Abstract
The daily pollen forecast provides crucial information for allergic patients to avoid exposure to specific pollen. Pollen counts are typically measured with air samplers and analyzed with microscopy by trained experts. In contrast, this study evaluated the effectiveness of identifying the component pollens using the metabolites extracted from an air-sampled pollen mixture. Ambient air-sampled pollen from Munich in 2016 and 2017 was visually identified from reference pollens and extracts were prepared. The extracts were lyophilized, rehydrated in optimal NMR buffers, and filtered to remove large proteins. NMR spectra were analyzed for pollen associated metabolites. Regression and decision-tree based algorithms using the concentration of metabolites, calculated from the NMR spectra outperformed algorithms using the NMR spectra themselves as input data for pollen identification. Categorical prediction algorithms trained for low, medium, high, and very high pollen count groups had accuracies of 74% for the tree, 82% for the grass, and 93% for the weed pollen count. Deep learning models using convolutional neural networks performed better than regression models using NMR spectral input, and were the overall best method in terms of relative error and classification accuracy (86% for tree, 89% for grass, and 93% for weed pollen count). This study demonstrates that NMR spectra of air-sampled pollen extracts can be used in an automated fashion to provide taxa and type-specific measures of the daily pollen count.