Explainable Artificial Intelligence and Machine Learning for Air Pollution Risk Assessment and Respiratory Health Outcomes: A Systematic Review

Abstract

Air pollution is a leading environmental risk that causes respiratory morbidity and mortality. The increasing availability of high-resolution environmental data and air pollution-related health cases have accelerated the use of machine learning models (ML) to estimate environmental exposure–response relationships, forecast health risks and call for the needed policy and practical interventions. Unfortunately, ML models are opaque, in a sense that, it is unclear how these models combine various data inputs to make a concise decision. Thus, limiting its trust and use in clinical matters. Explainable artificial intelligence (xAI) models offer the necessary techniques to ensure transparent and interpretable models. This systematic review explores online data repositories through the lens of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guideline to synthesize articles from 2020 to 2025. Various inclusion and exclusion criteria were established to narrow the search to a final selection of 92 articles, which were thoroughly reviewed by independent researchers to reduce bias in article assessment. Equally, the ROBINS-I (Risk Of Bias In Non-randomized Studies of Interventions) domain strategy was helpful in further reducing any possible risk in the article assessment and its reproducibility. The findings reveal a growing adoption of ML techniques such as random forests, XGBoost, parallel lightweight diagnosis models and deep neural networks for health risk prediction, with SHAP (SHapley Additive exPlanations) emerging as the dominant technique for these models’ interpretability. The extremely randomized tree (ERT) technique demonstrated optimal performance but lacks explainability. Moreover, the limitations of these models include generalizability, data limitations and policy translation. This review’s outcome suggests limited research on the integration of LIME (Local Interpretable Model-Agnostic Explanations) in the current ML model; it recommends that future research could focus on causal-xAI-ML models. Again, the use of such models in respiratory health issues may be complemented with a medical professional’s opinion.

1. Introduction

Air pollution impacts people’s health globally, and it is considered one of the leading risk factors for premature death worldwide [1]. Again, it is a major environmental and public health challenge in many countries [2]. Air pollution can severely impact the lungs of children and adults; however, it is often not considered a priority in the public health sector by policymakers in some developing countries [3]. Sankar et al. [4] indicate that the decline in air quality affects major cities, which negatively impact human health. One common air pollutant is particulate matter (PM_2.5), which suggests the need for an optimal and precise model that could be supported by a policy document.

Respiratory health outcomes refer to the state of the respiratory system, which involves the presence or absence of any disease, and it likely impacts an individual. Health outcome ranges from normal lung function to various respiratory illnesses and diseases, which are often measured using lung function tests, symptom reporting (like coughing and wheezing) and the presence of diagnosed respiratory conditions. Furthermore, respiratory illnesses and diseases can be in the form of asthma attacks, chronic obstructive pulmonary disease (COPD), respiratory infections and many more. Respiratory abnormality often puts a significant health burden on the individual, which requires prompt intervention from a health facility [5]. Notably, air pollution is a well-known contributor to asthma. Furthermore, air toxics are hazardous air pollutants with serious health effects [6].

Respiratory health outcome models are analytical models for quantifying the relationship between environmental exposures (especially air pollutants) and respiratory events. Respiratory health outcomes can significantly affect a person’s quality of life, impacting their ability to perform daily activities, exercise and participate in social events. Understanding respiratory health outcome is crucial to developing strategies to prevent respiratory illnesses and effectively manage existing conditions to improve the overall well-being of a person. Human exposure to air pollutants such as particulate matter, sulfur dioxide, ozone (O₃) and carbon monoxide is known to contribute to an adverse health hazard, which results in respiratory diseases, cardiorespiratory diseases, cancers, and worst of all, it can lead to death [7].

With the acceleration of industrialization, industrial smoke particles containing complex chemical compositions and varying particle sizes pose a serious threat to the environment and human health [8]. While AI-based models are increasingly being used in assessing the risks of exposure to hazardous air pollutants, the quality of a model’s decision is a subject for open debate. Arguably, users of AI-based models are more interested in how and why an AI model made a particular decision relative to a respiratory health outcome. Understanding these dynamics would help in making the choice of an AI technique for environmental and health impact assessment systems and, in some instance, help understand the need to combine federated learning models in impact-based health assessment models. Due to the proliferation of AI models in different contexts, model generalization is problematic and this could lead to the lack of trust in those generalized model in practical health cases [9]. AI models for respiratory impact assessment could include some elements of environmental variables such as the air quality index (AQI) and other meteorological variables [7]. Unfortunately, the geographical positioning of a country implies different underlying weather conditions, which makes model generalization challenging.

AI models are becoming a common commodity in every sector, from health to manufacturing to education, mining and many more. Such AI-driven models embody machine learning (ML) and deep learning (DL), which are employed in robotic-assisted surgery, thus helping to optimize the clinical workflows and improve patient outcomes [10] in categories of health outcomes such as any health symptoms, asthma, allergy and flu-like symptoms [11]. Machine learning models play a key role in assessing and predicting these health-related outcomes [11]. Examples of such models include logistic regression (LR) and random forest (RF) models.

The current dispensation of the integration of xAI models into ML/AI models called for a thorough review of machine learning models in air pollution risk assessment and respiratory health outcomes. Such a review is beneficial to health policymakers and health-related AI practitioners because it suggests the most optimal model that equally explains the how and why behind AI models’ decision relative to air pollution exposure risk and respiratory health outcomes. Furthermore, it benefits clinicians by giving them an alternative strategy to confirm their clinical health outcomes, thus mitigating any possible clinical negligence that comes with legal consequences. Thus, this research aims to systematically review the extant literature on xAI integration with ML models for air pollution risk assessment for respiratory health outcomes. This review contributes to the development of optimal data-driven models for the health sector, and the foresight gained from different countries are shared with the least developed ones, in terms of respiratory health outcomes and air pollution assessment models. In so doing, machine learning models can be tailored and trained on different sets of data other than dataset on air pollution and meteorological factors. Thus, this study bridges the research gap by finding the most optimal machine learning model for air pollution risk assessment and respiratory outcomes.

The optimality in this context refers to the accuracy of a model’s prediction. More importantly, policymakers are concerned about the model’s prediction accuracy to avoid negative health effect outcomes that may result in respiratory diseases, cardiorespiratory diseases, cancers, and worst of all, it can lead to death. In this paper, we attempt to systematically review recent xAI and machine learning models for air pollution risk assessment and respiratory health outcomes. The sections are organized as follows: Section 2 describes the methods and materials used for this review, Section 3 describes the results and discussion, and Section 4 concludes the systematic review.

2. Materials and Methods

In recent years, researchers have applied different methodologies to examine the extant literature on different contexts of air pollution. It is imperative to understand these methodologies to know the gaps to fill in our study.

Table 1. Methodological frameworks and research focus.

