In this project we examined different analysis techniques of EEG (electroencephalogram) signals, sampled from patients undergoing TMS (Transcranial Magnetic Stimulation) treatment.
The patients were diagnosed with varying levels of MDD (Major Depressive Disorder). The depression
levels were examined by a psychiatrist, using HDRS (Hamilton Depression Rating Scale).
The project goal is to find features selection from the patient’s EEG signals, so that a machine learning
algorithm with find a correlation between these features and the patient’s HDRS.
The main approach of the project was to use the empirical covariance matrices of the EEG signals, and
utilization of manifold learning techniques and nonlinear dimensionality reduction to map the matrices in another vector space, so that they can be fit to the HDRS. We implemented this approach by defining a basic algorithm and investigated the contribution of various signal processing techniques and operations on the covariances to find an efficient fit.
Most of the analysis techniques activated on the signals did not produce notable results. Nevertheless,
We observed that under an appropriate mapping, there is a difference in brain activity between patients.
This difference illustrates the need for an algorithm that will allow comparison of the different subjects,
for example, Parallel Transport (PT), which we used. Additionally, we found evidence that a separation to , 𝛽, 𝜃 brain waves holds information about each patient’s stage in the treatment.