Meyer David E, Smith Raymond L, Lanphear Elizabeth, Takkellapati Sudhakar, Chea John D, Ruiz-Mercado Gerardo J, Gonzalez Michael A, Barrett William M
Center for Environmental Solutions and Emergency Response, United States Environmental Protection Agency, Cincinnati, Ohio.
Eastern Research Group, Concord, Massachusetts.
J Occup Environ Hyg. 2025 May;22(5):375-385. doi: 10.1080/15459624.2024.2447320. Epub 2025 Jan 13.
Chemical release data are essential for performing chemical risk assessments to understand the potential exposures arising from industrial processes. Often, these data are unknown or unavailable and must be estimated. A case study of volatile organic compound releases during extrusion-based additive manufacturing is used here to explore the viability of various regression methods for predicting chemical releases to inform chemical assessments. The methods assessed in this work include linear Least Squares, Least Absolute Shrinkage and Selection Operator (LASSO) and Ridge regression, classification and regression tree, random forest model, and neural network analysis. Secondary data describing polymeric extrusion in multiple applications are curated and assembled in a dataset to support regression modeling using default parameters for the various approaches. The potential to add noise to the dataset and improve regression is evaluated using synthetic data generation. Evaluation of model performance for a common test set found all methods were able to achieve predictions within 10%-error for up to 98% of the test sample population. The degree to which this level of performance was maintained when varying the number and type of features for regression was dependent on the model type. Linear methods and neural network analysis predicted the most test samples within 10%-error for smaller numbers of features while tree-based approaches could accommodate a larger number of features. The number and type of features can be important if the desire is to make chemical-specific release predictions. The inclusion of release data from related processes generally improved test set predictions across all models while the use of synthetic data as implemented here resulted in smaller increases in test sample predictions within 10%-error. Future work should focus on improving access to primary data and optimizing models to achieve maximum predictive performance of environmental releases to support chemical risk assessment.
化学释放数据对于进行化学风险评估以了解工业过程中可能产生的暴露至关重要。通常，这些数据未知或不可用，必须进行估算。本文以基于挤出的增材制造过程中挥发性有机化合物的释放为例，探讨各种回归方法在预测化学释放以指导化学评估方面的可行性。本研究评估的方法包括线性最小二乘法、最小绝对收缩和选择算子（LASSO）回归、岭回归、分类与回归树、随机森林模型以及神经网络分析。整理并汇总了描述多种应用中聚合物挤出的二次数据，形成一个数据集，以支持使用各种方法的默认参数进行回归建模。利用合成数据生成评估向数据集中添加噪声并改善回归的潜力。对一个通用测试集的模型性能评估发现，所有方法都能够在10%误差范围内对高达98%的测试样本群体进行预测。当改变回归特征的数量和类型时，这种性能水平的维持程度取决于模型类型。对于较少数量的特征，线性方法和神经网络分析在10%误差范围内预测的测试样本最多，而基于树的方法可以容纳更多的特征。如果希望进行特定化学物质的释放预测，特征的数量和类型可能很重要。纳入相关过程的释放数据通常会提高所有模型的测试集预测结果，而本文所采用的合成数据的使用导致在10%误差范围内测试样本预测的增加幅度较小。未来的工作应专注于改善原始数据的获取并优化模型，以实现环境释放的最大预测性能，从而支持化学风险评估。
