Abstract
Odor sensory evaluation is crucial in the food, textiles, and cosmetics industries. Traditional methods rely on expert human evaluators but suffer from poor repeatability, while modern electronic noses (E-noses) effectively capture objective physicochemical characteristics of odorants but fail to account for human olfactory preferences. This study investigates the integration of human olfactory preferences encoded in electroen- cephalogram (EEG) signals, with objective physicochemical prop- erties captured by E-noses. To this end, EEG data was recorded while subjects smelled mono-molecular odorants, and vibrational spectra (VS), and fingerprint features (FPF) were generated for the same odorants. The fusion of EEG- derived features with VS and FPF enabled multimodal prediction, significantly improving performance compared to using each feature individually. To extract olfactory EEG features, we propose a novel TDA-NOSE framework—Topological Data Analysis (TDA)- based Neural Olfactory Signal Extraction (NOSE)—followed by a 1D-CNN- based classification network. This work is the first investigation of olfactory EGG analysis with TDA. TDA involves phase space reconstruction (PSR) to transform EEG time-series data into a phase space representation, followed by persistent homology to capture its topological properties. Along with olfactory perception prediction, we also show the effectiveness of TDA-NOSE features on olfactory EEG classification. TDA-NOSE features for single- channel EEG achieved 81.36% accuracy in a ten-class and 71% accuracy in binary olfactory EEG classification, significantly outperforming previous benchmarks of 13.5% and 57% achieved using all 64 EEG channels on the dataset. The findings highlight the potential of multimodal feature fusion in improving odor perception prediction and the use of TDA-based features in olfactory EEG analysis. The source codes and trained models are available at: https://github.com/durgesgameta/TDA-NOSE/