Classification of Irregular Heartbeat Patterns with Multilayer Perceptron Application on Ecg Data

Cardiovascular diseases are the leading cause of death globally, thus, giving rise to the need for effective methods for early detection and diagnosis with emphasis on unsupervised learning. However, since feature extraction from electrocardiography (ECG) relies heavily on medical expert knowledge, there have been few attempts to optimize an algorithm that can be deployed in heartbeat monitoring systems. The paper introduces an application of a multilayer perceptron (MLP) artificial neural network algorithm in the detection of irregular heartbeat patterns found in ECG data. The model was trained on the MIT-BIH Arrhythmia dataset with a 60-20-20, train-validation-test ratio. Common features found in more than one dataset were selected for the cross validation, and irregular heartbeat samples were merged into one class. A grid search was performed to identify hyperparameter values which yielded higher accuracies which favored two hidden layer systems and the hyperbolic tangent activation function. The highest testing accuracy recorded was 98.27% but was unable to predict the unknown heartbeat class due to its small sample size. Additionally, these results suggest that ECG datasets tend to overfit which requires more research to optimize an algorithm that can accommodate biases in newly recorded ECG data. More experimentation can also be done without the restriction of common features across datasets.