In the last posts, I wrote a summary of my (in-progress) dissertation and my proposal to the industry in developing search engines that, through wearable computing devices (such as Emotiv, Interaxon) would scan brain waves, compute the dimensions of emotions, and feed the results back into the search engines, in order to improve search results based on the positive/negative neurological feedback received, sort of a machine/computer learning process.
This I called the Smart Neuro (Affective) Search engine, that essentially would be an extension of the human brain onto search engines. In this post, I intend to provide an introduction of some aspects of these EEG signals and some of the data analysis involved.
In recognizing the dimensions of emotions through EEG devices (i.e. how these may correlate with high and low frequency brain waves), overall, studies appear to suggest that valence states may be measured through alpha asymmetry (present on states of low cognitive load) on the frontal lobes. On the other hand, a good indicator of the arousal states appears to be the ratio of beta-alpha waves (present on states of high cognitive load) on the frontal lobes.
For the purpose of my own research, I use the Emotiv neuroheadset. The very first thing that is needed is the raw EEG data. This data usually is not clean and some preprocessing steps are needed. These preprocessing steps include applying high-pass and low-pass filters. A high-pass filter helps remove the low frequencies and a low-pass filter helps remove the high frequency brain waves, such as Gamma waves.
Once the signals are preprocessed, we would need to divide them in chunks of time in order to extract features out of each one of these pieces. MatLab has many functions for filtering these signals where one could set band pass filters (e.g. alpha waves are between 8Hz and 12Hz).
It is also worth remembering that each of the sensory channels of the EEG neuroheadset, collects a spectrum of brain waves. These spectrums vary among the different types of brain waves. For example, one single channel may present different brain wave frequencies, each differing over time. Therefore, in order to extract these frequency bands, the spectrum should be computed through the Fourier Transform.
To sum up, careful data processing and analysis must be done when collecting EEG raw data sets. To use the EEG raw data, it would require us to 1) pass each channel through a high pass filter and 2) perform a transform on the data and 3) filter for a key frequency band.
In my future posts, I will cover additional details, as these data analysis can get quite complicated.
Read more about the exciting and emerging wearable computing and AI devices here: Brain controlled airplanes, neurogaming, robots that learn behavior, and MindRDR.