Classificação de arritmia cardíaca com aprendizado de máquina automatizado (AutoML)

Abstract

Cardiovascular diseases are the biggest killers in the world, with cardiac arrhythmia being one of the biggest causes. The test capable of detecting cardiac arrhythmia is the electrocardiogram, a test that contains noise and is difficult to interpret. In view of this, in order to contribute to research that facilitates the process of interpreting exams, Artificial Intelligence has a vast amount of work related to this objective, where researchers have developed strategies that make data more readable for algorithms and analyzed different model structures that better handle electrocardiogram data. Achieving the same objective, the present work implements and analyzes: segmentation sizes on Electrocardiogram (ECG) exam data; application of noise removal technique from the Neurokit library; Synthetic Oversampling Technique for Minority Classes (SMOTE) and Adaptive Synthetic Sampling Technique for Minority Classes (ADASYN); and uses automated machine learning to search for better hyperparameters that fit algorithms without manually establishing tests. Furthermore, the present work surveys the variations in hyperparameters generated for the models by the automated machine learning tool, AutoGluon. With the approaches implemented, this work concludes that for most of the algorithms used, the use of data with 360 segmentation allows the model to achieve greater performance when compared to data segmented in 2000. Machine learning models achieved better performances when classifying data that had the noise removal technique applied. Finally, Autogluon generated more variations for deep learning models, but despite this, the Light Gradient Boosting Machine (LightGBM) machine learning algorithm also achieved high performance in the testing phase. Therefore, the algorithms that achieved the best performance were neural networks and LightGBM.