Data Augmentation: Repositioning of atria in the torso to produce new 12-lead ECG signals
Atrial flutter (AFl) is a common reentrant arrhythmia, characterized by a self sustainable mechanism and an electrical signal that propagates along different pathways from that of a normal heartbeat. Although AFl is not a direct cause of death, it can lead to even fatal complications, such as stroke or heart attack. For this reason, it is essential to identify and recognize this condition, so that it can be promptly treated. Nowadays, to know how to best treat this pathology, it is needed to discriminate with high precision which type of AFl the subject is affected. To do this, invasive methods of signal acquisition are required, such as intracardiac catheters. It would be best to try to identify with more precision the type and the location of the AFl, using non-invasive methods, such as a classifier that only uses feature extracted from 12-lead electrocardiogram (ECG), in order to decrease the procedure time of the ablation therapy by helping the doctors plan the intervention.
To best train a classifier it is needed a very large train set of signals that represent all cases in exam. In the biomedical field, obtaining all these clinical data is very complex, for this reason it is possible to use synthetic data, therefore using realistic models of atria and torsos to simulate the different AFl scenarios and extract the respective 12-lead ECGs.
In this work, the student will aim to test a method of data augmentation by rotating the models of atria in the torsos, remaining in physiological positions, to produce new simulations, verifying that the new data generated does not cause overfitting in the dataset.
To do so, the student will reposition the atria models in the torsos, solve the forward analysis and extract the 12-lead ECG. The student, through a correlation analysis, will identify which is the minimum range of repositioning of the atria to avoid overfitting.