Norwegian Institute for Water Research (NIVA), Gaustadalléen 21, 0349 Oslo, Norway; Queensland Alliance for Environmental Health Science (QAEHS), University of Queensland, 20 Cornwall Street, Woolloongabba, QLD 4012, Australia.
Queensland Alliance for Environmental Health Science (QAEHS), University of Queensland, 20 Cornwall Street, Woolloongabba, QLD 4012, Australia.
Talanta. 2019 Apr 1;195:426-432. doi: 10.1016/j.talanta.2018.11.039. Epub 2018 Nov 15.
Guaranteeing clean drinking water to the global population is becoming more challenging, because of the cases of water scarcity across the globe, growing population, and increased chemical footprint of this population. Existing targeted strategies for hazard monitoring in drinking water are not adequate to handle such diverse and multidimensional stressors. In the current study, we have developed, validated, and tested a machine learning algorithm based on the data produced via non-targeted liquid chromatography coupled with high resolution mass spectrometry (LC-HRMS) for the identification of potential chemical hazards in drinking water. The machine learning algorithm consisted of a composite statistical model including an unsupervised component (i.e. principal component analysis PCA) and a supervised one (i.e. partial least square discrimination analysis PLS-DA). This model was trained using a training set of 20 drinking water samples previously tested via conventional suspect screening. The developed model was validated using a validation set of 20 drinking water samples of which 4 were spiked with 15 labeled standards at four different concentration levels. The model successfully detected all of the added analytes in the four spiked samples without producing any cases of false detection. The same validation set was processed via conventional trend analysis in order to cross validate the composite model. The results of cross validation showed that even though the conventional trend analysis approach produced a false positive detection rate of ≤5% the composite model outperformed that approach by producing zero cases of false detection. Additionally, the validated model went through an additional test with 42 extra drinking water samples from the same source for an unbiased examination of the model. Finally, the potentials and limitations of this approach were further discussed.
保障全球人口的清洁饮用水正变得越来越具有挑战性，因为全球范围内存在水资源短缺、人口增长以及人口化学足迹增加等问题。现有的饮用水危害监测靶向策略不足以应对如此多样化和多维度的压力源。在本研究中，我们开发、验证和测试了一种基于非靶向液相色谱与高分辨质谱（LC-HRMS）产生的数据的机器学习算法，用于识别饮用水中的潜在化学危害。该机器学习算法由一个组合统计模型组成，包括无监督组件（即主成分分析 PCA）和有监督组件（即偏最小二乘判别分析 PLS-DA）。该模型使用经过传统可疑物筛查测试的 20 个饮用水样本的训练集进行训练。使用包含四个不同浓度水平的 15 个标记标准的 20 个加标饮用水样本的验证集对开发的模型进行验证。该模型成功地检测到了四个加标样本中所有添加的分析物，没有产生任何误报。同样的验证集通过传统的趋势分析进行处理，以交叉验证复合模型。交叉验证的结果表明，尽管传统的趋势分析方法产生的假阳性检出率≤5%，但复合模型通过零误报的结果优于该方法。此外，该验证模型还经过了来自同一水源的 42 个额外饮用水样本的额外测试，以对模型进行无偏检查。最后，进一步讨论了该方法的潜力和局限性。
