Cardiac arrhythmia detection with deep learning architectures.

Time series classification (TSC) has an important role in medical diagnostics, providing decision support for vector-shaped data received from biomedical sensors. Traditional machine learning methods provide a sufficient baseline, while they need additional feature extraction procedures and result in lower accuracy compared to recent deep learning approaches. Providing more reliable results, deep learning architectures have become the golden standard for TSC tasks, evoking studies about which architecture provides better results with faster implementation. This study aims to provide a comparison between well-known deep learning architectures CNN and LSTM in comparison with traditional ANN, applying these classifiers for the MIT/BIH Arrhythmia Database, an Electrocardiogram (ECG) dataset that is publicly available. Results show that the best accuracy is achieved for CNN architecture used (96.17%), while LSTM resulted in comparable accuracy (94.42%) and traditional ANN (88.98%) could not compete with the more recent and complicated architectures. These outcomes indicate that although vector-shaped signals have relatively lower complexity compared to two or more-dimensional data like images, more complicated deep learning architectures outperform the traditional neural networks indicating exploration of high order patterns in one dimensional data improves classification accuracy. [ABSTRACT FROM AUTHOR]