Comparison of machine learning methods in forecasting and characterizing the birch and grass pollen season

Daniel Bulanda, Małgorzata Bulanda, Małgorzata Sacha, Adrian Horzyk and Dorota Myszkowska

PLOS ONE, 2026, vol. 21, issue 2, 1-22

Abstract: The primary approach to the treatment of seasonal allergic diseases involves minimizing exposure to allergens and initiating early personalized therapy. The medication should be introduced about 7 days before the start of the pollen season and intensified during the period of the highest concentrations of sensitizing pollen. Therefore, forecasts for the concentration of pollen that causes clinical symptoms are of indisputable value to both doctors and patients. The study was carried out in Krakow (Southern Poland) with birch (Betula) and grasses (Poaceae) pollen data collected using the volumetric method in 1991-2024. The following meteorological data were collected and used in the study: temperature (mean, minimum and maximum), humidity, cloud cover, sunshine duration, mean wind speed, mean pressure at sea level, global radiation and snow depth. Eight machine learning models from four distinct families (lazy, linear, tree-based, and deep learning) were chosen to estimate the probability of the occurrence of pollen concentration within specified categories. These predictions were based on meteorological data combined with pollen concentration levels in the preceding days. Using the occurrence of pollen concentration in the selected categories as the target variable, the top-performing models achieved accuracies of 92.2%, 88.3%, and 87.2% for 1-day, 4-day, and 7-day forecasts of Betula pollen, respectively. Similarly, for Poaceae pollen, the models achieved 86.1%, 81.8%, and 80.0% accuracy for predictions of 1 day, 4 days, and 7 days ahead, respectively. In addition, a feature importance analysis and an association rule mining were performed to explain the dependencies between pollen concentration and meteorological variables. The tested machine learning methods achieve results that allow for satisfactory efficiency in predicting pollen concentration for up to seven consecutive days. The best-performing machine learning methods were boosted trees, associative knowledge graphs, and deep neural networks with memory cells.

Date: 2026
References: Add references at CitEc
Citations:

Downloads: (external link)
https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0332093 (text/html)
https://journals.plos.org/plosone/article/file?id= ... 32093&type=printable (application/pdf)

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:plo:pone00:0332093

DOI: 10.1371/journal.pone.0332093

Access Statistics for this article

More articles in PLOS ONE from Public Library of Science
Bibliographic data for series maintained by plosone ().