Author(s)	Research Focus	Methodology	Data Repositories	Number of Articles Selected	Year Range
Chadalavada et al. [12]	Monitoring and prediction.	Systematic review: PRISMA guidelines.	PubMed (43), Scopus (224) and Web of Science (145).	64	2019 to 2023
Masood et al. [13]	Application, fundamentals and performance.	Systematic review based on the three-phased approach suggested by [14]: plan, conduct review and report the review.	Total of 300 articles from Scopus, Web of Science, Ei Compendex, PubMed and ScienceDirect, Harvard University library catalog (HOLLIS).	90	2003 to 2021
Subramaniam et al. [15]	Forecasting and human health.	Classic narrative review that follows the systematic review technique.	Google Scholar, Springer, PubMed, ScienceDirect, IEEE explore, Scopus and Web of Science (WoS).		January 2015 through June 2022
Liu et al. [16]	Air pollution and health.	Bibliometric techniques.	-	14,955	2001 to 2021
Leivaditis et al. [10]	Applications, benefits, limitations and future direction of AI focusing on thoracic surgery.	PRISMA and risk of bias assessment were conducted using the Cochrane Risk of Bias Tool and ROBINS-I for non-randomized studies.	PubMed, Scopus, WoS and Cochrane Library.	36	Studies published up to January 2025
Ye et al. [17]	Measurement methods of greenspace, their health outcomes and potential mechanisms.	Systematic review.	MEDLINE and EMBASE.	140	Up to 11 April 2022
Vachon et al., 2024 [18]	Compared statistical and machine learning models to predict NO2, ultrafine particles (UFPs) and black carbon (BC).	Systematic review.	WoS and Scopus.	38	Up to 13 June 2024

presents recent methodological frameworks and research focus.

As summarized in Table 1, prior reviews have focused on diverse but fragmented aspects of the field. Chadalavada et al. [12] concentrated on monitoring and prediction, while Masood and Ahmad [13] provided a broad methodological overview. Subramaniam et al. [15] connected forecasting to human health outcomes, and Liu et al. [16] emphasized bibliometric trends rather than methodological synthesis. Leivaditis, Maniatopoulos [10] narrowed the focus to applications in thoracic surgery, while Ye et al. [17] concentrated on greenspace measurement and health impacts. More recently, Vachon et al. [18] compared statistical and machine learning models for some selected air pollutants.

In contrast, our systematic review makes a distinct contribution by integrating these fragmented perspectives into a unified synthesis that explicitly connects methodological developments, application contexts, and performance outcomes. Unlike bibliometric or domain-specific reviews, our work (i) develops a new comparative framework across spatiotemporal and AI-based approaches, (ii) highlights underexplored intersections between pollution risk assessment, respiratory health-related outcomes, and transparent decision-making applications, and (iii) identifies three cross-cutting gaps in the literature: limited integration of xAI with spatiotemporal deep learning models, despite their promise for fine-grained air pollution forecasting; weak linkage between pollutant risk assessment and downstream decision support, such as exposure-aware routing, early warning systems or policy guidance—most reviews treat prediction and application separately; and lack of standardized evaluation protocols and cross-domain validation, making it difficult to compare models across regions, pollutants, and health outcomes.

By systematically addressing these gaps, our review not only consolidates existing knowledge but also extends beyond prior studies by proposing an integration of xAI in ML air pollution assessment models that bridges methodological advances with real-world implementation. Thus, our study adopts the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines to synthesis the most recent, relevant and related literature from 2020 to 2025. Furthermore, registration information is not applicable in this study. The inclusion criteria include the following: peer reviewed articles, conference papers, reviews, publication years and written in the English Language. The exclusion criteria include the following: book chapters, retracted papers, observational and animal studies. Figure 1 illustrates the PRISMA flow diagram detailing the systematic process: identification, screening and inclusion.

The following steps were undertaken to ensure quality assessment of the reviewed research article.

2.1. Research Questions

The research questions underpinning this systematic review are as follows:

• What machine learning techniques are used for air pollutants risk assessment and respiratory health outcomes?
• To what extent is xAI integrated into machine learning models for air pollution risk assessment and respiratory health outcomes?

While these research questions provided the hypothetical basis for the systematic review, an initial search keyword was conducted to explore the existing literature in the online repository.

2.2. Literature Search Strategy

This review followed the PRISMA guidelines, and the search strategy ensured a thorough article search through online data repositories: PubMed, ScienceDirect, Scopus and IEEE Explore. The search was carried out using the following keywords: “Air Pollution”, “Air pollution risk assessment”, “Respiratory Health Outcomes”, “Machine Learning” and “Explainable Artificial Intelligence”. The articles were extracted from each data repository on the 24 August 2025. The Boolean operators such as “AND” and “OR” were used in the search string to identify articles. Appendix A.1 (Supplementary S1) shows the search string used for the systematic review.

2.3. Risk of Bias Assessment and Reproducibility Strategy

The potential biases include publication bias and the lack of transparency in ML model reporting. To minimize risks of bias and enhance reproducibility, several strategies were adopted. Firstly, two independent academic researchers were engaged in the data collection process. The first researcher was responsible for data retrieval from the online data repositories, while the second merged the extracted datasets and removed duplicates using the EndNote reference manager. Discrepancies and uncertainties were resolved through group discussions, ensuring consistency and agreement on the final set of included studies.

Secondly, subject matter expert input was critical in confirming the relevance of studies included in the final analysis and to avoid bias in classification. Again, search keywords were predefined and agreed upon by the independent researchers to ensure topic alignment within the scope of the review.

Finally, to assess study quality and transparency, we drew upon established criteria (e.g., PRISMA guidelines, ROBINS-I (Risk Of Bias In Non-randomized Studies of Interventions)) for risk of bias in non-randomized studies and best practices in ML model reporting). These steps enhance the reproducibility of the review process and mitigate potential biases, including publication bias and selective reporting.

In terms of ROBINS-1, the domains relevant in this review focused on biases in confounding, reviewer selection, missing data and the selection of reported outcome. Firstly, bias due to confounding described the differences in contextual factors that could affect the review which includes ML, xAI and respiratory health outcomes. Secondly, expect view was crucial in avoiding misclassification of articles. Thirdly, bias due to missing data in terms of DOI was addressed by retrieving from Google Scholar, where missing author names were removed. Finally, bias in the selection of reported results ensured that all the specified outcomes were reported and that ML/xAI model performance metrics were transparently presented in the articles reviewed. Performance metrics, in terms of the model’s result accuracy, are measured with RMSE, MSE and R² values. An optimal model refers to a model that guarantees high results accuracy.

2.4. Articles Extraction

Articles were saved in EndNote file type, where 177 articles were finally extracted, with their related articles downloaded for a thorough literature review. After removing 25 duplicate article titles, 153 papers with unique titles were screened further. In the second stage, documents that were not within the scope were excluded from the final document review. The final stage required a thorough review of the articles (92) to clearly understand the issues relating to the research questions. Because of the possibility of missing articles during extraction, Google Scholar helped to retrieve the missing articles/documents using the DOI, because they contain some useful information.

