**8. Discussion and conclusion**

Brain is a self-sustained oscillator where individual neurons oscillate at certain harmonics. Major rhythms of motor outputs generate through bifurcation. Several linear (spectral coherence and cross-correlation) and nonlinear (phase synchronization, mutual information and entropy) measures have been adopted to measure the oscillations [100]. Structural and functional connectivity of the brain works in coherence to perform a common action. Structural connectivity relates to the physical connection between different regions of the brain, while functional connectivity is the correlation between various regions over time that shows dynamic behavior [101].

During cognitive tasks requiring attention, certain brain regions become more active while the other regions activity decreases. A flashing or flickering visual stimulus eliciting event or evoked potential (P300 or SSVEP) increases activity in frontal and visual cortex. Due to more repetitive mental tasks, brain activity increases in the specific region, whereas activity in the other regions is reduced. The reason for reduction may be due to unrelated or difficult tasks [102]. This increase in brain activity corroborates growth in working memory of brain illustrating brain dynamic states. Brain changes its state according to the environment similar to an artificially intelligent machine which adapts to learn from the input attributes without being explicitly programmed. So, a BCI illiterate at one point of time may adapt to learn with continuous trials due to dynamic brain states [103].

Human brain is a nonlinear dynamic system behaving as a chaotic and fractal system [104]. Therefore, EEG is a complex, nonlinear, and nonstationary signal. However, EEG signals have been analyzed based on linear/nonlinear and stationary/nonstationary techniques for feature extraction and classification. Fourier transforms, wavelet decomposition, power spectral density, autoregression, CCA, LDA, SVM, and CSP are some of the linear methods for EEG classification. However, only commonly used linear classifiers in hybrid BCI are discussed here. Due to the dynamical nature of the brain and the associated EEG signal, widely used linear approaches are not enough to obtain promising results. Therefore, the nonlinear dynamical behavior of EEG should be carefully considered during brain signal analysis. EEG signals need to be analyzed along with the dynamic states to reveal additional features that cannot be assessed with the linear methods.

EEG signals were analyzed dynamically in [105] to identify and code the attractors related to mental states using artificial neural network. It was shown that binary patterns of attractors resulting from neural firing of identical cognitive or sensory stimuli are similar, but they might appear as distinct features with different stimuli. The chaotic behavior of attractors highlights the fact that the neural signals are coherent. Indeed, brain dynamics has unveiled an emerging area of research to quantify information using attractors and fractals from EEG signals for useful operations applying hybrid BCI. It should be noted that attractors and fractals are the dynamic variables to measure complexity (correlation dimension and Hurst exponent) and stability (Lyapunov exponent and entropy) of EEG data [106–108]. The phenomenon of brain receiving the sensory inputs, storing the information, and processing output is still unknown. Hybrid BCI can be an efficient tool to transform the interesting brain dynamics into actions.

In this chapter, hybrid BCI is reviewed, and advancements from single BCI system to hybrid BCI systems, associated signal processing methods, usages, shortcomings and future scopes are discussed. The common hybrid systems based on signal combinations as well as operation methods, their performances, and improvements are mentioned. Statistical analysis of BCI and hybrid BCI related to P300 and SSVEP are illustrated based on publications. Transitioning from laboratory to the possible commercial applications is well discussed along with the limitations onward. This review illustrates P300, SSVEP, and MI which are mostly used EEG signals for BCI. Simultaneous operation is very common in P300-SSVEP hybrid and sequential operations are incorporated mostly in MI-related hybrid experiments. Average accuracy and ITR among reviewed hybrid BCI papers are 90% and 28 bits/min, respectively, which demand the need for a more efficient hybrid BCI system.
