Interpretabilidade de modelos de aprendizado de máquina para classificação de arritmias cardíacas

Abstract

Cardiovascular diseases are a public health problem throughout the world and are the leading cause of death in many countries. Among cardiovascular diseases, cardiac arrhythmias are the most common and, for this reason, their precise classification has been of great interest in biomedical studies. The analysis of ECG is one of the most effective tools for detecting and classifying heart disease, and the use of machine learning models can help in this process. However, the interpretability of these models is a challenge, since they can be considered “black boxes”, i.e. it is not possible to understand how they reach certain conclusions. In this context, this work addresses the issue of the explainability of machine learning models for classifying cardiac arrhythmias in ECGs. To this end, it proposes an approach for building interpretable models, which allow the identification of the main characteristics of ECGs that influence the detection of cardiac arrhythmias, as well as carrying out explanations at model, class and ECG signal level. The explainability of the models is achieved by generating visual explanations that allow doctors and other health professionals to understand how the model reaches certain conclusions. The results show that the variability of RR intervals, the median of RR and PR intervals and the signal-to-noise ratio are crucial for classifying cardiac arrhythmias.