Winning Posters Explore Personalized Medicine, Political Polarization and Atomic Imaging
The Foundations of Data Science Center kicked off the fall semester with a poster competition, drawing 16 students from Columbia Business School, Columbia Engineering, Columbia University Medical Center (CUMC), Columbia’s Graduate School of Arts and Sciences (GSAS) and Graduate School of Architecture, Planning & Preservation (GSAPP). Three posters were picked for special recognition by a panel of judges—Ali Mehmani, Ken Ross, Sharon Sputz, John Wright and Tian Zheng. The posters, and the judges comments, are summarized below.
First Place Meal-time Decision Support for Type 2 Diabetics
Many health apps rely on their users to make discoveries and take action. GlucOracle, an app developed by CUMC researchers Lena Mamykina, Dave Albers and George Hripcsak, puts users in control by computing personalized, nutrition-based glucose forecasts in real-time. The poster presented by research associate Matthew Levine demonstrates the mathematical framework and preliminary findings. The authors embed mechanistic models of the glucose-insulin system in an unscented Kalman filter, and demonstrate that the models can be trained on as little as a week’s worth of patient-collected self-monitoring data. With this data, they found that the proposed models were able to match or beat forecasts made by clinical experts, and could infer hidden glucose dynamics.
Judges: We like that the platform fulfills the need of providing individualized diet management for diabetic patients. It combines the technology of a mobile computational platform, real-time predictive capability of artificial intelligence, and patient-collected data from wearable gadgets. Both the poster and the 2-minute presentation are well explained and clear.
Second Place: A Model Tracking Partisan Divide in the U.S. Senate
As the country becomes more ideologically divided, party affiliation has grown more important in national politics. That’s the conclusion of a statistical model based on U.S. Senate voting records developed by researchers Gen Li and James Wilson at CUMC and University of San Francisco respectively. The poster presented by graduate student Jihui Lee presents a new statistical method developed by the team — a functional exponential random graph model (FERGM) that measures how dynamic networks change through time. An extension of an exponential random graph model, it’s able to capture trends through time.
Judges: We like how this poster explains a modification of statistical model for studying trends and changes in dynamic networks. Using the co-voting network among senate members, it shows that the new model can efficiently describe the narrowing and widening of partisan division among senators. The poster provides clear explanation of the literature background and contribution of the new method.
Third Place An Algorithm to Sharpen Views of Atomic Structure
Invented in the 1980s, the scanning tunneling microscope (STM) is still the best way to analyze the structure of individual atoms to gain insight into their real-world properties. Increasingly, optimization algorithms are being applied to STM images to extract critical information in the search for novel materials. A learning algorithm developed by researchers John Wright in Columbia Engineering and Abhay Pasupathy in GSAS uncovers hidden information in STM images. In their poster, graduate students Yuqian Zhang and Sky Cheung presented results showing atomic features in the superconductor material NaFeAs previously obscured by statistical noise. These features could lead to new insights into the material’s behavior. The method improves on previous Fourier-based techniques and has application for computer vision and neuroscience.
Judges (excluding Wright who recused himself on the vote): The poster addresses an interesting data-intensive situation (high resolution images) with sparse signals (defects). The proposed method is motivated directly by a mathematical translation of the signal convolution process and efficiently addresses the challenging non-convex optimization computation issue. The presentation focused on the difficult data problem and clearly detailed existing methods and the proposed innovation.