A Comparative Framework Integrating Hybrid Convolutional and Unified Graph Neural Networks for Accurate Parkinson’s Disease Classification

Md Abubakkar

Abstract:

Parkinson’s disease (PD) is a progressive neurode-
generative disorder, affects motor function and is often chal-
lenging to diagnose due to the complex interplay of clinical
features. This study integrates a comparative framework in-
tegrating hybrid Convolutional Neural Networks (PCNN) and
graph-based models (GCN, GAT) to enhance Parkinson’s disease
(PD) diagnosis using structured medical data. PD, a progressive
neurodegenerative disorder affecting motor function, poses di-
agnostic challenges due to complex clinical feature interactions.
The PCNN employs 1D convolutions to capture local feature
patterns, while GCN and GAT model intricate interdependencies
between clinical variables by representing the dataset as a graph.
Notably, GAT’s attention mechanism dynamically prioritizes
important features, improving interpretability and diagnostic
precision. Through hyperparameter optimization with Optuna
and addressing class imbalance using SMOTE, our approach
achieved a peak accuracy of 97.44%, surpassing traditional
methods. The comparative analysis reveals that while PCNN
excels in classification accuracy, GAT’s attention-based feature
selection offers superior interpretability. This makes it a valuable
tool for more precise Parkinson’s disease detection in clinical
applications. The integration of these models provides a com-
prehensive framework for PD diagnosis, leveraging both local
and global feature extraction techniques. This study represents a
significant advancement in applying advanced machine learning
to neurodegenerative disease diagnostics, offering improved early
detection and personalized treatment potential for Parkinson’s
disease.