This article explores methods for recognizing neurophysiological signals from the electroencephalogram related to the preparation and the execution of voluntary movements in the human brain. The experiment involved participants performing a complex self-initiated movement, comprising pressing a button and touching a marker located around a transparent partition. Collected data underwent preprocessing, including filtering, Laplacian transformations, and the extraction of temporal epochs lasting 1500 ms prior to the button press. Frequency and temporal features were extracted from the prepared epochs. Various machine learning models were employed, including regularized linear models, random forests, and support vector machines. Hyperparameter optimization was performed using cross-validation to ensure a robust and reliable evaluation of the models. To address class imbalance, methods, such as upsampling of minority classes, and adding penalties in the loss function were applied. Balanced accuracy was chosen as the target metric, which accounts for the disparity between the number of positive and negative epochs and provides an unbiased evaluation of classification performance. The SHapley Additive exPlanations method, which leverages Shapley values to explain machine learning model predictions, was utilized to provide insights into feature importance and their contributions to classification outcomes. The best-performing model demonstrated a balanced accuracy of 72% for right-hand presses and 77% for left-hand presses when averaged among all participants.