3. Results and Discussion

Air pollution is a global health hazard which affects the health of an individual in several ways [12]. Respiratory diseases ranging from severe acute respiratory syndrome (SARS), Middle East respiratory syndrome (MERS), HIV, influenza A (H1N1) and COVID-19 can cause lung infections [19]. Huang, Zhou [20] identified the air pollutant that primarily contributes to mortality risk. They also characterized individual and joint pollutants (i.e., multi-pollutants) and examined the association of long-term air pollutant exposures with cardiopulmonary mortality. Liu, Li [21] indicate that limited information exists on the interaction effects between air pollution and influenza vaccination on allergic respiratory diseases. They conducted a large population-based study to evaluate the interaction effects between influenza vaccinations. Atzeni et al. [22] acknowledged the complexity in managing asthma issues and the many environmental factors that play a critical role in asthma management. It is critical to monitor and assess personal exposure to air pollutants to examine its impact on health and respiratory outcomes aside from asthma conditions. Neo, Hasikin [7] conducted a systematic review on air quality impact on human health, focusing on the advancement of artificial intelligence (AI) and the use of federated learning.

Among the studies that meet the inclusion criteria but were excluded include the study by Rowland et al. [23] which lacks ML and xAI but focused on gut feelings relative to fatty acids for the regulation of respiratory health; the study by Renouf, Sutanto [24], though focused on asthma, was largely on laboratory research, lacking machine learning and/or xAI content.

By combining independent screening, keyword-based selection, expert input, and formal bias assessment with ROBINS-I, we strengthened both the transparency and reproducibility of our review, while mitigating risks such as publication bias and selective reporting. The following subsections present the results of the systematic literature review to address the key research questions.

3.1. Machine Learning for Air Pollution Risk Assessment and Respiratory Health Outcomes

Ye et al. [17] suggested health outcomes such as obesity/overweight, myopia, lung health, circulatory health, cognitive function and general health in children and adolescents. Vachon, Kerckhoffs [18] indicated that machine learning techniques such as tree-based methods are superior to linear non-regularized techniques for predicting the ambient concentrations of NO₂, UFPs and BC. They also suggested that the most frequently used models are linear non-regularized methods and RF.

Geo-statistical models have been applied to assess fine-scaled air pollution exposures; however, they have not been used to assess non-regulated air pollutants, and this led to the use of a spatial model to assess both regulated and non-regulated pollutants (such as ultrafine particles, black carbon and benzene) [25]. This spatial model applied partial least squares regression (PLS) and a dimension reduction algorithm to predict pollution exposure. Upon comparison with other machine learning models, the RF and PLS emerged as the most effective models that can predict the short-term average concentration of toxic air pollutants. After evaluation, both models guaranteed a high cross-validation R² score and the lowest root mean square error value.

Li et al. [26] indicated that understanding how a model estimates error can help determine its impact on model development. Unfortunately, their experiment suggested a negative impact on model development when the RF model was used to predict the personal PM_2.5 exposure. Furthermore, an atmosphere-to-scenario-based ratio and linear regression model were used for the essential parameters of the scenario-based ratio exposure model. Their finding, however, suggested that personal PM_2.5 exposure was higher than atmospheric PM_2.5 concentration, with an R² value between 0.65 and 0.93.

Maio et al. [27] examined the long-term effect of multi-pollutants on human health. The health effects that were examined include the prevalence of respiratory/allergic symptoms and diseases. During their study, questionnaires were administered to obtain the risk factors and their health outcomes from participants; thereafter, the association between air pollutants and health effects were assessed with machine learning models such as LR (for single pollutant) and principal component LR model (for multiple pollutants). The pollutants were PM_2.5, PM₁₀, NO₂ and O₃. Their findings suggested that PM₁₀ and PM_2.5 contributed to an increased health effects in different percentages: rhinitis (14–25%), asthma (23–34%) and night awakening (30–33%); NO₂ was rhinitis (6–9%), asthma (7–8%) and night awakening (12%); and O₃ contributed 37% increased odds for asthma attacks [27].

Tang et al. [28] studied the major factors that control emission of particle-bound hydrophobic organic compounds (HOCs) in lung fluids. They focused on particles from barbecue and smoking; then a machine learning model with a vitro method assessed the most significant factor that affected inhalation bio-accessibility of polycyclic aromatic hydrocarbons (PAHs), followed by the organic carbon and elemental carbon contents. Their study suggested significant particle deposition efficiency and bio-accessible fractions of HOCs in the risk assessment.

Yang et al. [29] indicated that multiple air pollutants often interact, which suggests that there is an inter-correlation of air pollutants, and as such, single pollutant estimation models often fall short in estimating this inter-correlation. This challenge was addressed via a joint multi-pollutant retrieval model (referred to as physics-informed multi-task deep neural network (phyMTDNN)). During the experiment, six air pollutants, namely PM_2.5, PM₁₀, SO₂, NO₂, CO and O_3, were used and the evaluation with cross-validation R2 value suggested a variation from 0.7 to 0.90 [29].

Zhang et al. [30] opined that PM_2.5 has a significant impact on health and climate. During their impact analysis, the inter-annual trends in particulate air pollutants and the response to anthropogenic emissions were examined using RF algorithm. This machine learning model quantified key drivers of PM_2.5. Furthermore, the health-risk air quality index (HAQI) value appeared to have decreased with significant contribution from anthropogenic emission reduction. It was concluded that the contribution of PM_2.5 to HAQI was reduced, and there was a significant contribution from O₃ [30].

Zhou et al. [31] examined the joint effect of the indoor air pollution index and ambient particulate matter on fetal growth. Prenatal exposure to PM_2.5 and PM₁ was assessed using a machine learning technique, which further computed the indoor air pollution index. The finding suggested that indoor air pollution and ambient PM exposure had individual and joint negative effects on fetal growth [31]. Ahmadian, Rajabi [32] assessed outdoor air pollutants and their consequent health risks using machine learning models. The pollutants considered were PM_2.5, SO₂, NO₂, and CO. During the assessment, the EPA health risk assessment model was used to understand the exposure to human health. In using the model, the mean hazard quotients (HQ) for PM_2.5 and NO₂ were established, where the HQ value exceeded 12 h exposure, which raises health risk concerns. The machine learning models used for the analysis are RF, K-Nearest Neighbors (KKN), decision tree (DT) and Naive Bayesian (NB). RF model emerged as the best prediction model with the highest accuracy of R² value of 0.99.

Chen et al. [33] developed an RF model to predict the daily air pollutants such as PM_2.5 and PM₁₀, NO₂ and O₃ from multisource at an 8 h interval. This model was evaluated with out-of-bag R^2, where PM_2.5 (0.85), PM₁₀ (0.77), O₃ (0.85) and NO₂ (0.85).

Guo et al. [34] examined the association of PM_2.5 and acceleration of aging (AA) using a two-staged machine learning technique at one km² resolution. The machine learning technique was based on the post-matching linear mixed model (LMM), where the determination of aging was based on the variation between Klemera–Doubal biological age and chronological age. The outcome suggested that long-term exposure to PM_2.5 may increase the biological aging among the adult population, for instance, in China. Furthermore, when the adult population was exposed to an increased level of PM_2.5, it also increased their aging process [34].

