Dynamical system based compact deep hybrid network for classification of Parkinson disease related EEG signals

Electroencephalogram (EEG) signals accumulate the brain's spiking activities using standardized electrodes placed at the scalp. These cumulative brain signals are chaotic in nature and vary depending upon current physical and/or mental activities. The anatomy of the brain is altered when dopamine releasing neurons die because of Parkinson Disease (PD), a neurodegenerative disorder. The resulting alterations force synchronized neuronal activity in β frequency components deep within motor region of the brain. This synchronization in the motor region affects the dynamical behavior of the brain activities, which induce motor related impairments in patient's limbs. Identification of reliable bio-markers for PD is active research area since there are no tests or scans to diagnose PD. We use embedding reconstruction, a tool from chaos theory, to highlight PD-related alterations in dynamical properties of EEG and present it as a potentially reliable bio-marker for PD related classification. We use Individual Component Analysis (ICA) to demonstrate that the strengthened synchronizations can be cumulatively collected from EEG channels over the motor region of the brain. We use this information to select the 12 EEG channels for classification of On and Off medication PD patients. Additionally, there is the strong synchronization between amplitude of higher frequency components and phase of β components for PD patients. This information is used to improve the performance of this classification. We apply embedding reconstruction to design a new architecture of a deep neural network called Dynamical system Generated Hybrid Network. We report that this network outperforms the state of the art in terms of classification accuracy of 99.2%(+0.52%) with approximately 24% of the computational resources. Apart from classification accuracy, we use well known statistical measures like specificity, sensitivity, Matthews Correlation Coefficient (MCC), F1 score, and Cohen Kappa score for the analysis and comparison of classification performances.