همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

NeuroPC-KGNN: A Causality-Driven Graph Neural Network Framework for Early Diagnosis of Parkinson’s Disease

نویسندگان :

Fatemeh Ahouz¹ Ahmad Ali Abin² Amin Golabpour³

1- دانشگاه شهید بهشتی 2- دانشگاه شهید بهشتی 3- دانشگاه علوم پزشکی شاهرود

کلمات کلیدی :

causal feature selection،PC algorithm،graph neural networks،Parkinson’s disease،voice analysis

چکیده :

Abstract—Early diagnosis of Parkinson’s disease is essential for effective intervention, and voice signals provide a simple non-invasive tool for this purpose. However, many acoustic features are correlated rather than causal, limiting the robustness and generalizability of conventional machine-learning models. This study introduces NeuroPC-KGNN, a framework that integrates the PC causal discovery algorithm with a Graph Convolutional Network (GCN) operating on a graph constructed from samples connected to their k-nearest neighbors. The PC algorithm identifies features that exhibit direct or indirect causal relationships with disease status, ensuring that non-causal and noisy variables are excluded. A GCN trained on a kNN-constructed graph then models the structural dependencies among samples. Using the UCI Parkinson dataset, NeuroPC-KGNN achieved 90% accuracy and 91% sensitivity on the test set. Permutation-based sensitivity analysis revealed that spread1, PPE, and MDVP:Flo(Hz) were the most influential features, showing the greatest impact on model performance when perturbed. The strong alignment between the causal graph and feature-importance results highlights the reliability and interpretability of the proposed framework. Our results demonstrate that NeuroPC-KGCN provides a robust, interpretable, and clinically meaningful approach for early PD diagnosis from voice data.