Jiang et al. [35] designed a navigation system to aid users in commuting at least PM_2.5-exposed paths, and the system automatically sends a risk warning to its users. Their navigation system employs machine learning models, namely RF, LightGBM, Adaboost and XGBoost, where XGBoost outperformed the comparative models with R² value exceeding 0.90 and with root mean square error (RMSE) (~15.74 μg/m³). Thus, commuters who engage in outdoor activities can avoid locations that are highly exposed to PM_2.5 concentration [35].

Lv et al. [36] assessed the spatiotemporal dispersal of PM_2.5 source-specific impacts in locations without PM_2.5 compositions using machine learning models. XGBoost showed outstanding prediction performance for nitrogen dioxide (NO₂), carbon oxide (CO) and elemental carbon (EC), with cross-validated R² values ranging between 0.87 and 0.92 [36].

Masseran, Safari [37] classified the categories of air pollution events (that is, extreme and non-extreme) using a probabilistic machine learning technique, which is based on naïve Bayes technique. These events are grouped based on the magnitude of severity. The naïve Bayes model guaranteed high accuracy, specificity and sensitivity on the trained and test datasets [37].

Chen et al. [38] suggested that ultrafine particles (UFPs) (with a diameter ≤ 100 nm) are airborne particles which can pose a significant health risk, including cardiovascular disease, respiratory and cancer. Their study found that while experimental approaches are unable to capture the temporal fluctuation, computational models are also deficient in capturing the complex nature of the ultrafine particle depositions, thus making it challenging to understand the impact on human health. Some of the sources of UFPs include traffic, combustion and industrial processes [38].

Amini et al. [39] indicates that European cities (such as Copenhagen, Denmark) lack models to understand the long-term average particle size (APS), lung deposited surface area (LDSA) and particle number concentration (PNC). Thus, a machine learning model based on RF and bagged trees was suggested. Their model achieved a high R² value (0.93) and low root mean square error [39].

Cao et al. [40] focused on both water and air pollution as they pose significant threats to public health. They used traditional biological methods, which are time consuming, costly and limited in scale. Subsequently, a machine learning model, that is, Knowledge-guided Pre-training of Graph (KPGT), was used as a pollution risk assessment model, which predicted the carcinogenicity of pollutants with an AUC (0.83) [40].

Chen et al. [41] indicates that air pollution and climate change are key challenges in recent times. Unfortunately, there is an inadequate assessment of the impact of extreme temperatures and air pollution from different regions. Air pollution includes daily PM_2.5, PM₁₀, NO₂, and O₃ concentrations. They design machine learning models to estimate the daily concentration within the required location-specific time-varying temperature threshold [41].

Gao et al. [42] examined the association between mixed pollutants in PM_2.5 and reproductive health risk using the quantitative structure–activity relationship model (QSAR) integrated machine learning algorithm. In their study, the mixed reproductive health risks associated with phthalates (PAEs) and organophosphates (OPEs) exposure were assessed to understand how strongly the compounds bind to estrogen receptors (ER) and androgen receptors (AR). Thus, it will help gain insight into how the mixed reproductive exposure risk of PAE and OPE is associated with PM_2.5 [42]. Phthalates (PAEs) from indoor dust impact human reproductive health [43].

Holloway et al. [44] indicated that collecting satellite data on PM_2.5 and NO₂ are beneficial in developing a health risk assessment model to track both long-term and short-term trends on air quality information and related health records.

Kuo et al. [45] proposed the machine learning (ML)-measurement model fusion (MMF) framework to quantify air pollutants using Chemical Transport Modeling (CTM) data and to assess the burden of disease (BD) from exposure to different sources of air pollutants. Among the air pollutants considered are PM_2.5 and O₃. The proposed ML-MMF successfully improves CTM-modeled for PM_2.5 and O₃ with R² values, respectively, ranging from 0.41 to 0.86 and 0.48 to 0.82 [45].

Lee et al. [46] developed a comprehensively integrated air-risk index (CARI) to assess the adverse effect of hazardous air pollutants (HAP) such as particulate matter (PM). They suggest that including a machine learning model enhances the spatiotemporal resolution of the CARI model for large industrial locations.

Liu et al. [47] examined how atmospheric particulate matter (PM) affects human health using a Lung Performance Optimization-based XGBoost (LPO-XGBoost) model, which leverages adaptive optimization principles inspired by lung function to enhance classic PM source apportionment. Comparative analyses with models such as RF, Support Vector Machine (SVM), and their LPO-enhanced variants confirm that LPO-XGBoost provides the most reliable performance in estimating pollution source contributions, offering scalability and robustness ideal for high-time-resolution observational data and having a coefficient of determination R² (0.88) [47].

Wenjie et al. [48] indicates that air pollution and tobacco smoke are one of the significant risk factors. The presence of Benzopyrene (BaP) in these pollutants was explored using a combination of different models, such as network toxicology, machine learning (Support Vector Machine, RF and LASSO regression), and molecular docking analyses. This resulted in the creation of a robust periodontitis risk assessment model to understand environmental pollutants on periodontal health [48].

Yang et al. [49] attempted to model the association between varied types of eye-level greenness and pediatric blood pressure using the deep machine learning model and generalized linear mixed-effects regression model. The mediation analysis results show that hypertension is significantly mitigated by lower levels of air pollutants, including PM_2.5, PM₁₀, SO₂ and NO₂. This outcome suggests that eye-level greenness, especially trees, was associated with lower prevalence of hypertension in children, with air pollution exhibiting mediating effects [49].

Zhao et al. [50] clearly identifies the source of atmospheric particulate matter concentration to help in planning the needed intervention to reduce the concentration, which can be harmful to human health. They further stated that, though traditional methods such as chemical mass balance models are effective, machine learning models are more efficient and effective. Such effective models include the source apportionment pipeline (RX model), which combines computer vision and machine learning to trace particle sources. This can help build particulate matter health risk assessment models for atmospheric pollution source analysis [50].

According to Paul et al. [51], assessing the acute and chronic health effects of biomass smoke exposure require reliable estimates of PM_2.5 concentrations during the wildfire season and throughout. The RF machine learning model uses potential predictor variables integrated from multiple data sources and estimates daily mean (24 h) PM_2.5 concentrations at a 5 km × 5 km spatial resolution [51].

Brooks et al. [52] examined how brick kilns contribute to fine particulate matter (PM_2.5) and estimated its association with child asthma symptoms, chronic obstructive pulmonary disease (COPD) and general respiratory symptoms [52].

Hu et al. [53] examined prenatal and early-life exposure to air pollution and extreme temperatures and how they are associated with childhood asthma and wheeze. By using the distributed lag nonlinear models (DLNMs), they suggested that early-life air pollution exposure contributes to the development of childhood asthma and wheeze, while exposure to temperature showed mixed associations with respiratory outcomes.

The study by Okello, Nantanda [54] indicates that ambient air pollution (such as PM_2.5) and weather have an impact on respiratory diseases. They applied generalized additive models (GAM) to adjust for rainfall, temperature and humidity, and also assessed the association between air pollution and healthcare facility events. The finding suggested that PM_2.5 impacts healthcare facility and that their association are influenced by meteorological factors [54]. Bouma et al. [55] acknowledged the use of different risk exposure assessment methods for outdoor air pollution with asthma onset and lung function in children [55].

