Hyper-Efficient Neural Network Pruning via Adaptive Sensitivity-Based Gradient Projection for Scalable Biomedical Signal Processing

**Abstract:** Biomedical signal processing (BSP) datasets are increasingly voluminous and exhibit high dimensionality, posing significant computational challenges for real-time analysis and deployment in resource-constrained environments. Traditional deep learning models, while achieving high accuracy, often suffer from excessive parameter count and computational complexity. This paper introduces a novel approach, Adaptive Sensitivity-Based Gradient Projection (ASGP), for neural network pruning within the context of BSP, specifically targeting electrocardiogram (ECG) arrhythmia classification. ASGP dynamically adjusts pruning intensity based on the sensitivity of each neuron's gradient to the loss function during training, enabling selective removal of redundant parameters without significant loss of accuracy. This results in a compressed model with improved computational efficiency, suitable for edge-device deployment in real-time healthcare applications. We demonstrate its effectiveness through rigorous experiments on publicly available ECG datasets attaining a 10x reduction in parameters while maintaining competitive classification accuracy.