Modelo para detecção de arritmias cardíacas baseado em coeficientes de predição linear e rede neural convolucional 1D

Abstract

In recent decades, health problems resulting from heart disease have become the biggest causes of death in the world. When considering the mortality rate and the large number of people suffering from heart disease, the need for an early diagnosis becomes quite relevant. In order to contribute to the diagnosis, it is proposed the development of a computational system of detection and classification of heartbeats for some types of arrhythmias. To achieve this objective, the proposal was developed in two stages. First, an algorithm was developed and implemented to estimate heart rate from ECG signals using the Linear Prediction Coefficient (LPC) algorithm. At this stage, accuracy above 98% was obtained in estimating heart rate. In the second moment, a CNN network architecture capable of learning the dynamic characteristics of the heartbeats generated by the LPC algorithm and the morphological characteristics of the ECG signal was chosen. To train the network, ECG signals obtained from the database of the Massachusetts Institute of Technology (MIT-BIH Database of Arrhythmias) with 2480 signals and from PhysioNet/Computing in Cardiology Challenge 2017 with 3720 signals were used. To expand the data collection, a data augmentation technique was applied. After the training phase, the results were 93.35% for accuracy, 93.12% and 93.08% for accuracy and recall, respectively, for the MIT-BIH data. And for the Physionet data, the model reached 99.87% for both accuracy, precision and recall, respectively. The results of the classification model when combining the information obtained by combining the LPC and the Unidimensional Convolutional Neural Network are quite expressive when compared with other techniques published in the scientific community.