Deep Neural Network-Based Concentration Model for Oak Pollen Allergy Warning in South Korea.

Area of Science:

• Environmental Science
• Allergy and Immunology
• Data Science and Machine Learning

Background:

• Oak pollen is a significant allergen in Korea, with high sensitivity rates among the population.
• Conventional regression models struggle to accurately predict oak pollen concentrations and season dynamics.
• Accurate oak pollen forecasting is crucial for managing allergic rhinitis and improving public health.

Purpose of the Study:

• To develop and evaluate a deep neural network (DNN)-based model for estimating oak pollen concentration.
• To overcome the limitations of traditional regression models in predicting allergenic pollen levels.
• To assess the DNN model's performance against regression and support vector regression (SVR) models.

Main Methods:

• A DNN model was developed using weather factors as input and pollen concentrations as output.
• A bootstrap aggregating ensemble method with 30 members was employed to prevent overfitting and underestimation.
• Model performance was compared using data from 2007-2016, evaluating predictions of concentration, risk, and season length.

Main Results:

• The DNN model achieved a mean absolute percentage error of 5.04% for pollen concentration, significantly lower than regression (11.18%) and SVR (10.37%).
• The DNN model provided more accurate estimations for the start and end dates of the pollen season compared to regression and SVR.
• While overall performance was superior, the DNN model requires improvement in predicting peak pollen concentrations.

Conclusions:

• The developed DNN ensemble model demonstrates superior performance in estimating oak pollen concentrations and season length.
• Further refinement of the DNN model and improved data quality are necessary to enhance peak pollen concentration predictions.
• This advanced modeling approach offers a promising tool for allergy forecasting and public health management in Korea.