limited by data availabilityTemporal adjustmentDay of the week, seasonality term, long-term trend splineEssential to control for short-term, seasonal, and long-term trendsIn case-crossover settings the temporal adjustments are not required as the design intrinsically controls for seasonality and long-term trends through time-stratified matching Proposed core input structure for DLNM-based models assessing temperature–health associations In case-crossover settings the temporal adjustments are not required as the design intrinsically controls for seasonality and long-term trends through time-stratified matching It is important to emphasise that this proposal represents a minimal specification intended to facilitate cross-study comparability, rather than a rigid framework. Alternative modelling choices, such as the use of different thermal exposure indicators (e.g., maximum temperature, apparent temperature, wet-bulb globe temperature) remain valid and appropriate when epidemiologically justified and consistently motivated. This core input structure reflects the minimal set of covariates necessary to adjust for confounding and temporal dynamics in time-series DLNM-based designs. For thermal exposure we suggest the use of mean temperature. Its broad availability, computational simplicity, and stable representation of overall daily thermal conditions make it a practical and widely comparable baseline indicator. Maximum and minimum temperatures are strongly correlated with mean values and may exaggerate specific segments of the exposure distribution without improving estimation. Composite indicators can offer conceptual advantages [74] but their added value for effect estimation is generally limited [163], so mean temperature remains the most efficient and comparable choice for temperature-health related analysis. For the non-thermal meteorological variable we suggest relative humidity as it is the most widely adopted measure of atmospheric moisture, mainly due to its broad data availability, historical use in epidemiological research, and comparability across studies. However, we also have to highlight that it may not always represent the most physically consistent metric. Absolute or specific humidity better reflects the actual vapour content in the air and often yields more stable and monotonic associations with health outcomes [181]. Nonetheless, relative humidity remains an appropriate choice in season specific analyses, where temperature variability is narrower and its physical interpretation is internally consistent within each seasonal context. The inclusion of air pollution variables such as , , , and may be useful where reliable data are available, its impact on heat- or cold-related risk estimates is not always substantial [182], their inclusion should therefore be guided by local environmental profiles and data accessibility. Additional environmental variables such as wind speed, solar radiation, or air pressure typically operate as effect modifiers rather than independent confounders. Their inclusion may improve model fit or exposure characterisation, but their influence on the ET-health nexus is often marginal and context-specific. A key distinction should also be noted across analytical designs. Ordinary time-series models (GLM, GAM, DLNM) require explicit temporal controls to capture seasonality and long-term trends, whereas case-crossover designs inherently account for these through their matching structure, making additional temporal covariates unnecessary. Overall, this flexible core input structure should be understood as a methodological benchmark that promotes transparency, comparability, and reproducibility, while recognising the heterogeneity of data environments and epidemiological contexts that characterise temperature–health research. Strengths and limitations This review presents several strengths that contribute to its methodological robustness and its value for advancing research on the ET–health nexus. First, to our knowledge, this is the first systematic review to focus specifically on the role of input variables in epidemiological models, rather than on outcome measures or effect sizes. By addressing how exposures, confounders, and effect modifiers are defined and integrated into model structures, this research fills a critical gap in the literature and proposes a new dimension for comparative assessment. Second, the review follows a transparent and rigorous methodology consistent with PRISMA standards, including multi-database searches, multi-stage screening, and explicit eligibility criteria, which support the replicability and credibility of its findings. Third, the dataset spans a broad range of geographical contexts—including studies from Asia, Europe, North and South America, and Africa—thereby capturing the diversity of modelling practices across climatic and institutional settings. Fourth, the functional classification of model inputs—based on their role in statistical estimation—provides a coherent and replicable framework that enhances comparability across studies and may guide future modelling efforts. Nonetheless, this study has several limitations. First, despite the comprehensiveness of the search strategy, the extensiveness of the review process may have unintentionally led to the exclusion of some relevant studies. This is particularly plausible in cases where the input structure was not clearly documented or where studies employed hybrid modelling frameworks that fell outside predefined eligibility boundaries. However, the characteristics of the excluded studies suggest that their omission is unlikely to materially affect the main patterns and interpretations presented in this review. Second, the review does not formally distinguish between simplified model structures adopted for analytical parsimony and those resulting