[dipl] Perustieteiden korkeakoulu / SCI

Permanent URI for this collectionhttps://aaltodoc.aalto.fi/handle/123456789/21

Browse

Browsing [dipl] Perustieteiden korkeakoulu / SCI by Subject "2-person neuroscience"

Filter results by typing the first few letters
Now showing 1 - 1 of 1
  • Relief-based feature evaluation and selection for hyperscanning MEG
    (2020-10-19) Ruotsalainen, Ari
     Perustieteiden korkeakoulu | Master's thesis
    In this thesis, I have investigated the possibility of utilizing Relief-based feature selection algorithms (RBA) for the analysis of hyperscanning magnetoencephalography (MEG) dataset. The Relief-based algorithms are a family of multivariate feature evaluation algorithms that take high-order feature interactions into account during the evaluation process. The hyperscanning MEG or 2-person MEG is a recently proposed method, where the neural signals of two subjects are recorded simultaneously, for example, during social interaction. The RBAs property of taking high-order feature interactions into account during feature evaluation process should prove advantageous in finding interesting points-of-interests such as synchronous inter-brain activation patterns in the hyperscanning MEG data. This result could be achieved by narrowing the inspection of features extracted from the hyperscanning MEG data into a handful of features deemed highly relevant by the used RBA algorithm. The validity of the proposed concept is being investigated by utilizing two RBA algorithms: Relief-F and Tuned Relief-F (TuRF). The algorithms are used to evaluate features extracted from the hyperscanning MEG dataset arranged into leader–follower and joint-leadership classification task. The extracted features that are being evaluated include sensor-level average power estimates in five frequency bands: theta, low alpha, high alpha, beta and gamma. The performance of feature evaluations obtained with the two RBAs are inspected by training a number of support vector machines (SVM). In addition, the stability and topography of highly evaluated features are investigated. Finally, the behavioral data analysis using finger attached accelerometers was used to ensure that the observed data contained valid class labels for the intended classification task. The results show that the SVMs are able to achieve encouraging levels of predictive accuracy in the leader–follower and joint-leadership classification task by utilizing the highly ranking features evaluated with both RBAs. The two RBAs show minimal difference in the performance during dyad-level analysis, but the increase in the amount of available data during group-level analysis seem to favor TuRF algorithm. In the future, the quality of interactions between highly evaluated features should be investigated.