An Assessment of Modeling Dynamics in EEG Signals Using Learned Sparse Nonlinear Representations
Dragos Constantin Popescu, Andrei Mateescu, Ioana Livia Stefan, Ioana Miruna Vlasceanu, Ioan Ştefan Sacală, Ioan Dumitrache · 2024
This work focuses on the challenge of modeling the nonlinear dynamics components of EEG signals, with the long term goal to extract valuable information regarding brain state behavior or emotional and psychological condition. At the core of the approach is the SINDy (Sparse Identification of Nonlinear Dynamics) method, in which the task of choosing the adequate nonlinear signal transformations and hyperparameters of the sparse regression is accomplished by using a Genetic Algorithm and a Bayesian Optimization. The performance of the two solutions is analyzed and compared against the MindBigData-EP dataset. The first results show that very good accuracy in modeling the dynamics of EEG signals, of over 80% on 60 seconds long recordings, is achieved, however, the precision is a trade-off against the complexity.