Synthesis of realistic human heart-beats with time-series forecasting techniques and arrhythmia classification using Wavelet Scattering Transform

Cardiovascular diseases are the first cause of death, globally. In such contexts an early diagnosis can be life-saving for most of the cases considered. The real problem is that, for a lot of arrhythmia types, there are not many data available, even in the most notorious dataset used in this matter. For these reason it is fundamentally important to create new beats that are similar to the available, but not too much in order to preserv the privacy of the patient. This is the task tackled in this work: use time-series forecasting techniques in order to create new heart-beats, simulating especially types of beats with some sort of cardiac anomaly. In this work four classes of beats were considered: the normal class did not need any synthesis process, which was reserved for the other three. Four kind of time-series forecasting techniques were used: exponential time smoothing (ETS), ARIMA and two neural forecaster, a Recurrent Neural Network (RNN) applied to time-series and a Long Short-Term Memory (LSTM), which works on the same principles but it appears to be a little optimized. After a successful augmentation process, data was transformed using the Wavelet Scattering Transform, an evolution of Fourier transform which has a lot of very good mathematical properties, such as invariance to scaling and to translations. Transformed data was then classified using k-nearest neighbors, which allowed to obtain results comparable to the state-of-the-art in this kind of tasks.