Kuo Cheng-Pin, Fu Joshua S, Liu Yang
Industrial Technology Research Institute, Hsinchu, Taiwan.
Department of Civil and Environmental Engineering, University of Tennessee Knoxville, Knoxville, TN, United States.
Front Public Health. 2025 Mar 12;13:1436838. doi: 10.3389/fpubh.2025.1436838. eCollection 2025.
As air pollution events increasingly threaten public health under climate change, more precise estimations of air pollutant exposure and the burden of diseases (BD) are urgently needed. However, current BD assessments from various sources of air pollutant concentrations and exposure risks, and the derived uncertainty still needs systematic assessment. Owing to growing health and air quality data availability, machine learning (ML) may provide a promising solution. This study proposed an ML-measurement-model fusion (MMF) framework that can quantify the air pollutant biases from the Chemical Transport Modeling (CTM) inputs, and further analyze the BD biases concerning various sources of air pollutant estimations and exposure risks. In our study region, the proposed ML-MMF framework successfully improves CTM-modeled PM (from R = 0.41 to R = 0.86) and O (from R = 0.48 to R = 0.82). The bias quantification results showed that premature deaths in the study region are mainly biased by boundary conditions (Improvement Ratio, IR = 99%) and meteorology (91%), compared with emission and land-use data. The results of further analysis showed using observations only (PM: 17%; O: 56%) or the uncorrected CTM estimations (PM: -18%; O: 171%) contributed more BD biases compared with employing averaged risks without considering urbanization levels (PM: -5%; O: -4%). In conclusion, employing observations only, uncorrected CTM estimations, and homogeneous risks may contribute to non-negligible BD biases and affect regional air quality and risk management. To cope with increasing needs of finer-scale air quality management under climate change, our developed ML-MMF framework can provide a quantitative reference to improve CTM performance and priority to improve input data quality and CTM mechanisms.
在气候变化背景下，空气污染事件对公众健康的威胁日益增加，因此迫切需要更精确地估计空气污染物暴露情况和疾病负担（BD）。然而，目前基于各种空气污染物浓度来源和暴露风险的BD评估以及由此产生的不确定性仍需要系统评估。由于健康和空气质量数据的可用性不断提高，机器学习（ML）可能提供一个有前景的解决方案。本研究提出了一种ML-测量模型融合（MMF）框架，该框架可以量化化学传输模型（CTM）输入中的空气污染物偏差，并进一步分析与各种空气污染物估计来源和暴露风险相关的BD偏差。在我们的研究区域，所提出的ML-MMF框架成功地改善了CTM模拟的PM（相关系数从R = 0.41提高到R = 0.86）和O（相关系数从R = 0.48提高到R = 0.82）。偏差量化结果表明，与排放和土地利用数据相比，研究区域内的过早死亡主要受边界条件（改善率，IR = 99%）和气象条件（91%）的偏差影响。进一步分析的结果表明，与不考虑城市化水平采用平均风险相比，仅使用观测数据（PM：17%；O：56%）或未校正的CTM估计值（PM：-18%；O：171%）会导致更多的BD偏差。总之，仅使用观测数据、未校正的CTM估计值和同质风险可能会导致不可忽视的BD偏差，并影响区域空气质量和风险管理。为了应对气候变化下对更精细尺度空气质量管控日益增长的需求，我们开发的ML-MMF框架可以为提高CTM性能提供定量参考，并为提高输入数据质量和CTM机制提供优先事项参考。
