1) Team: Team Live and Let Liver (Kenny Groszman, Lucas Cadalzo, Lorenz Gahn)
Title: Predicting Treatment Response of Hepatocellular Carcinoma Patients
Hepatocellular Carcinoma (HCC) is a lethal type of cancer, with curative therapies unavailable to more than 80% of patients. Treatments for HCC are extremely painful and intensive, and currently no methods exist to predict whether the treatment will be effective for a particular patient. Patients' time to progression (TTP), a measure of time until the disease worsens, typically ranges from 30 weeks to 200 weeks for patients that respond well to treatment. In this project, we develop a model for predicting patient responses to transarterial chemoembolization (TACE), a common treatment for terminally ill HCC patients. First, multiple logistic regression is used to classify the patients as either responders or non-responders to TACE. Following classification, multiple linear regression predicts patients' TTP. This approach was shown to be considerably more accurate than comparable machine learning techniques. Using a dataset provided by Dr. David Fuentes at MD Anderson, we found that our approach has a classification accuracy of 92%, which allows oncologists to make more informed decisions about whether or not to administer TACE to a given patient. Additionally, responder TTP is predicted with a median error of 8 weeks. As a result, this work could significantly improve outcomes for patients with HCC.
2) Team: Team Priam (Ariana Morgan, Eric Gong, Wesley DeLoach)
Title: Predictive Model of Temporal Lobe Epilepsy Seizure Focus
Epilepsy affects approximately 1% of all Americans. Of these 3 million Americans, about one million are affected by drug-resistant epilepsy. Of these one million, roughly half are MRI-negative. In other words, half suffer from seizure-causing brain tissue damage that is not identifiable through standard brain imaging techniques. For patients of pharmacoresistant epilepsy, the difference between detection and nondetection is significant. Compared with their MRI-negative counterparts, MRI-positive patients are twice as likely to be seizure free post-surgery. The problem of locating diseased tissue in MRI-negative patients can be viewed as a classification problem, where each voxel of a given brain is classified as healthy or unhealthy. Advancements in algorithms and computational power make this problem well-suited for machine learning. This project looks at the steps necessary, from start to finish, to translate the original problem into a semi-supervised classification problem, as the data needed to formulate it as a supervised classification problem is not readily available. It explains the image pre-processing required for converting brain MRI scans into a format suitable for computational use. Ultimately, it seeks to find and train an effective semi-supervised machine learning algorithm that can classify brain tissue, while documenting what does and does not work along the way.
3) Team: Team Arcs (James Lee, Judy Zhang, Doria Du)
Title: Dynamics of Arches
State-of-the-art engineering and manufacturing practices often require the construction of materials that have competing design constraints. In particular, many components of aircrafts are required to be light to reduce fuel consumption during flight, but also robust enough to withstand high pressure and temperature gradients of the flight. Aircraft panels of hypersonic vehicles undergo extreme acoustic excitation and thermal loading and the accurate prediction of their response is crucial to their success.
A key challenge is creating a computationally efficient model that can accurately predict the extreme nonlinear response of the panels, particularly when they undergo a phenomenon called snap-through. Snap-through behavior is the complex physical phenomenon that occurs when transverse forces, coupled with stiffening and softening response of the material, cause a dynamic instability manifested as a transition from small to large amplitude oscillations, along with a reversal of curvature. Properly identifying snap-through behavior is important to predicting the long-term durability of aircraft components.
Current finite element method (FEM) models require prohibitively long simulation times to solve the nonlinear problem accurately, so there is considerable effort to designing simpler models that retain the accuracy of FEM models. In particular, reduced-order modeling (ROM) can be a way to describe the snap-through behavior of shallow curved panels more efficiently. Our goal is to utilize a specific ROM based on Fourier approximations to efficiently and accurately identify the parameter boundaries that differentiate snap-through behavior.
4) Team: Team Junior Varsity (Luke Hall, Hannah Park, Jeremy Vollen)
Title: Wi-Fly 2.0
Wi Fly 2.0 is a continuation of a previous CAAM Senior Design project. The product is an Android application that, given a user-drawn 2-D floor plan and router location, evaluates the signal strength at all points inside the floor plan. The app is currently functional, but certain limitations keep it from being put to use in a realistic sense. Our project's goal is to address these limitations, expanding the generality of the problems the application can solve. A specific goal of our project is to allow the user to specify the material of their walls. Currently, the walls are assumed to be made of concrete, meaning wave transmission through walls is not allowed. In order to work toward this goal, we started with simpler problems such as the Laplace boundary value problem in order to help us understand the more complicated inhomogeneous Helmholtz problem as well as the transmission problem.
5) Team: Team to the Heart (Andrew Weatherly, Shan Zhong)
Title: Creating a 2D Heart Model for a Full Body Blood Flow Simulation
Team To the Heart's project was to create a 2D heart model for the full body blood flow simulation being developed by TCH and Rice. The goal of our project is to obtain a two–dimensional (surface) model of elastic dynamics, of a layer of material which represents the wall of a heart chamber (ventricle or atrium), where the underlying assumption is that the thickness of the cardiac wall is small enough to simplify the dynamics across the layer, and only compute explicitly the dynamics along the defining surface. The expected outcome is to accelerate the generation of a numerical solutions by at least an order of magnitude.