Arrhythmia Detection Using Spiking Variable Projection Neural Networks

Edge-based Machine Learning (ML) has a pivotal role in revolutionizing smart healthcare by introducing a tangible improvement in the secure and discrete medical data analysis. This paper presents a novel neural network architecture by combining Variable Projections (VP) and Spiking Neural Networks (SNN). VPs are nonlinearly parameterized orthogonal projections whose weights have physical meaning, whereas SNNs are biologically plausible neural networks that operate on both spatial and temporal information. In the proposed hybrid topology, VP layer is coupled with spiking layers to encode input space into a compact and interpretable latent feature space. The latent space, encoded into spike trails, enables the subsequent SNNs to be trained with a low bandwidth. The effectiveness of the proposed VPSNN architecture is assessed in binary classification of normal and ventricular ectopic beats (VEBs) in ECG recordings of the PhysioNet MIT-BIH Arrhythmia Database. In our experiments, ECG records are apportioned into balanced training and test sets with approximately 60/40 ratio. Results show that a VPSNN variant detects VEBs with an overall classification accuracy of 97.16%, with a highly shallow topology consisting of only 242 parameters. The compact topology of VPSNN, makes it a suitable candidate for neuromorphic computing.