Data mining reveals patterns in brain’s reactions to good and bad smells
Activation in piriform cortex and amygdala not the same for everyone
image: Activation patterns in the amygdala in response to pleasant smells. (A to D) Brain scans showing activation patterns. Histograms showing Shapley values of experimental participants. (E) Venn diagram showing number of experimental participants contributing to the four activation patterns shown. Pie chart showing that 40 of the 42 participants contribute to at least one of these patterns.
Credit: Maëlle Moranges et al.
Seeking to shed light on inconsistent brain activation patterns observed in previous studies of smell, a team of researchers in Lyon, France, used data mining techniques to analyze the pleasant or unpleasant odor sensations and corresponding brain scans of a group of 42 participants. They found characteristic patterns in the piriform cortex and amygdala. The group’s research was published August 1 in Intelligent Computing, a Science Partner Journal.
The exceptional model mining method of data analysis, which looks for “exceptions” in data patterns, showed not only which specific brain areas responded to pleasant and unpleasant smells, it also allowed the analysis of differences in brain response between individuals and between subgroups split according to age and sex. In future studies, the same approach could be used for analysis of brains in healthy and pathological states to understand and treat disease.
Everyone is different, thus there are individual differences in brain responses to pleasant and unpleasant smells. However, such differences can be erased or overlooked when data are considered in the aggregate, thus the researchers used exceptional model mining “to identify partitions of data where a model fitted to the target variables is significantly different from this same model applied to the entire dataset.”
To collect the dataset, functional magnetic resonance imaging was conducted on 42 experimental participants in three different age groups while they reacted to six different smells. A device was used to deliver puffs of odorized air timed to coincide with the participant’s breathing. Ratings on a five-point scale of pleasantness and other characteristics were collected for analysis alongside the brain scans.
The data mining process consisted of six steps:
- preprocessing of brain imaging data and participant ratings,
- modeling to transform the data into a graph structure consisting of nodes and edges,
- use of the C-energetics data mining algorithm to identify the activated brain regions corresponding to different ratings,
- statistical filtering to highlight relevant patterns,
- calculation of individual participant contributions to identified patterns using Shapley values, and
- visualization of results.
The study of brain activation in response to smell is complex. It encompasses multiple brain subregions: most notably the anterior and posterior piriform cortex and the amygdala. Moreover, these brain regions respond to a combination of aspects of a smell: valence, which is a measure of the pleasantness or unpleasantness of a smell; quality, which describes the type of smell; and intensity, which is a measure of the strength of the smell. Differing characteristics of experimental participants add to the complexity, which is why it is useful to be able to conduct analysis that takes individual differences into account.