Enhancing classification accuracy of HRF signals in fNIRS using semi-supervised learning and filtering

Abstract

This paper introduces a novel approach to enhance the classification accuracy of hemodynamic response function (HRF) signals acquired through functional near-infrared spectroscopy (fNIRS). Leveraging a semi-supervised learning (SSL) framework alongside a filtering technique, the study preprocesses HRF data effectively before applying the SSL algorithm. Collected from the prefrontal cortex, HRF signals capture variations in oxyhemoglobin (oxyHb) and deoxyhemoglobin (deoxyHb) levels in response to odor stimuli and air state. Training the classification model on a dataset containing filtered and feature-extracted HRF signals led to significant improvements in classification accuracy. By comparing the algorithm's performance before and after employing the proposed filtering technique, the study provides compelling evidence of its effectiveness. These findings hold promise for advancing functional brain imaging research and cognitive studies, facilitating a deeper understanding of brain responses across various experimental contexts.

Keywords: Feature extraction; Filter; Hemodynamic response; Olfactory; Semi-supervised learning.