HomeAbstract of 7th-ccgconfAn Interpretable Artificial Intelligence Framework to Investigate the Interactions Between Meteo-Pollution Data and Emergency Department Admissions

An Interpretable Artificial Intelligence Framework to Investigate the Interactions Between Meteo-Pollution Data and Emergency Department Admissions

  • Jul 22, 2026
  • Posted by:
  • Category: Abstract of 7th-ccgconf
No Comments


Abstract Book of the 7th World Conference on Climate Change and Global Warming

Year: 2026

[PDF]

An Interpretable Artificial Intelligence Framework to Investigate the Interactions Between Meteo-Pollution Data and Emergency Department Admissions

Prof Dr Vito Telesca

ABSTRACT:

This study proposes an interpretable Machine Learning framework to analyze the nonlinear relationships between meteorological parameters (T, P, RH), air pollutants (CO, NO₂, O₃, PM₁₀), and Emergency Department admissions for cardiorespiratory diseases (CRD), cardiovascular diseases (CVD), respiratory diseases (RD), and total admissions (TOT) in the Metropolitan City of Bari over the period 2013–2024. Daily data were preprocessed using a 7-day Exponential Moving Average (EMA-7) to reduce background noise. The modeling framework is based on a rigorous validation protocol comparing a tree-based model (XGBoost) and a Deep Learning architecture (GAMI-Net), validated through Monte Carlo simulations with 11 independent seeds and 5-fold Cross-Validation. The XGBoost model achieved test-set R² values ranging from 0.76 to 0.86; similarly, RMSE and MAPE ranged from 1.14 to 13.50 and from 0.06 to 0.18, respectively. XGBoost interpretability was assessed using Explainable Artificial Intelligence (XAI) techniques, identifying CO, T, and P as the main predictors. SHAP (SHapley Additive exPlanations) and PDP (Partial Dependence Plots) analyses were employed to evaluate risk direction and derive epidemiological thresholds, identified as CO ≥ 0.5 mg/m³, P ≤ 1006 hPa, and T ≤ 11 °C. GAMI-Net highlighted T×O₃ as the dominant interaction among environmental predictors, with increases of up to +3.33 patients/day. These findings support the use of interpretable AI approaches for identifying environmental thresholds that may contribute to public health early-warning systems.

Keywords: Epidemiology; Explainable Artificial Intelligence; Feature Importance; health risk thresholds; Machine-Deep learning