Category: CMS-Mathematics & Statistics

Thomas Dunn Undergraduate Researcher in Mathematics

Thomas Dunn with research poster.Thomas Dunn is a sophomore at the University of Central Oklahoma majoring in mathematics. He started working on research with Dr. Tyler Cook and Dr. Emily Hendryx in the spring semester of his freshman year. Interested in applications of deep learning, Thomas has been rapidly learning the Python coding language to put these methods into practice. During his sophomore year, Thomas has been particularly focused on using autoencoders to differentiate between normal and anomalous electrocardiogram (ECG) beats. ECG data can provide a wealth of information to physicians about the health of a patient. However, even experienced clinicians may struggle to distinguish normal from anomalous ECGs in cases when the differences are subtle or distributed over long periods of time. Thomas’s autoencoder model takes steps toward developing an initial screening tool for ECG beats. An autoencoder (AE) is a machine learning model which learns to compress input data into a low dimensional vector representation and then reconstruct an approximation. The goal for such a model is to learn to reconstruct the input data as well as possible, which requires learning an effective representation for the data. An AE trained to reconstruct one class of data (normal ECG beats) will have a high error when trying to reconstruct data from another class (anomalous beats). His model uses reconstruction error to discriminate between different beat types which can then be assessed by a clinician or further classified using additional models. He investigates whether AE-based anomaly detection methods are viable tools in application to real ECG data, comparing the performance of the autoencoder against traditional classification methods. Thomas presented his results at this year’s Oklahoma Research Day at Southwestern Oklahoma State University. He also hopes to continue his research on deep learning by exploring the use of generative adversarial networks for 3D image generation.