Over the past decades, the threat posed to human health and the aquatic environment by emerging micropollutants (MPs), such as pharmaceuticals, pesticides, hormones, and perfluorinated compounds (PFCs), has increased significantly (Jian et al. , 2017; Jiang et al., 2013; Le et al., 2013). The presence of these pollutants has been reported in the water cycle at trace levels ranging from ng/L to μg/L, making them particularly difficult to detect and monitor (Aguera et al., 2013; Lin et al. , 2016). Therefore, the scientific community's interest in efficient and accurate quantitative measurement of MPs has increased significantly. In particular, the use of gas chromatography-liquid chromatography combined with/or liquid chromatography without mass spectrometry has been widely studied and successfully implemented for MP analysis (Aguera et al., 2013; Thomas et al., 2022).
Although chromatography is a reliable and powerful analytical technique, the performance and quantitative accuracy of this method is highly dependent on various factors such as sample preparation (Niu et al., 2018; Shen et al., 2022; Tran et al., 2013). For example, syringe filters are traditionally used in sample preparation to remove impurities, thereby preventing column clogging. However, the materials used to manufacture the syringe can have a significant impact on measurement accuracy, leading to an underestimation of MP concentrations (Dong et al., 2022; Hai et al., 2011; Hebig et al., 2014). For example, Sörengård et al. (2020) reported significant losses (up to 100%) of 21 poly- and perfluoroalkyl substances depending on the type of syringe filter and water matrix used. Dong et al. (2022) observed losses of 7–52% of various antibiotics such as quinolones and macrolides in aqueous phase samples when using hydrophobic polytetrafluoroethylene syringe filters. Additionally, losses of pesticides, hormones, and various organic MPs were observed in syringe filters made of glass fiber, cellulose acetate, and nylon (Carlson and Thompson, 2000; Godby and Conklin, 2016; Hebig et al., 2014 ).
Most of the previous studies focused on investigating the loss of drugs, hormones, and PFCs in a single filter or across multiple filters (Chandramouli et al., 2015; Dong et al., 2022 ; Hebig et al., 2014). Several studies have considered trends in syringe filter losses depending on pore size and water matrix (Hebig et al., 2014; Sörengård et al., 2020). However, there has not been a thorough systematic analysis of the characteristics that influence the loss based on the physicochemical properties of various MPs, such as disinfection byproducts and toxic substances. Additionally, the effects of other operating parameters such as pH, flow rate, and filter diameter have not been well studied. Therefore, limitations exist in determining the relative importance and ranking of each feature and the operating parameters.
To fill this knowledge gap, in our study, we implemented a novel data analysis approach, machine learning, and combined it with experimental results. This provided new insights into the factors influencing MP loss. Additionally, this method can reduce analysis time and labor costs by identifying similarities between MPs, importance and ranking of each element, and suitability of filters. For this purpose, Random Forest Classifier, a supervised machine learning algorithm with good accuracy, ability to handle mixed feature types, and inference performance on tabular datasets, was adopted (Shwartz-Ziv and Armon, 2022). The results of the feature importance analysis were cross-checked using the Kruskal-Wallis test, a non-parametric statistical test. This test verifies the reliability of the analysis by comparing the results of the machine learning approach with traditional methods (statistical tests) to prevent unexpected errors in the analysis. It was carried out. The importance of features in selecting appropriate filters was analyzed using SHapley Additive exPlanation (SHAP) analysis. The combined evaluation of SHAP analysis and Kruskal-Wallis test facilitates the ranking of features such as physicochemical properties and operating parameters that influence MP loss.
Overall, the objectives of this study are to (1) investigate the effects of eight different syringe filter materials on the loss of 70 structurally diverse MPs; Almost half of them have not been studied before. (2) Use machine learning approaches to classify and provide insight into physicochemical properties and operating parameters and the importance of features in selecting appropriate filters. (3) evaluate the influence of filter characteristics (such as pore size and diameter) and operational parameters such as pH, flow rate, sample volume, solvent type, prewash, and water matrix on MP loss;