The study by Leão et al. [56] indicates that an increase in temperature will have impacts on air particulate matter and health outcomes [56]. Shams et al. [57] suggested the need to link deep learning-based air quality forecasting models with health and socioeconomic outcomes. They proposed the combination of best-performing models, such as Deep convolutional neural networks (Deep-CNN), Long Short-Term Memory (LSTM), and deep neural networks (DNN) to forecast the daily maximum O3 concentrations [57]. Many researchers have acknowledged the impact of air pollution on human health and have adopted different methods to assess the health risk outcomes [56,58,59,60]. Understanding the spatiotemporal dynamics of Population Exposure to PM_2.5 (PEP) and its relationship with socioeconomic activity (SEA) is crucial to reducing exposure risks and health dangers [61].

Pei et al. [62] studied the environmental pollutant fluctuation and its potential health impact, particularly viral hepatitis. They focus on 2005 and 2020 viral hepatitis datasets and 2012 to 2020 pollutants datasets (CO, NO₂, and O₃). Three time series models, Holt–Winters model, seasonal auto-regressive integrated moving average (SARIMA) and generalized additive model (GAM), were applied for prediction and to assess the effect of pollutants on viral hepatitis. Furthermore, ML models were also used for pollutant analysis. While Holt–Winters was superior in predicting the incidence of viral hepatitis, SVM and GPR models were superior in analyzing pollutant data. The findings suggest that patients who are infected with HAV and HEV were influenced by PM₁₀ and CO, whereas SO₂ and PM_2.5 impact each other significantly. Again, persons aged between 35 and 64 years are more susceptible to these pollutants.

VOC is one of the air pollutants that is very harmful to human beings, and it is associated with risks of acute and chronic disease, such as cancer and leukemia [63]. Given the complex nature of screening the biological particles for VOC exposures, machine learning models and bio-information techniques were found to be limited in validating the biomarkers and toxicity of VOC exposure.

Wang, Wang [64] measured the PM_2.5 and PM₁₀ concentrations both indoors and outdoors. While sensors provide accurate measurements, it suggests that the RF model is an effective technique to validate the measurement as compared to the multivariate linear model (ML), K-nearest neighbor model (KNN), support vector machine model (SVM), decision tree model (DT), neural network model (that is, Multi-Layer Perceptron), R² and RMSE performance validation models. Appendix A.2 (Supplementary S1) provides a summary of the related literature on ML for air pollution risk assessment;

Table 2. Model categories and related algorithms.

Model Type	Type of Algorithm	Typical Datasets Used	Validation Approach	Comparability
Tree-based models	Random Forest (RF), XGBoost, LightGBM	Air quality monitoring station data; meteorological variables.	K-fold cross-validation; train/test split	Widely used; strong performance, but feature importance varies across studies.
Deep learning models	CNN, LSTM, DNN [57]	Spatiotemporal pollution data; satellite-based remote sensing; health outcomes.	Hold-out validation; temporal split	Handle complex patterns; often limited transparency; high data demands.
Hybrid models	LUR + ML, SARIMA + ML	Land use regression data; time series pollution data.	Rolling-origin or temporal CV	Improve accuracy but add complexity; results are difficult to generalize.
Statistical baselines	SARIMA, regression models	Time series pollutant concentration.	Train/test split; rolling validation	Serve as benchmark; often outperforms ML in small datasets.

also presents model categories and related algorithms.

The heterogeneity of methods summarized in Table 2 highlights a key challenge for comparability across studies. Differences in modeling approaches (e.g., tree-based ensembles vs. deep neural networks), dataset sources (monitoring stations, satellite retrievals, clinical records) and validation strategies (cross-validation, temporal splits, rolling-origin evaluation) make it difficult to directly compare reported performance metrics. For example, deep learning models often require large-scale spatiotemporal datasets, while tree-based models perform reliably even with smaller, heterogeneous inputs. Similarly, temporal validation generally produces lower performance estimates compared to random K-fold cross-validation, which may inflate comparability. This methodological diversity underscores the need for standardized evaluation protocols and reporting practices to enable fair comparison and benchmarking of models for air pollution risk assessment and respiratory health outcomes.

In view of this literature, ML techniques have been extensively applied to air pollutants risk assessment and respiratory health outcomes. Machine learning models such as extremely randomized tree (ERT) proves to outperform both XGBoost and RF using the PAHs pollutant, thus leading to the exposure of lung cancer risks; the hybrid LUR and CNN performed optimally in predicting the impact of UFPs on mortality; RF performed optimally in PM_2.5 Bound Mercury (PBM_2.5) prediction which lead to the findings that population density and power generation stations contribute largely to PBM_2.5 concentration. Furthermore, to a very limited extent, PBM_2.5 in 2020 posed a non-carcinogenic health risk. This is supported by the hazard quotient (HQ), which remained well below the threshold of concern, with values mostly <0.02 during winter. Such levels indicate that the risk was negligible, meaning exposure did not reach levels typically associated with adverse non-carcinogenic health outcomes. Estimating PM_2.5 using TDLPI + XGBoost to enhance the LUR model suggested a moderate health risk in winter. Again, a satellite-based ML model with a time-varying Cox hazard model proves to establish the association between mortality risks due to lung cancer with PM_2.5.

To a significant extent, exposure to PM_2.5 is associated with an elevated risk of cardiorespiratory mortality [65]. Epidemiological studies show that fine particulate matter can penetrate the lungs of an individual, trigger systemic inflammation, oxidative stress and vascular dysfunction. The risk may be more pronounced among vulnerable groups such as the elderly, children and individuals with some pre-existing conditions. However, the extent of risks varies depending on pollutant concentration, duration of exposure, location and population susceptibility. Therefore, while evidence strongly supports a causal link, the degree of mortality risk remains context-dependent, influenced by environmental and demographic factors.

3.2. Explainable AI with Machine Learning Models for Air Pollution Risk Assessment and Respiratory Health Outcomes

Explainable AI (xAI) helps with model interpretability and transparency. While most machine learning models have been widely used to predict air pollution concentrations, they lack interpretability [64]. Zhang, Yi [66] indicated that environmental pollution contributes to cancer etiology; however, the expensive nature of traditional experiments and their time-consuming nature suggest the need for an alternative approach, such as deep learning models. For instance, the interpretable DL model (CarcGC for chemical carcinogenicity) was proposed, which was based on graph convolutional neural network (GCN) that employed molecular structural graphs as input and for interpretability. Their study further indicated the association of air pollution with the incidence of lung cancers, and, therefore, accurate prediction was suggested.

Rajesh, Babu [67] developed a real-time air quality assessment and health risk prediction model for multi-locations. Their techniques were Gradient Boosting, RF, XGBoost and LSTM. These helped systems to predict air pollutant concentrations and then classify the air quality levels with high temporal accuracy. Furthermore, SHAP (Shapley Additive exPlanations) was used for interpretability of their model.

