O uso de aprendizado profundo para predição de diagnósticos de doenças cardiovasculares

Abstract

Cardiovascular diseases (CVD) are a group of diseases of the heart and blood vessels. These diseases are the leading cause of death worldwide. CVD is affecting more and more young adults, due to unhealthy habits such as not exercising, being overweight, obesity, smoking, stress, depression and eating a diet high in fat and processed foods. In this context, early detection of CVD plays an essential role so that management with counseling and medication can begin. Most heart conditions can be diagnosed using electrocardiogram (ECG) signals. The classification of the ECG is highly relevant today, as there are many medical applications in which this problem can be evidenced. In the second half of the 20th century, Machine Learning (ML) evolved as a sub-field of Artificial Intelligence (AI), capable of developing self-learning algorithms that form knowledge from data in order to make predictions. Although there are many AM methods that present solutions capable of helping with ECG interpretation, predicting CVD is one of the most complex impasses in the field of clinical data analysis. That said, carrying out analyses with methods that have not yet been explored for this purpose could make significant contributions to the health field. With this in mind, this work proposes the use of deep learning to classify CVD diagnoses, developing Transformer and LSTM neural network models and evaluating them using the metrics of accuracy, recall and f1-score, in order to compare the performance of the methods that have been used to predict cardiovascular disease diagnoses. The results of the experiments showed accuracy values of approximately 70% to 84% for the LSTM neural network and accuracy values of 48% or less for the Transformer neural network. For the recall metric, Transformer obtained values of up to 62% and LSTM obtained results of up to 77%. For the f1-score metric, LSTM had values between 41% and 72% and Transformer had lower values of up to 36%.