Using An Artificial Neural Network to Detect Activations During Ventricular Fibrillation

Abstract

Ventricular Fibrillation (VF) is a severe cardiac arrhythmia that canresult in sudden death, a leading cause of death in the United States. DuringVF, the electrical activity of the heart becomes disordered, the ventriclescontract erratically, and an insufficient supply of blood is pumped to the body.Identification of the electrical activation sites during VF is important for theunderstanding and improved treatment of the disorder. Unipolar electrogramsof four pigs were recorded following the induction of VF. The data from theVF recordings was preprocessed using a Rule-Based Method (RBM) , a Current Source Density (CSD) method, and a Transmembrane CurrentMethod (TCM) to separate local activations from distant activity. RBM usesthe magnitude of the derivative of the voltage to identify activations. CSD is ascalar quantity that represents the magnitude of the current source or sink.TCM estimates a value proportional to the transmembrane current. Afeedforward artificial neural network (ANN) using backpropagation was trained to identify the local activations in the electrograms of VF based on the RBM and CSD calculations. Another feedforward ANN usingbackpropagation was trained using data preprocessed with not only RBM andCSD, but also TCM. In order to improve the ability of the ANNs to detect localactivations, a new training method, called staged training, was utilized. Instaged training, the ANNs were trained in stages using different sets oftraining examples. Examples were included in a particular training set basedon the minimum magnitude of the voltage derivative. When training was done in stages, both the ANN with RBM and CSD data only and the ANN with RBM,CSD, and TCM data were able to more accurately distinguish activations.Overall, the ANN, which used only RBM and CSD data, produced better results than the ANN that also included TCM data.