Razavi-Termeh, Sadeghi-Niaraki [68] examined the lack of interaction between asthma-prone locations and urban environmental factors and suggested the use of spatiotemporal modeling techniques with an xAI technique like SHAP. Their modeling technique leveraged XGBoost (eXtreme Gradient Boosting), optimized using Bat algorithm (BA). Upon evaluating the asthma-prone area map with Receiver Operating Characteristic (ROC), different accuracies were obtained for different seasons and percentages, such as spring (97.3%), summer (97.5%), autumn (97.8%) and winter (98.4%). This result suggested that rainfall and temperature have the most significant impact on asthma. Furthermore, air pollutants in certain seasons, such as PM_2.5 (in spring), CO (in summer), O₃ (in autumn) and PM₁₀ (in winter) showed the most substantial effect among other factors of air pollution.

Agbehadji and Obagbuwa [69] examined how to enhance the noise in measuring the concentration of SO₂ prediction. They indicate that the noise can increase the error of estimation in a dynamic time-varying air pollutant such as SO₂. In this regard, an xAI technique using SHAP was integrated with an LSTM-based Adaptive Kalman Filter model. However, the aspect of respiratory health-related outcomes was not considered in their study. Furthermore, Agbehadji and Obagbuwa [70] leveraged xAI (e.g., LIME) in graph neural network that also include A* search strategy (GeoxAI-GCN-LSTM-A*) to find an optimal navigation path for predicted CO concentration in multi-dimension time step. However, the respiratory health-related outcome was not considered in their approach.

Kibria, Hossain [19] pointed out the common lung infection detection techniques like X-ray image analysis, computed tomography (CT) scans and reverse transcription–polymerase chain reaction (RT-PCR). These techniques have their limitations, thus leading to the proposition of parallel lightweight diagnosis model based on depth-wise separable CNN (LW-PDS-CovidNet) model. Additionally, SHAP and gradient-weighted class activation mapping (GRAD-CAM) were the xAI techniques that showed the most significant features. Their model achieved a high accuracy of 98.06% (for two-class), 97.43% (three-class) and 92% (four-class) classification with a low number of parameters. Finally, it was suggested that respiratory diseases, including severe acute respiratory syndrome (SARS), Middle East respiratory syndrome (MERS), HIV, influenza A (H1N1) and COVID-19, can cause lung infections.

Sunder, Tikkiwal [71] indicated that human beings are mostly exposed to fine PM_2.5, which has adverse health impacts related to cardiovascular and respiratory-related diseases. They suggested that machine learning models proved to be a good predictor of air quality. To validate their suggestion, they provided a framework based on temporal-based prediction using RF, Gradient Boosting (GB) regression, and Tree-based Pipeline Optimization Tool (TP) Auto ML that worked with meta-heuristic methods such as genetic algorithm. TP produced Global Performance Index (GPI) values of 7.4, which was the highest GPI value for August 2016, while the lowest GPI value (−0.6) was recorded in June 2019. These outcomes were explained with xAI technique, which helped to investigate the fidelity of feature importance of machine learning models used to extract information on rhythmic shift in the PM_2.5 concentration patterns.

Chang, Liu [72] demonstrated the use of machine learning to predict pediatric acute respiratory infections (ARIs), which are common in children. The machine learning model predicted six common respiratory pathogens, such as adenovirus, influenza virus (types A and B), parainfluenza virus (PIV), Mycoplasma pneumoniae (MP) and respiratory syncytial virus (RSV). Feature importance was assessed with SHAP values. The findings suggest that event patterns are useful for influenza virus prediction, and C-reactive protein had the highest SHAP value for adenovirus infections. Thus, such models could improve patient outcomes and reduce unplanned medical expenses when integrated into clinical workflow.

Chi, Moghadas-Dastjerdi [73] examined the diagnosis code associated with the highest mortality risk using a deep SHAP (D-SHAP) model. They indicated that in most instances, clinicians are more interested in learning how input (from raw dataset) affects prediction. The results suggest the top five most important diagnosis codes as respiratory failure, sepsis, pneumonia, shock, and acute kidney injury. Fortunately, these results were consistent with the opinion of the physicians. On the contrary, the diagnosis for urinary tract infection showed a discrepancy between D-SHAP values and clinical opinion, because of the low frequency of the disease and its outcome in combination with other comorbidities. Appendix A.3 (Supplementary S1) summarizes the xAI technique integrated with ML for air pollution risk assessment in the literature.

In conclusion, to some extent, xAI has been integrated into ML air pollution concentration prediction models. Among such models are XGBoost with SHAP for air pollutants such as UFPs and PM_0.1; LightGBM-based Bayesian with SHAP for PM_2.5, OM, NO₃⁻, NH₄⁺, BC and SO₄²⁻; XGBoost-based NDVI with SHAP for PM_2.5. The performance evaluation of the XGBoost with SHAP based on RMSE guaranteed 0.79, suggesting a need to enhance the model’s performance. Also,

Table 3. Methodological heterogeneity of xAI-integrated ML studies for air pollution and respiratory health outcomes.

Author	Model(s)	Dataset/Application	xAI Technique(s)	Outcome Focus	Key Findings
Zhang et al. [66]	GCN (CarcGC)	Chemical carcinogenicity (molecular structural graphs)	Model-intrinsic interpretability	Air pollution and lung cancer	GCN-based interpretability identified structural predictors of carcinogenicity.
Rajesh et al. [67]	Gradient Boosting, RF, XGBoost, LSTM	Real-time multi-location AQ and health risk	SHAP	Air pollutant prediction + AQ classification	Accurate AQ prediction with interpretable pollutant contributions.
Razavi-Termeh et al. [68]	XGBoost + Bat algorithm	Asthma-prone locations (spatiotemporal)	SHAP	Asthma risk + seasonal pollutant drivers.	Identified seasonal pollutant importance (e.g., PM2.5 in spring, CO in summer).
Agbehadji et al. [69]	LSTM + Adaptive Kalman Filter	SO2 concentration prediction	SHAP	Air pollution (no direct health outcomes)	Improved interpretability of SO2 estimates; highlighted noise impact.
Agbehadji et al. [70]	(GeoxAI-GCN-LSTM-A*)	CO prediction.	LIME	Air pollution	Optimal navigation route with RMSE (0.4059 for 8 h and 0.4124 for 16 h).
Kibria et al. [19]	LW-PDS-CovidNet (CNN)	Lung infection detection (CT, X-ray, RT-PCR)	SHAP, Grad-CAM	Respiratory infections (COVID-19, SARS, MERS, etc.)	High accuracy (>90%); highlighted key imaging features.
Sunder et al. [71]	RF, Gradient Boosting, TP AutoML.	PM2.5 temporal prediction	xAI (feature importance, SHAP-like methods)	Air pollution	Meta-heuristic TP AutoML achieved highest GPI; interpretability aided temporal feature analysis.
Chang et al. [72]	ML classifiers	Pediatric ARI prediction	SHAP	Respiratory pathogens (adenovirus, influenza, etc.)	Identified biomarkers (CRP most important for adenovirus).
Chi et al. [73]	D-SHAP (Deep SHAP)	Clinical diagnosis codes	Deep SHAP	Respiratory failure, sepsis, pneumonia, etc.	Explanations aligned with physician judgment, improving trust in predictions.

presents the methodological heterogeneity of xAI-ML studies.

