A brain-computer interface based on an electroencephalogram is an emerging tool for conducting neurorehabilitation therapies for people with motor disabilities. The electroencephalogram recorded in the sensorimotor area has characteristic rhythms such as beta (12–30 Hz) and mu (8–12 Hz) that change when a person attempts to perform a movement, resulting in a decrease in amplitude called event-related desynchronization (ERD). In this paper, different methods for estimating the power spectral density of the electroencephalogram to detect ERD are investigated, both parametric and non-parametric methods. Own recordings, BCI Competition III database and a linear Fisher discriminant were used as classifiers and the true positive rate metric was used for performance evaluation for the improvement of brain-computer interface systems. The study aims to improve feature extraction techniques in BCIs for motor neurorehabilitation. The study shows that the best performance in PSD estimation methods was achieved with parametric ARMA methods. In particular, the Burg method with an order of 8, the covariance method with an order of 6, the Yule-Walker method with an order of 6, and an ARMA model with a numerator of order 4 and a denominator of order 8 showed the most promising results. These parametric methods outperformed the non-parametric methods, with the ARMA model being particularly effective in detecting ERD.