Automated Diagnosis of Parkinson’s Disease from Electroencephalogram Signals Using Adaptive Dilated Convolution Capsule Network with Weighted Optimal Feature Fusion

Electroencephalogram (EEG) signals have been recently used for Parkinson’s Disease (PD) diagnosis, and they have proven to add significant features related to early detection. EEG is a complex signal with various non-linear and non-stationary elements. This work develops an automated PD diagnosis model for managing several symptoms related to the disease and automates the PD detection process for timely detection and diagnosis. At first, the EEG signals are gathered from benchmark sources. Further, the frequency division is applied to the EEG to get delta, theta, alpha, and beta waves. Subsequently, the spatial and temporal features are extracted using four frequency divisions via an Attention-based Conditional Autoencoder (A-CAE). In addition, the spectral features and statistical features are extracted. Significant patterns associated with PD are effectively analyzed using the extracted features. The optimal weighted feature fusion process is performed via the Improved Random Uniform Number-based Rock Hyraxes Swarm Optimization (IRUN-RHSO) by tuning the weights optimally. The weighted features are used for the PD detection process via the proposed Adaptive Dilated Convolution Capsule Network (ADCCN). Thus, the presented model automatically learns from the signals and provides detected outcomes for better progression of Parkinson’s disease.