- 學位論文
機器學習預測居民暴露特徵—以土壤及地下水脆弱度為例
Using Machine Learning to Characterize and Predict Exposure Behavior- A Vulnerability Assessment Method for Soil and Groundwater Exposure
摘要
健康風險評估是用來估計人們暴露於危害物質時，所可能承受的不良健康效應的重要工具，然而執行完整之健康風險評估相當耗費資源，特別是在土壤及地下水場址中，污染物的調查與整治都相當困難。另外在土壤及地下水健康風險評估程序中，對於變異性的處理始終不足，變異性屬於族群的自有特徵，來自於個體間在暴露行為上及對污染物之敏感性的不同，將變異性特徵化是相當重要的。本研究使用資料探勘與機器學的概念與技巧，處理大量的健康風險訪查問卷，並結合社會經濟公開資料，找出受體基本資料(年齡、性別、體重、職業及受體型態)、社會經濟資料(收入、自來水普及率、教育程度、人口密度、老年人口比例及耕地率)與暴露行為間的關聯性。本研究使用機器學習方法中的決策樹法，以樹狀結構呈現暴露行為與各因子間之關係，以特徵化民眾對於土壤及地下水之暴露行為之變異性。另外本研究建立了一套完整的脆弱度評估方法，結合暴露性因子與敏感性因子後產出脆弱度地圖，以台中市作為脆弱度地圖的示範地區，並呈現不同接觸型態之暴露性及脆弱度地圖，可提供管理者制定減輕及預防對策。 本研究定義脆弱度為暴露性乘上敏感性，研究方法遵循資料探勘的步驟，包含資料蒐集、資料萃取、模式建立及模式結果闡述與驗證處理暴露性因子，敏感性因子則以統計方法處理年齡資料，分別產生暴露性地圖及敏感性地圖後再結合成脆弱度地圖。 本研究使用決策樹法針對10種不同土壤及地下水暴露行為進行特徵化，結果顯示土地利用為土壤及地下水接觸的重要特徵；農業用地及職業為農業的資料通常具有較高機率的接觸性與較高接觸量；自來水普及率在淋浴泡澡揮發吸入及皮膚接觸之接觸型態關聯性較高；教育程度的提昇可能有助於預防或減輕民眾對於地下水的接觸；土壤與地下水的接觸特徵其實不太相同，需分開討論；地下水接觸判定決策樹AUC值為0.9051，具有優良預測能力。本研究以台中市作為示範區，製作了8種不同接觸型態之暴露性及脆弱度地圖，其中沙鹿區、北屯區、南屯區及西屯區等敏感性較高的區域，脆弱度通常也較高。 本研究改善了健康風險評估中耗費資源且無法對大範圍地區進行評估的缺點，只要有社會經濟公開資料就可以快速地進行評估。本研究方法與結果可使用於高污染性工廠或工業區之管理參考，當確認那些區域的確具有較高之脆弱度時，相關政策必須進行補足或矯正，針對最脆弱之受體進行保護，以實現環境正義。
並列摘要
Health risk assessment is a process to estimate the adverse health effects in human who may be exposed to chemicals in contaminated environmental media. Health risk assessment also plays an important role in facing soil and groundwater contamination events. However, health risk assessment often takes a lot of time and money, especially in soil and groundwater cases, which is more difficult in pollution survey and remediation. In addition, variability is an inherent characteristic of a population; people vary substantially in their exposures and susceptibility to harmful effects of the exposures. Addressing variability is critical for health risk assessment, and it should be better characterized. A simplified assessment method is proposed in this study. To characterize the variability in exposures of a population. Decision tree, one of the most popular methods in machine learning, is used to extract information from 4188 historical questionnaires. Then we choose the most important factors that affect exposure behavior as exposure factors and susceptibility factors from the social and economic factors database, including acceptor basic information (age, sex, weight, occupation and acceptor type) .and socio-economic factors (income, rate of population served by tap water, education, population density, elders and farm rate). This study defines vulnerability as the combination of exposure factors and susceptibility factors, completes a vulnerability assessment, and produces exposure maps and vulnerability maps of 8 different exposure type in Taichung. 10 different exposure behaviors of soil and groundwater are characterized, the result shows that landuse is a important characteristic of soil and groundwater exposure; landuse of farm and occupation of farmer are both more likely to have high exposure; rate of population served by tap water is highly related to exposure type such as inhale exposure when taking a shower or bath and dermal exposure; education may help predict exposure to groundwater; the exposure characteristics of soil and groundwater are not the same and needs to be discuss separately; decision tree of whether exposure to groundwater has good prediction ability, the AUC is 0.9051. The study addressed one of the problem of conducting health risk assessment, which is related to resourse demand and difficulty to assess large area. Vulnerability assessment can be done quickly by socio-economics open data. Vulnerability assessment can be used as guidance for high pollution factories and industrial area management. In addition, when a place is determined to be more vulnerable, relevant policies should be put in place in order to protect the most vulnerable population and achieve environmental justice.