I am trying to analyze data from EEG electrodes to understand how brain activity changes in different coginitive states (for simplicity, assume that there are only 2 states: baseline (A) and chanelled attention (CA) - a state which was induced in study subjects by making them play a video game that required high focus). The goal of my analysis is to predict in which of states the subject is based on EEG signal.
When I plotted the data for both states separately, I noticed that the main difference between them is frequency of signal. My assumption is that it corresponds to different types of brain waves (that alpha/beta or other, for example alpha waves are observed in people who are meditating). My plan is to:
- First split training data into many time series with respect to state (A or CA) and participant and for each series use Fourier Transform to obtain the dominant frequency.
- If everything goes as planned, observe that some frequencies are specific to certain state.
- For testing data use detected frequency to predict the cognitive state.
The problem is that testing dataset is not a set of timeseries but rather one big timeseries. I am supposed to give probability of each cognitive state for each timepoint in test data. How can I compute "local frequency" at one timepoint? Or detect sudden changes in frequency? My approach would be to use time window around each timepoint, but I'm not sure if it's good approach - for example, what if changes are too brief?