High-Accuracy Detection of Odor Presence from Olfactory Bulb Local Field Potentials via Deep Neural Networks

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Odor detection is crucial for food safety, environmental monitoring, and medical diagnostics.
• Existing artificial sensors struggle with complex odor mixtures, and non-invasive recordings lack single-trial reliability.

Purpose of the Study:

• To test if spectral features of local field potentials (LFPs) are adequate for robust single-trial odor detection.
• To determine if olfactory bulb signals alone are sufficient for odor detection.

Main Methods:

• An ensemble of complementary one-dimensional convolutional neural networks (ResCNN and AttentionCNN) was developed.
• The model decoded odor presence from multichannel olfactory bulb LFPs in awake mice.
• The framework was tested on 2349 trials across seven mice.

Main Results:

• The ensemble model achieved a mean accuracy of 86.2%, an F1-score of 85.3%, and an AUC of 0.942.
• Performance substantially outperformed previous benchmarks.
• t-SNE visualization confirmed the capture of biologically significant olfactory signatures.

Conclusions:

• Robust single-trial odor detection is feasible using extracellular LFPs.
• Deep learning models show potential for a deeper understanding of olfactory representations.
• The study validates the sufficiency of olfactory bulb signals and spectral LFP features for odor detection.