Edmund Do

I’m a Master’s student in Computer Science at the University of Vermont. My advisor is Prof. Byung Lee. My research interests are broadly in network science, data mining, and machine learning. I am particularly interested in applications to social networks.

Currently, I am working on research relating to organizational dynamics and job performance. Previously, I did some research on classifying 12-lead electrocardiograms with machine learning.

selected publications

submitted

12-lead imbalanced beat classification using time series resnet

Do, E., Boynton, J. W., Lee, B. S., and Lustgarten, D.

Abs

This paper reports the performance study of classifying 12-lead ECG beat segments in the face of severe imbalance in the class sizes, which is typical of ECG data. The efficacy of data augmentation for class size balancing to improve the classification accuracy is well known. In the ECG domain, however, it has not been overlooked or handled inadequately. We propose an amplitude-alteration approach to add ECG heartbeats selectively as needed in individual ECG classes while not disrupting the timeline of the ECG signals. In addition, augmentations of training dataset and test dataset receive separate attentions, and four cases of data augmentation are considered depending on whether each dataset is augmented or not. The effects of the augmentation scheme was evaluated using ResNet, the deep learning technique reputed for its remarkable accuracy through unique skipping connections between layers; specifically, a time series version of ResNet was used. The results confirmed the key benefit of class-balancing the training dataset through the proposed data augmentation scheme and, additionally, showed some extra benefit of augmenting the test dataset through “time-test augmentation”. Further, we adopted the class activation map (CAM) to identify heat map signatures that would explain ECG beat classes. The results demonstrated that the CAM could be an effective visual aid to classifying ECG beats, especially with the proposed data augmentation scheme in place. In this paper the augmentation scheme, the associated experiments, and their results are discussed concretely.