Automated detection of Myocardial Infarction (MI) using ECG signals with artificial intelligence

Koh, Zi Kang (2024) Automated detection of Myocardial Infarction (MI) using ECG signals with artificial intelligence. Final Year Project, UTAR.

Preview

PDF Download (10Mb) | Preview

Abstract

Myocardial Infarction (MI) is known as heart attack, it is one of the most life-threatening cardiovascular diseases. During infarction, the coronary artery which is responsible for the delivery of blood, oxygen and nutrients, is fully or partially blocked, and the heart muscle will die of ischemia. Percutaneous Coronary Intervention (PCI) is a nonsurgical technique to treat MI, the faster the patient receives PCI treatment, the higher the survival rate. The heart activity (pumping blood) is controlled by the electrical current generated by itself, therefore 12-lead electrocardiogram is an excellent tool to capture the activity of the heart, the pattern of a complete heart cycle is referred to as the PQRST cycle. MI will cause a morphological change to this pattern, therefore this can be used to diagnose the MI. In order to avoid the intra-/inter-observer effect caused by manual human interpretation, many researchers proposed machine-learning-based methods and then nowadays many deep-learning-based methods have emerged to perform automatic and end-to-end classification. Nevertheless, many studies that emphasized deep learning models did not care about the data split method during their experiment, this led to a misleadingly supreme performance due to information leakage problem. The models might be trained to memorize which subjects have MI heartbeats instead of learning the features related to the disease itself from the amplitude and time (in a sequential model). Thus, this research proposed three models: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) and 1-dimensional Convolutional Neural Network (1D-CNN) with the implementation of intra-patient and inter-patient data split techniques. In FYP 2, the architecture for each model had been improved compared to FYP 1 to augment the performance, then regularization and dropout techniques were applied to increase the generalization ability and finally, one transformer model had been developed to test its potential in processing ECG signal. From the perspective of the inter-patient method, the LSTM model obtained 90.53% accuracy, while the 1D-CNN and GRU models obtained 85.82% and 86.65% accuracy respectively. On the other hand, for all the intra-patient models, LSTM and GRU obtained a similar 95.4% accuracy while 1D-CNN obtained a 97.68% accuracy. The transformer model achieved 82.28% and 91.15% in intra-patient and interpatient analysis. Obviously, this has proven that the intra-patient models can produce a misleadingly high result. Another dataset is obtained from the open database on the Internet, but unfortunately, the testing result has shown that all of the models failed to generalize.

Actions (login required)