Estimating Inhaled Nitrogen Dioxide from the Human Biometric Response

Abstract

Breathing clean air is crucial for maintaining good human health. The air we inhale can significantly impact our physical and mental well-being, influenced by parameters such as particulate matter and gases (e.g. carbon dioxide, carbon monoxide, and nitrogen dioxide). Building on previous research that explored the effects of particulate matter (PM) in specific environments, analyzed using biometric indicators and machine learning models; this work focuses on the effects and estimation of inhaled nitrogen dioxide (NO<sub>2</sub>). This study involved a cyclist equipped with sensors to monitor various biometric parameters. In addition, an electric car following the cyclist measured the ambient NO<sub>2</sub> levels using an onboard sensor. A total of 329 biometric variables have been taken into account, of which 320 biometric variables are cognitive responses extracted using an electroencephalogram (EEG) and 9 biometric variables are physiological responses extracted using several sensors. Inhaled NO<sub>2</sub> levels are first estimated initially by making use of all 329 variables, then using 9 physiological responses and finally using only 6 of the 9 physiological responses. The study also uses a ranking method to pinpoint which biometric variables most significantly estimate inhaled NO<sub>2</sub> levels. Furthermore, it investigates the linear and non-linear relationship between certain variables and inhaled NO<sub>2</sub>. The general precision of the prediction for the data set was moderate, as indicated by the coefficient of determination (R<sup>2</sup>) and the root mean square error (RMSE) between the true and estimated values of NO<sub>2</sub> to be 0.35 and 5.41 ppb, respectively, in the test set. A higher accuracy in the prediction of lower values of NO<sub>2</sub> levels was qualitatively observed using a scatter diagram and a Quantile-Quantile plot where the data were more plentiful. For more robust conclusions, additional data and refined machine learning models are necessary.