Table 3 highlights the methodological heterogeneity in how explainable AI has been integrated into machine learning applications for air pollution and respiratory health outcomes. While previous reviews have focused primarily on either model performance or broad applications of ML in environmental health, our synthesis uniquely maps model architectures, datasets, xAI techniques, and health/pollution outcomes in a comparative framework. This approach reveals not only what models were used, but also how interpretability was operationalized (e.g., SHAP, Grad-CAM, deep SHAP), in what contexts (spatiotemporal prediction, imaging, clinical diagnosis) and with what implications for respiratory health and policy translation. By presenting both a structured comparison and a critical appraisal, this systematic review provides clarity on where methodological diversity strengthens generalizability and where it limits comparability, offering actionable insights for researchers aiming to select, validate and apply interpretable ML models in further air pollution health studies. To some extent, on respiratory outcomes, high PM_2.5 contributes significantly to an increased asthma risk. Again, a high urban heat island index (UHI) and high PM_2.5 concentration levels also increase asthma risk. In the categorization of health risk outcomes [11], asthma was revealed to be the most featured in the literature to the neglect of others, such as any health symptoms, allergy and flu-like symptoms. PM_2.5 has been identified as a significant air pollutant contributing to premature mortality [65]. From an epidemiological perspective, most studies establish associations but not precise causal thresholds; however, SHAP algorithm identifies the concentration thresholds at which the components of PM_2.5 could affect mortality. Thus, indicating a long-term mortality and morbidity risk.

Table 4. Reported performance metrics (selected examples) across ML/xAI studies.

Model(s)	Dataset/Task	Reported Metric(s)	Performance	Findings
RF, XGBoost, LSTM [68]	Real-time AQ and health risk prediction	R2 (0.82–0.89), RMSE (between 4.2 and 6.1 µg/m3)	High temporal accuracy.	SHAP improved interpretability of pollutant drivers.
XGBoost + Bat Alg. [69]	Asthma-prone locations	AUC (between 0.973 and 0.984)	Seasonal variations noted.	PM2.5, CO, O3, PM10 dominant factors.
LSTM + Adaptive Kalman Filter [70]	SO2 prediction	RMSE (0.79)	Reduced noise error.	No direct health outcomes.
CNN (LW-PDS-CovidNet) [19]	Lung infection detection	Accuracy between 92–98%.	High diagnostic accuracy.	xAI (SHAP, Grad-CAM) explained imaging features
Pediatric ARI ML [72]	Clinical infections prediction.	AUC between 0.86 and 0.92.	Event patterns predictive.	SHAP highlighted biomarkers.
D-SHAP model [73]	Diagnosis codes and mortality risk.	AUC > 0.90	Clinician-aligned explanations.	Improved clinical trust.

presents the reported performance metrics.

Although several studies report high performance metrics (e.g., R² > 0.80, AUC > 0.90), the lack of standardized datasets, inconsistent validation approaches, and variable outcome measures preclude direct effect-size comparisons. For example, deep learning models often report superior accuracy in image-based classification, while tree-based ensembles achieve robust performance in spatiotemporal pollution prediction; however, these results are not directly comparable due to differences in input features and evaluation designs. Accordingly, our conclusions emphasize relative performance trends within contexts (e.g., tree-based models for tabular spatiotemporal data, CNNs for imaging) rather than identifying a universally “best” model. These findings suggest that systematic reviews would benefit from harmonized performance reporting and benchmark datasets, which would enable quantitative meta-analysis of ML/xAI models in this domain.

3.3. Challenges and Limitations

Though ML models have been applied within the context of respiratory health outcome models. It is evident that the focus mostly lies on asthma. One of the major challenges in environmental pollution assessment is that health consequences are difficult to evaluate scientifically and accurately [74]. Exposure to particulate air pollution is one of the greatest environmental risk factors for adverse human health outcomes [75].

The challenges in terms of machine learning models are as follows: Firstly, to some extent, models perform differently, which impacts the accuracy of prediction. In view of this, the model’s performance may be good in one location or air pollution dataset, and that same model might perform sub-optimally in another location within the context of air pollution risk analysis and its respiratory health-related outcomes. Thus, generalization of a model would be problematic. The second challenge is that these models are more focused on the prediction of air pollutant concentration and its health risk-related factors, including mortality risk [20]. Unfortunately, causal inference may limit real policy application in the health sector, particularly on respiratory health outcome issues. This, however, may be the reason for reliance on the traditional methods of respiratory health outcome analysis. In this regard, machine learning models could complement or validate expert opinion as well. This suggests a symbiotic relationship between ML models and expert opinion.

The methodological challenge stems from data collected from the online repositories for this systematic review. For instance, not all available data repositories are considered, and it is equally possible that there could be very useful information on those repositories. Another methodological limitation is that the research articles extracted were restricted to articles, review articles and conference paper, neglecting other equally important and relevant materials like book chapters and others.

Generally, most cited studies focus on developed countries including Asia, Europe and North America. Thus, there was limited research data coverage of low- and middle-income countries, where air pollution and respiratory risk is highest.

The challenge relating to the integration of xAI with ML is the possibility of discrepancy in results obtained from the xAI model (e.g., D-SHAP) and clinical results [73]. This suggests that the interpretation of results from ML models should be augmented with an expert view. Thus, clinical validation of xAI model’s outcome should be encouraged due to the generalization challenge of ML models. Furthermore, different air pollution and respiratory datasets could be used to validate these models to enhance model generalizability. While SHAP was the most frequently reported interpretability method in the reviewed studies, alternative xAI techniques such as LIME (Local Interpretable Model-Agnostic Explanations), Grad-CAM (Gradient-weighted Class Activation Mapping), counterfactual explanations and integrated gradients are increasingly applied. SHAP provides consistent, game-theoretic attributions that are generally more stable and globally interpretable than LIME, but it can be computationally expensive and may still struggle with correlated features. LIME, by contrast, is faster and more flexible but often produces explanations that are unstable across runs, which limits its reliability in high-stakes clinical settings. Grad-CAM and related saliency-based methods are particularly useful in image-based tasks (e.g., chest X-rays, CT scans), offering visual heatmaps of important regions, but they provide less quantitative interpretability compared to SHAP or LIME. Counterfactual methods have the advantage of generating “what-if” scenarios that can directly inform clinicians about actionable changes but remain less explored in environmental epidemiology.

From a clinical decision-making perspective, these trade-offs are critical. For example, SHAP’s robust feature attributions can support policymakers in identifying consistent air pollution drivers across populations, whereas LIME’s instability could undermine trust in regulatory applications. Conversely, in diagnostic imaging, Grad-CAM may be more intuitive for physicians since it visually overlays explanations on familiar radiological images.

