Jakob Torgander: Straight to the Heart: Classification of Multi-Channel ECG-signals using Residual Neural Networks

Abstract.

The Residual Neural Network (ResNet) model represents a significant milestone in Deep Learning (DL) research and has since its introduction achieved state-of-theart results on a range of different data classification problems. In this thesis, we investigate the ability of the ResNet-model to successfully classify multi-channel electrocardiogram (ECG) signals. More specifically, we will examine whether a trained ResNet-model can accurately classify the correct positioning of an electrode array used to record an ECG-signal, when given the recorded signal as input. Since this electrode array in turn is placed on the standard feeding tube, which is provided to the majority of patients who are undergoing intensive care treatment, a model that accurately can classify the position of the array could hence be used to support medical practitioners during the insertion and positioning of the feeding tube.

 In the first part of the thesis, we will describe the process for fitting a ResNetmodel to our given data set, which consists of 8-channel ECG-recordings, originating from a sample of 27 patients who are all undergoing intensive care treatment. Evaluating our trained ResNet-model on unseen patients, we receive a F1-score close to 90% and a ROC-AUC score over 95% – significantly exceeding the performance from a baseline logistic regression model. In the second part, we investigate how Shape-aware Stochastic Neighbor Embeddings can be used to visualize the intermediate outputs of our trained ResNetmodel in 2D-space, such that global structures of the original high-dimensional feature space are preserved. Doing this, we show that there exists an intermediate feature space of our model, which we will call the Critical layer of our network, where the target classes become sufficiently separated. We furthermore show that, at the Critical layer, the output data points are ordered according to their predicted class probabilities along a line in the projected 2D-space. We will illustrate that by inspecting the layer outputs along this line, we can receive a better understanding of the decision process of our trained ResNet-model. Keywords: deep learning, machine learning, residual neural networks, dimensionality reduction, ECG, signal processing.