Determination of heart diseases from ECG signals using dense neural networks

Introduction:

Thanks to the machine learning and deep learning libraries developed in the past couple of decades, early detection of the cardiac abnormalities and the development of automation tools for the diagnosis has seen significant growth in the past 10 years. Despite these results conventional machine learning approaches require lots of time and effort for feature selection [1]. It is necessary to select the best features as input for the supervised classification algorithms which is found by a lot of trials, and might be time consuming. Advantage of deep learning over traditional machine learning is, that it doesn’t need feature extraction and feature selection. The decision-making algorithm can consider all available evidence. Even though the theory and the applications of deep learning are still evolving, several published [3-6] studies have shown that it possesses the capability of faster and more reliable diagnoses in physiological signals. In some case, the deep learning architectures have shown their usefulness by surpassing the performance of traditional supervised classification machine learning techniques [5]. This predicts a trend, where we might shift away from the currently used decision support methods, such as (SVM) and K-Nearest Neighbour (K-NN), towards deep learning methods [4]. For instance, Acharya et al. [4] reported an accuracy of 98%. Furthermore, their system outperformed traditional approaches and it has the added benefit of not having to perform feature extraction and de-noising of the signals.

Currently application of deep learning to physiological signals uses Convolutional Neural Networks and Autoencoders [1,4]. The objective of this work is to use deep learning methods (CNN) applied to ECG signals to detect heart disorders in the measured subjects. The labeled dataset for training the CNN was obtained using Electro-Cardiogram (ECG) measurements, which are typically performed by placing electrodes on the chest [1]. This signal reflects the functioning of the heart and it has well distinguishable features. However, the difficulty in ECG interpretation is to spot the morphological changes which indicate a particular cardiac problem or diabetes [2]. These abnormalities may be minute and very often they may be transients or present all the time. A supervised learning based approach may be useful to automatically detect such abnormalities in the ECG. The report is organized as follows. The information about training dataset, the preprocessing techniques, and the architecture of the deep neural network used in this work are described in Section II. The results of the training dataset and of optimization of some of the hyper-parameters are discussed in Section II. Finally, Section IV concludes the report.

Models and Methods

We obtained the dataset of ECG signals along with the labels from the datasets provided in the Machine Learning class of 2017 (Kaggle, ML-2017 [12]). The signal is sampled at 150Hz. The dataset consists of 6822 ECG signals out of which 4040 signals are obtained from healthy individuals and 2192 ECG signals are taken from the individuals with a heart condition (Atrial flutter or Atrial fibrilation). The rest of the signals (590) correspond to ventricular diseases which we ignore in this work. In this way, the dataset includes 4040 healthy signals (label 0) and 2192 signals with heart condition (label 1). We divided our dataset in 3 parts, the 80% of the signals were used for the training, and 20% of the signals were used for testing. Note, that we perform this split before training. That is, the labels of the testing dataset were exclusively used for final test score, not for training. For the having a better accuracy of the training set, we performed cross-validation in this data partition [13]. From the training data, 90% was used for training and 10% was used for validation.