A Novel Algorithm for Early Detection of Junctional Ectopic Tachycardia in Patients with Congenital Heart Disease
Texas Children’s Hospital, Medical Informatics Corporation, & Rice’s Data2Knowledge Lab
September – December 2019
Paper ⋅ Presentation
Arrhythmias can be lethal for children in the period following cardiac surgery, and junctional ectopic tachycardia (JET) is considered the most common type of tachycardia seen during early post-operative care. We present a novel classification algorithm that detects the JET onset based on electrocardiogram (ECG) waveforms. Our algorithm obtains an average cross-validation sensitivity of 87.07%, specificity of 87.12% and area under the receiver operating characteristic curve (AUROC) of 90.71% on a dataset of 9 patients from Texas Children’s Hospital. In addition, we present a “human in loop” development pipeline by creating a novel waveform visualization system. This pipeline will enable cardiac surgeons to better recognize and label arrhythmia onset.
Recurrent R2U-Net for Video Segmentation
Lawrence Livermore National Laboratory – Data Science Institute
May – August 2019
Poster ⋅ Website
Our client has classified videos that they would like to segment to look for specific objects and events. We embed a R2U-Net into a Convolutional LSTM to make it recurrent across time, thus making it suitable for event detection. As the target dataset was classified, we built our own dataset to test on, the Moving Colorful MNIST dataset. This dataset consists of videos of two colored digits moving across a colored square. Our model achieved over 99% accuracy on the dataset and was passed on to the team with appropriate clearances to apply to the classified videos.
Unsupervised Creation of Robust Environment Representations for Reinforcement Learning
Perception, Control and Cognition Lab at Brigham Young University
January 2018 – May 2019
Paper
Reinforcement learning (RL) methods traditionally have suffered from large sample complexities, which reduces their widespread effectiveness. This is particularly true for the beginning stages of deep RL methods, where the methods must learn what the objects in an environment are in addition to what actions to take. We aim to improve the sample complexity of deep RL methods by disentangling these two tasks, using unsupervised learning methods to complete the first task. In particular, we propose a method that embeds learned environmental representations into a Deep Q-Network (DQN), so that the DQN does not have to learn the environment itself. Specifically, we provide a Rainbow DQN with robust environment representations of Atari video games created through a conditional variational ladder autoencoder with coordinate convolutions.
Using Markov Chains in Facility Location Management
Mathematics Department at Brigham Young University
April – December 2017
Paper
We consider using Markov chains to solve facility location problems. We find that Markov chains, coupled with realistic assumptions, can find optimal facility locations defined in terms of four metrics. Namely, the mean first passage time, consumer throughput, top-tier, and the Kemeny constant. As a case study, we present a novel Markov chain construction that finds the optimal location for a new religious center of a Christian denomination.
Brigham Young University Speeches Popularity Predictor
Brigham Young University
August – December 2017
Paper
Natural language processing is currently an active area of research in machine learning. A particular subgroup of natural language processing strives to predict the “popularity” of a given text. Brigham Young University (BYU) maintains an archive of speeches that are given weekly on campus. We extracted a variety of features from the speeches and ran several machine learning models to predict which talks were popular (according to page views). All the models performed better than the baseline, but the J48 decision tree was the most effective. We did feature reduction using a wrapper algorithm, but this didn’t improve the overall performance of our models. We discovered which features were most helpful by examining the decision tree’s output.
Analysis of the Rigid Motion of a Developable Conical Mechanism
Mathematics Department at Brigham Young University
June – December 2016
Paper ⋅ Animation
We demonstrate analytically that it is possible to construct a developable mechanism on a cone that has rigid motion. We solve for the paths of rigid motion and analyze the properties of this motion. In particular, we provide an analytical method for predicting the behavior of the mechanism with respect to the conical surface. Moreover, we observe that the conical developable mechanisms specified in this paper have motion paths that necessarily contain bifurcation points which lead to an unbounded array of motion paths in the parameterization plane.
The nationwide landscape of K-12 school websites in the United States: Systems, services, intended audiences, and adoption patterns
Instructional Psychology Department at Brigham Young University
September – December 2016
Paper
This study sought to collect URLs (web addresses) of all K-12 schools in the United States (N = 98,477) and analyze website home page system and service data for all available U.S. institutional websites (n = 65,899). Building upon previous research related to Web 2.0 educational potentials, this first-of-its-kind study sought (a) to provide descriptive results of system and service adoption and website data for all schools in the United States and (b) to detect theorized differences based upon school demographics and service/system type (e.g., open source vs. proprietary). Results indicated that proprietary and purchased systems were much more common than free and open systems, that adoption patterns were generally not meaningfully influenced by demographic data (except for charter school status), and that K-12 institutional adoption of Web 2.0 seems to be more focused on educational uses of these tools that might not strictly be considered pedagogical (e.g., community outreach).
Religious Identity, Expression, and Civility in Social Media: Results of Data Mining Latter‐Day Saint Twitter Accounts
Instructional Psychology Department at Brigham Young University
January – June 2016
Paper
This study explores religious self‐identification, religious expression, and civility among projected Latter‐Day Saint Twitter accounts (201,107 accounts and 1,542,229 tweets). Novel methods of data collection and analysis were utilized to test hypotheses related to religious identity and civility against social media data at a large scale. Results indicated that (1) projected LDS Twitter accounts tended to represent authentic (rather than anonymous or pseudonymous) identities; (2) local minority versus majority status did not influence users’ willingness to religiously self‐identify; (3) isolation stigma did not occur when users religiously self‐identified; (4) participants exhibited much lower degrees of incivility than was anticipated from previous studies; and (5) religious self‐identification was connected to improved civility. Results should be of interest to scholars of religion for better understanding participation patterns and religious identity among Latter‐Day Saints and for exploring how these results may transfer to other groups of religious people.