DYNAMIC ANALYSIS OF EEG TIME SERIES DISTINGUISHES PATIENTS WITH PARKINSON’S DISEASE FROM HEALTHY INDIVIDUALS USING MODEL-LEVEL GRAPH TRIPLET NEURAL NETWORK OPTIMIZED WITH CRAYFISH OPTIMIZATION ALGORITHM

Abstract

Parkinson's Disease (PD), a degenerative neurologic disorder, has a marked impact on motor and cognitive skills. Early and correct diagnosis is key to successful treatment and disease management. In light of limitations inherent in existing EEG-based PD classification techniques like insufficient feature representation, poor classification efficiency, and inadequate generalization a new dynamic analysis of EEG time series method using a Model-Level Graph Triplet Neural Network optimized with Crayfish Optimization Algorithm (MLGTNN-COA) is introduced. The approach starts with EEG signals from San Diego EEG dataset are first preprocessed with a Deformable Kernel Network (DKN) to eliminate artifacts and cluster the signals into 10 brain regions to perform localized analysis. A Fast hybrid Vision Transformer (FhVT) is then employed to extract discriminative spatial-temporal features from preprocessed EEG signals. For classification, the architecture uses a Model-Level Graph Triplet Neural Network (MLGTNN), which integrates a Model-Level Graph Network for learning inter-region dependencies with a Triplet Attention Network (TAN) for improvement on attention on salient features. The architecture works well in classifying EEG data into three binary PD diagnostic tasks: healthy vs. PD, PD-on medication vs. PD-off medication and healthy vs. PD-off medication. For further improvement in performance, weight and loss values are optimized by applying Crayfish Optimization Algorithm (COA). The developed model has better classification metrics: accuracy (99.93%), recall (99.74%), precision (99.83%), specificity (99.59%), F1-score (99.67%), and high area under the ROC curve, indicating robustness and diagnostic accuracy. The complete pipeline shows great promise for clinical application in diagnosing Parkinson's conditions based on non-invasive EEG recordings.