From prediction to policy: Explainable AI bridges gap in air pollution risk assessment

The study identifies explainable AI as critical for bridging the gap between data-driven predictions and evidence-based public health action. By making AI decisions more interpretable, xAI helps scientists, clinicians, and policymakers understand not just what the risks are, but why they occur.

A new review provides one of the most comprehensive assessments to date of how explainable artificial intelligence (xAI) is being deployed to track air pollution and forecast respiratory health outcomes.

The study, titled "Explainable Artificial Intelligence and Machine Learning for Air Pollution Risk Assessment and Respiratory Health Outcomes: A Systematic Review" and published in Atmosphere highlights that while xAI-powered models are proving useful in predicting pollution-related health risks, challenges in transparency, validation, and policy adoption still limit their full potential.

Why explainability matters in AI-driven air pollution models

The authors argue that standard machine-learning models have made progress in forecasting health risks linked to poor air quality but often remain opaque in how they reach their conclusions. This lack of transparency has slowed adoption in clinical and policy settings, where trust and interpretability are crucial.

The review screened thousands of publications and narrowed them down to 92 studies published between 2020 and 2025, using the PRISMA framework and risk-of-bias assessment tools. The authors found that many models achieved high predictive accuracy but failed to report explainability outputs, limiting their usefulness for decision-makers in healthcare and environmental management.

The study identifies explainable AI as critical for bridging the gap between data-driven predictions and evidence-based public health action. By making AI decisions more interpretable, xAI helps scientists, clinicians, and policymakers understand not just what the risks are, but why they occur.

State of the Art: Methods and their gaps

The review highlights that random forests, XGBoost, and deep neural networks dominate the field for predictive modeling of air pollution exposure and respiratory health outcomes. Among explainability tools, SHAP (SHapley Additive exPlanations) is the most commonly used, enabling models to highlight which pollutants or environmental factors most influence outcomes.

However, the authors warn that several gaps remain:

• Generalization limits: Models often perform well in the region or dataset where they were developed but fail to maintain accuracy in different geographies or seasons.
• Data constraints: Many regions still lack comprehensive, high-quality pollution and health data, undermining both model performance and validation.
• Underused methods: While SHAP is common, other approaches like LIME and causal xAI techniques are rarely integrated, leaving gaps in understanding causal pathways between pollution exposure and health risks.

The study calls for wider adoption of causal xAI-ML to move beyond correlation and generate actionable insights for interventions.

A roadmap for reliable and trusted AI in public health

The authors propose a set of best practices to ensure that AI-powered models are not only accurate but also trusted and usable in real-world contexts.

Key recommendations include:

• Harmonizing metrics: For regression tasks, use R², RMSE, and MAE; for classification tasks, report AUC, sensitivity, and specificity. Consistency in metrics is critical for comparing models.
• Transparent validation: Adopt rolling-origin evaluation for forecasts and nested cross-validation for smaller datasets to reduce bias and improve reproducibility.
• Publishing explainability outputs: Share SHAP values, feature rankings, and counterfactuals so domain experts can verify whether model explanations align with established knowledge.
• Triangulating approaches: Apply multiple xAI methods to the same problem and validate results with expert input to strengthen confidence in model-driven decisions.
• Combining uncertainty with performance: Incorporate uncertainty quantification alongside predictive scores to help decision-makers judge model reliability.
• Human-in-the-loop oversight: Keep clinicians and environmental experts involved in interpreting and acting on AI-generated insights, especially in high-stakes decisions.

The authors point out that models must be both transparent and operationally validated before they can meaningfully guide clinical care or public health interventions.