Machine learning-based feature combination analysis for odor-dependent hemodynamic responses of rat olfactory bulb

Abstract

Rodents have a well-developed sense of smell and are used to detect explosives, mines, illegal substances, hidden currency, and contraband, but it is impossible to keep their concentration constantly. Therefore, there is an ongoing effort to infer odors detected by animals without behavioral readings with brain-computer interface (BCI) technology. However, the invasive BCI technique has the disadvantage that long-term studies are limited by the immune response and electrode movement. On the other hand, near-infrared spectroscopy (NIRS)-based BCI technology is a non-invasive method that can measure neuronal activity without worrying about the immune response or electrode movement. This study confirmed that the NIRS-based BCI technology can be used as an odor detection and identification from the rat olfactory system. In addition, we tried to present features optimized for machine learning models by extracting six features, such as slopes, peak, variance, mean, kurtosis, and skewness, from the hemodynamic response, and analyzing the importance of individuals or combinations. As a result, the feature with the highest F1-Score was indicated as slopes, and it was investigated that the combination of the features including slopes and mean was the most important for odor inference. On the other hand, the inclusion of other features with a low correlation with slopes had a positive effect on the odor inference, but most of them resulted in insignificant or rather poor performance. The results presented in this paper are expected to serve as a basis for suggesting the development direction of the hemodynamic response-based bionic nose in the future.