SNN with Gradient-based Backpropagation algorithm for ECG arrhythmia classification with LIF neuron and AdEx neuron

Signal processing is used by many to diagnose cases of Arrhythmia in a person’s heart by analyzing the Electrocardiogram (ECG) of the patient. Recently, ML algorithms and neural networks have been adopted to process the time-varying ECG signal and diagnose medical conditions. As part of this, Spiking neural networks (SNN) have gotten significant attention for their potential in analyzing time-varying signals like ECG. This project explores the SNN’s ability to process data with a few neuron models that can mimic the behavior of the biological neuron. This project focuses on finding the efficacy and complexity of the SNN’s that are modeled for the classification of the ECG into cases of Arrhythmia with the parameters that are extracted from the ECG signal. In this work, two 3-layered fully connected SNN models (3 x n x 5 architecture), where n is the number of neurons in the hidden layer, are created for 5 classes of arrhythmia cases (N, V, S, A, R) according to AAMI standard with 2 different neuron models namely LIF and AdEx, and input features extracted from the ECG data from the MIT-BIH database. The model was trained using the backpropagation algorithm to compute the loss gradient for updating the synapse weights. The model is designed to maximize accuracy and specificity with less complexity. An accuracy of 85.94% is achieved with AdEx neurons with 10 neurons in the hidden layer (n = 10) and 81.65% is achieved with the LIF neurons with 15 neurons in the hidden layer (n = 15).