Sagatdinov Artur Rinatovich

RESEARCH, TEACHING, or OTHER INTERESTS

Human-Computer Interaction, Artificial Intelligence, Signal Processing, Computer Science

Scopus Publications

Temporal and frequency features play different role in recognizing preparation of voluntary movements from electroencephalogram
Artur Sagatdinov, Mikhail Lipkovich, Veronika Knyazeva, and Aleksander Aleksandrov
Springer Science and Business Media LLC



Detection of the intention to perform a two-staged movement from EEG signals
Artur Sagatdinov, Veronika Knyazeva, Mikhail Lipkovich, and Aleksander Aleksandrov
IEEE
The article explores methods for recognizing neu-rophysiological signals related to the preparation and execution of voluntary movements in the human brain. The experiment involved participants performing a complex self-initiated move-ment, comprising pressing a button and touching a marker around a transparent partition. Various machine learning models were utilized, including linear models with regularization, random forests, and support vector machines. The introduction of a “stacking” model facil-itated the incorporation of new feature types without complete retraining of base models. Hyperparameter optimization was conducted using cross-validation. To address class imbalance, upsampling methods and penalties in the loss functions were applied. Balanced accuracy was chosen as the target metric, considering the disparity between the number of positive and negative epochs. The SHAP method was employed for results interpretation. The best-performing model demonstrated a balanced accuracy of 72% for right-hand presses and 77% for left-hand presses.


Evoked Potentials Detection During Self-Initiated Movements Using Machine Learning Approach
Mikhail Lipkovich, Veronika Knyazeva, Aleksander Aleksandrov, Nadezhda Shanarova, Artur Sagatdinov, and Alexander Fradkov
IEEE
This paper focuses on the issue of determining the intention to initiate a movement using machine learning approach. In the experiment under consideration, the subjects make spontaneous movements, and their brain activity is recorded using an electroencephalogram. The problem is formulated as a binary classification problem where the model should determine whether the given segment of signal precedes the movement. In order to perform this task the features from both the time and frequency domains were extracted. The process of feature extraction is parametrized through a special procedure, and the parameters of this procedure are selected through a grid-search technique along with model hyperparameters. The best metrics were obtained using Random Forest model that had a balanced accuracy of 77% on the test set. Moreover, the impact of the attention system during the conducted experiments was analyzed. Two paradigms were employed: the oddball paradigm, where deviant stimuli, which occurens leads to the involuntary attention activation, were introduced, and the control paradigm, which omitted any deviant stimuli. Experiments have shown that models perform better under oddball paradigm.


Application of the "Stripe" Algorithm for Online Decoding of the EEG Patterns
M. M. Lipkovich and A. R. Sagatdinov
New Technologies Publishing House
In this paper, we consider the problem of determining the hand with which the subject intends to make a movement according to the signals of the electroencephalogram. The relevance of the task is due to the wide spread of brain-computer interfaces, where electroencephalography is one of the main non-invasive methods for obtaining signals from the brain. To solve the problem, temporal and frequency features are selected from the segments of signals preceding the movement, which are fed to the input of the classification machine learning model. In contrast to the standard supervised learning setup, it is assumed that there is no predefined training data set and the training samples for the model are received one after another. Thus, a situation is simulated in which the model must work with a new subject and adjust to them in real time. The traditional method for training linear models in such a paradigm is stochastic gradient descent. Previously, it was shown that the "Stripe" algorithm developed by Yakubovich for a certain problem has a higher convergence rate than stochastic gradient descent. However, this is achieved by performing algorithm step on each feature of the sample. Thus, that version of "Stripe" is not suitable for working with high-dimensional data. This article discusses another version of "Stripe" that does not have this drawback. It is shown that the proposed algorithm has a higher rate of one learning step compared to traditional linear models based on stochastic gradient descent on the BCI competition II dataset.