Quaternion Neural Network with Temporal Feedback Calculation: Application to Cardiac Vector Velocity during Myocardial Infarction
AbstractQuaternion neural networks have been shown to be useful in image and signal processing applications. Herein, we propose a novel architecture of a neural unit model characterized by its ability of encoding 3-dimensional past information and that facilitates the learning of velocity patterns. We evaluate the implementation of the network in a study of the cardiac vector velocity and its usefulness in early detection of patients with anterior myocardial infarction. Experimental results show an improvement of the performance in terms of convergence speed and precision when comparing with traditional methods. Furthermore, the network shows successful results in measuring the velocity reduction that is usually observed in vectorcardiogram signals in the presence of myocardial damage. Through a linear discriminant analysis, a pair of 100% / 98% of sensitivity / specificity is met with only two velocity parameters. We conclude that this method is a very promising development for future computational tools devoted to early diagnosis of heart diseases.