Structure aware graph community cluster pruning for efficient neural network compression in Parkinson's disease diagnosis.

Deep neural networks in medical and edge environments often face computational and memory constraints, which necessitate effective model compression. Neural network pruning is a widely used solution, but conventional methods often rely on magnitude-based or random criteria that can remove entire functional groups of neurons and reduce model performance. This study introduces Graph-Community Cluster Pruning (GCCP), a structured pruning framework that constructs neuron activation similarity graphs and employs Louvain community detection to identify functionally cohesive groups. Within each community, neurons are ranked using gradient-weighted saliency, enabling the removal of redundancies while retaining representative units essential for preserving functional diversity. We applied proposed GCCP to Parkinson's disease detection using the publicly available UCI benchmark voice dataset and achieved a 37.23% reduction in parameters and a 37.84% reduction in Floating Point Operations (FLOPs), while maintaining test accuracy (94.87%) and AUC (0.969). Comparative evaluations demonstrate that GCCP consistently outperforms conventional pruning methods, particularly under high compression ratios, and provides interpretable insights into neural specialization through community visualization. These results show that GCCP is a robust and interpretable compression method for developing resource-efficient and clinically reliable diagnostic models.