An AutoML Pipeline for the Classification of Noisy and Clean Electrocardiogram Signals Using Bayesian Optimization

Electrocardiogram (ECG) recordings are affected by notorious source of noise which can limit the reliability of the clinical decisions. Several methods have been implemented to automatically identifying noisy signals. Recently Machine Learning (ML) has helped in the automatic classification of ECG signals. In particular, Automated Machine Learning (AutoML), allows to overcome the limits of more common ML approaches.In this study, we aimed to implement an AutoML pipeline able to classify ECG signals as noisy or clean.Several databases were considered which were annotated on a 2 s window basis.After pre-processing and balancing the dataset, AutoGluon was applied for signals classification.For the model with the best trade-off between accuracy and inference time, Bayeasian Hyperparameter Optimization (HPO) was then applied.The Neural Network model was the best performing model on both the training and test set.HPO increased the performance without increasing significantly the inference time.This study showed the efficiency of an AutoML pipeline for ECG signal classification encouraging its applicability in a clinical setting.