Serverless data parallelization for training and retraining of deep learning architecture in patient-specific arrhythmia detection

Stacked Denoising Autoencoders (SDA) are deep networks which have superior generative properties and therefore can be trained and retrained to learn the structure of a patient's heart beat signal with minimal training data. This approach is particularly useful in continuous remote devices because they gather large amounts of data for longer periods of time. Serverless applications are the desired way of building applications due to its cost effectiveness after advancements in commercially available serverless host providers like Amazon AWS. This work proposes a serverless architecture for the training and retraining of SDA, for classification of arrhythmias in a patient-specific manner. This work also proposes a technique for data parallelization in the serverless architecture to achieve a speedup of up-To 13x in training time. This work uses MIT-BIH Arrhythmia database. Retraining with this architecture shows high classification accuracies for Ventricular Ectopic Beats (VEB) (97.41%) and Supraventricular Ectopic Beats (SVEB) (98.77%).