Abstract:
Myocardial infarction (MI) is one of the leading causes of death around the world. MI can be diagnosed from the Electrocardiography (ECG) of the patient. The ECG is crucial for the patients survival in the early hours of MI as diagnosis from elevated serum cardiac enzymes takes 5-7 hours. This thesis presents a Convolutional Neural Network (CNN) architecture which takes raw Electrocardiography (ECG) signal from three ECG leads (lead II, III, AVF) and differentiates between inferior myocardial infarction (IMI) and healthy signals. A 5 layer deep network architecture consisting of inception blocks is developed. The discriminating strength of the features extracted by the convolutional layers by means of geometric separability index and Euclidean distance is analyzed and compared with the benchmark model. The performance of the CNN is also evaluated in terms of accuracy, sensitivity, and specificity and compared with the benchmark. The proposed model achieves superior metrics scores when compared to the benchmark model using hand engineered features. The detection models in the existing literature focused on ST segment elevation. But, studies showed that there is significant information in the leads containing ST segment depression. Hence, the IMI detection capability of different combinations of these leads needs to be investigated which is a computationally challenging task. We analyze the discriminating strength of the features extracted by the convolutional layers by means of geometric separability index and Euclidean distance and compare it with the benchmark model. Additionally, a comparison of the predictive capability (in terms of accuracy, sensitivity, and specificity) of different lead combinations is carried out. Experiments show that the combinations of leads that capture ST segment elevation and depression often outperforms the combinations of leads that capture ST segment elevation only.