from limited data availability. This ambiguity may affect the interpretation of model adequacy, particularly when comparing studies conducted in high- versus low-resource settings. Third, the absence of a quantitative synthesis of effect estimates such as a meta-regression, limits the ability to assess how different combinations of input variables influence the strength or direction of the reported associations. Fourth, although input variables were systematically classified, the coding process was conducted manually and based on the information reported in the articles, which introduces a potential source of subjectivity, especially when model specifications were incomplete or ambiguously described. Fifth, while a comparison with studies applying the IPCC risk assessment framework would have been valuable, especially given its relevance in low-data settings this was beyond the methodological scope of a systematic review, which requires predefined and focused research questions. Such integration may be more appropriate for other types of reviews, such as scoping or narrative reviews. Lastly, this review did not include output variables; these are instead addressed in a follow-up study, which focuses exclusively on outcome definitions and their implications for estimating health effects. Conclusion This systematic review examined how epidemiological models investigating ET-health nexus are specified with respect to both input selection and modelling strategies. By focusing not on outcome heterogeneity but on the structural components of model design (particularly the inclusion, classification, and operationalisation of input variables) this study offers a novel and methodologically grounded contribution to the growing field of climate-health epidemiology. It advances the understanding of how choices made “on the right-hand side” of the model equation shape the robustness, comparability, and policy relevance of resulting evidence. The analysis of 119 studies published between 2014 and 2024 reveals a number of recurring practices, alongside substantial inconsistencies. Daily mean temperature was the most common exposure metric, though rarely justified over alternatives. Relative humidity was frequently included, while other environmental and pollution variables were inconsistently used and poorly motivated. Temporal covariates were generally standardised, but details on parameter choices were often lacking. Socio-demographic and built environment indicators appeared mainly in non-DLNM studies whereas in DLNM-based frameworks they were more commonly introduced in later analytical stages, particularly through stratified analyses or meta-regressions aimed at explaining between-location heterogeneity. These differences highlight a persistent gap: socio-economic and contextual determinants, although central to explaining who is most at risk and why, remain marginal in the current modelling practice, constraining the field’s ability to inform equitable adaptation policies. Strengthening the integration of socio-economic and infrastructural variables is therefore not only a methodological refinement but a conceptual shift toward a more equitable understanding of climate–health relationships. Incorporating such inputs allows models to capture differential exposure, sensitivity, and adaptive capacity across populations, linking epidemiological evidence to the social structures that drive vulnerability. Embedding this perspective within modelling frameworks would enhance the relevance of ET-health nexus research for targeted adaptation and public health equity. Additional insights emerged regarding geographical variation, with input structures reflecting differences in data availability and institutional priorities. As input specification directly influences the estimation of health effects, this review provides the necessary foundation for Part B, which explores how these modelling choices shape outcome definitions, effect estimates, and the broader conclusions drawn in the ET–health literature. These findings highlight a structural problem: the absence of a shared framework for input selection and model construction. More critically, the field still lacks a clear understanding of what constitutes a minimum model specification—that is, the set of core adjustments and structural components required to ensure the reliability of estimated effects. In many cases, simplified models are accepted without evidence that their specification meets minimal adequacy thresholds. This gap in methodological guidance undermines the comparability of findings, increases the risk of biased estimates, and complicates efforts to synthesise evidence for policy purposes. Future research should prioritise the development of consensus-based guidelines for input selection and model construction in the temperature–health domain. Such guidance should strike a balance between epidemiological rigour and pragmatic feasibility, particularly in settings where data availability is limited. This includes defining a core set of inputs (thermal, environmental, temporal, and socio-demographic) that should be routinely considered, while also allowing for local adaptation. Equally important is the need for greater transparency in modelling decisions, including formal sensitivity analyses and explicit reporting of data constraints. Only through harmonised, methodologically transparent practices can the field produce robust, comparable, and policy-relevant evidence capable of informing adaptation strategies in a warming world. Supplementary Information Supplementary Material 1 Funding This study was supported by Ministry of Health within the PNC "HEALTH AND EQUITY CO-BENEFITS IN SUPPORT OF CLIMATE CHANGE RESPONSE PLANS IN ITALY" project—CUP J55I22004450001. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1