Table 5. Comparison of xAI techniques for air pollution and respiratory health applications.

xAI Technique	Strengths	Limitations	Best-Use Scenarios
SHAP (Shapley Additive Explanations)	Provides consistent, game-theoretic feature attributions; both global and local interpretability; widely adopted in environmental/epidemiological studies.	Computationally expensive; sensitive to correlated features; explanations may be complex for non-technical users.	Tabular spatiotemporal data; epidemiological studies; policy translation requiring robust feature attribution.
LIME (Local Interpretable Model-Agnostic Explanations).	Fast; flexible; model-agnostic; intuitive local approximations.	Unstable across runs; explanations vary with sampling; less reliable for high-stakes contexts.	Quick interpretability checks; exploratory analysis; low-stakes decision support.
Grad-CAM (Gradient-weighted Class Activation Mapping).	Visual explanations overlayed on images; intuitive for clinicians; highlights regions of interest.	Limited to CNNs; qualitative rather than quantitative; can miss non-visual factors.	Medical imaging (X-rays, CT scans); radiology applications.
Integrated Gradients.	Provides attribution scores along a path from baseline to input; less noisy than saliency maps; theoretically grounded.	Requires careful baseline selection; computationally intensive.	Deep learning models with structured or sequential data (e.g., respiratory signals, spatiotemporal patterns).
Counterfactual Explanations.	Intuitive “what-if” reasoning; directly actionable; useful for clinical/policy interventions.	Computationally complex; less standardized; limited adoption in air pollution.	-

presents the comparison of common xAI techniques for air pollution and respiratory health applications.

3.4. Uncertainty Quantification

A critical consideration for the translation of ML models into air quality regulation and respiratory health policy is the treatment of uncertainty. While many studies report performance metrics such as R², RMSE or AUC, relatively few explicitly quantify predictive uncertainty. This omission limits the reliability of model-informed decision-making in regulatory settings where precautionary principles and health risk thresholds are central. Uncertainty in ML predictions arises from several sources: (1) data uncertainty includes measurement errors, missing data and sensor calibration issues; (2) model uncertainty includes structural choices, parameter variability and overfitting/under fitting); (3) outcome uncertainty includes inherent variability in respiratory health responses to pollutant exposures.

For policymakers, explicit uncertainty estimation methodological approach (e.g., bootstrap resampling and cross-validation, ensemble methods, among others) are crucial to balance precautionary measures, communicate risk transparently and prioritize interventions. For example, if an ML model predicts high respiratory risk from PM_2.5 with wide uncertainty intervals, regulators may adopt stricter emission standards or reinforce monitoring in hotspot areas. Conversely, narrow and stable uncertainty bounds can justify targeted, cost-effective interventions. Therefore, integrating uncertainty quantification approach into xAI-ML-based air pollution assessments enhances not only scientific robustness but also regulatory credibility and public trust.

3.5. Implications of the Research Finding

The implication focused on three core areas: practical, policy and theoretical. The practical implication is to determine how the findings could be used in real-world practice. In this regard, the findings suggest methodological considerations to guide healthcare application developers to focus on the best-performing models such as ERT in addition to XGBoost and RF. Again, uncertainty estimation methods, when integrated into xAI-based ML models, provide some robustness in existing respiratory health outcome applications. Furthermore, healthcare sector manages volumes of patient records using legacy systems built with statistical and/or hybrid models, and integrating the optimal model identified in this review is beneficial. Furthermore, it helps health practitioners find alternative strategies aside from laboratory work to validate their laboratory results with machine learning methods. It is imperative for models (causal-xAI-ML/DL) that seek to find causal relationship with some explanation to be considered for practical implementation.

Policy implications seek regulatory reforms within the health sector that leverage environmental data and respiratory impact. In this regard, the findings suggest the adoption of hybrid models (ML/DL and xAI) in the work process of the health sector, which require the re-design of work structure and staff training on the use of new models for respiratory risk assessment and its outcomes. Again, policy that enforces uncertainty quantification in xAI-based ML/DL in addition to the current performance metrics is encouraged for AI models on health.

Theoretically, the findings are highly relevant and timely, given the increasing adoption of AI techniques in environmental epidemiology and the growing demand for interpretability in clinical and policy contexts.

3.6. Future Directions

With continued advancements and responsible integration, AI will play a pivotal role in shaping the next generation of precision thoracic surgery [10]. To some extent, the use of other explainable AI models, such as LIME could be explored within the context of air pollution risk assessment models for respiratory health outcomes. While Table 3 provides a comprehensive lens on methodological diversity (models, xAI methods, application domains), and it also highlights their gaps (e.g., limited direct integration with respiratory health outcomes, reliance on post hoc SHAP). Future research is focused on digital twin technology, federated learning, and xAI to improve AI interpretability, reliability, and accessibility.

Considering the lack of causal inference, future work should explore causal ML approaches (e.g., causal forests, Directed Acyclic Graphs (DAG)-based models). DAG models the complex causal structures and improve causal inference in various health-related scenarios. While causal forests identify effect heterogeneity with tree-based models, DAG methods offer a structural method to capture causal relationships before using an estimation technique. Furthermore, developing a hybrid causal-xAI-ML model offers interpretability in clinical context.

To enable future meta-analyses and more robust cross-study comparisons, this review recommends the development of a standardized benchmarking framework with the following elements. (1) Dataset standardization: establish benchmark datasets (e.g., multi-city, multi-season air quality and health records) that can serve as reference points for model evaluation, and encourage open-access repositories to facilitate reproducibility and fair comparison. (2) Harmonize performance metrics which require consistent reporting of core metrics (e.g., R², RMSE/MAE for regression; AUC, sensitivity, specificity for classification), mandating clear definitions of metrics (e.g., temporal vs. random cross-validation splits) to avoid misinterpretation. (3) Validation protocols promote standardized validation strategies such as rolling-origin evaluation for time-series forecasting and nested cross-validation for small datasets, explicitly documenting training/validation/test splits, including temporal coverage, to improve reproducibility. (4) Transparent xAI reporting requires reporting of performance and interpretability outputs (e.g., SHAP values, feature rankings, counterfactual examples), and alignment of xAI explanations with domain expertise (e.g., clinicians, environmental scientists) is encouraged to validate interpretability claims.

4. Conclusions

In conclusion, this review provides valuable insight into machine learning and explainable models in air pollution risk assessment for respiratory health outcomes. This review highlights several issues including methodological issues relating to xAI integration in ERT algorithm, fusion of uncertainty quantification methods into best-performing AI models, federated learning and causal-xAI-ML models. These provide some level of assurance to policymakers within the arena of healthcare application. Furthermore, the challenges identified and the implication of the findings contribute greatly to helping AI-based healthcare application developers find alternative strategies aside from the traditional methods of risk and exposure–response assessment models within the respiratory health outcome arena. Ultimately, no single xAI technique is universally optimal; the choice depends on data modality, computational constraints and end-user needs. Future studies should aim to triangulate findings using multiple xAI methods and to validate interpretability outputs against expert judgment to enhance trustworthiness in clinical and policy translation. Again, future research should systematically benchmark uncertainty quantification techniques alongside performance metrics to enable evidence-based policymaking.