Pre-operative risk-stratification algorithms for cardiac surgery patients

Accurate pre-operative prediction of patient risk for adverse outcomes following surgery is important to ensure that patients are given the best possible treatment option. We develop hospital-specific pre-operative risk-stratification models for patients receiving isolated aortic valve replacements (the standard of care for treatment of severe aortic stenosis). However, because we have insufficient data for this task, we investigate approaches to best make use of related data (source data) that do not directly pertain to the task of interest (target data). These related data are taken from other surgeries and other hospitals, and therefore exhibit differences from the the target data. Naive pooling of all of the data into training improves performance, but can sometimes result in negative transfer. We hypothesize that appropriately weighting instances based on how related they are to the target data can result in improved performance.


Predicting adverse events after acute coronary syndrome


Acute coronary syndrome (ACS) affects millions of Americans every year. Patients diagnosed with ACS are at increased risk for future adverse cardiovascular events such as fatal arrhythmias or congestive heart failure. The ability to risk stratify patients with ACS would enable treatments appropriate for the level of risk. Our goal is to use electrocardiographic (ECG) data, which are inexpensive and routinely collected, to identify patients at higher risk for future adverse cardiovascular events. To do this, we apply signal processing techniques, machine learning, and other algorithms based on the (patho)physiology of the heart.


Video classification based on minute details


What distinguishes a made free throw from a missed one? What about a birdie golf putt from a near-miss? This project seeks to provide individually-tailored answers to such questions by identifying subtle differences in videos of nearly-identical actions (e.g., putting successfully vs unsuccessfully).


Leveraging hierarchy in medical codes for predictive modeling


Medical diagnoses are inherently hierarchical--if one has "acute systolic heart failure", he or she also has "systolic heart failure" and, most generally, "heart failure." If one ignores this hierarchy, one could falsely conclude that a patient with "acute systolic heart failure" is dissimilar to one with "chronic systolic heart failure", since these are different diagnoses. However, if one simply says that both have "heart failure", important subtleties may be lost. This project takes an intelligent approach to leveraging hierarchy to improve disease prediction.


Early prediction of baseball game outcomes


Predicting the outcomes of baseball games before they finish would allow adaptive strategy adjustments by teams and richer analysis by commentators. Unfortunately, such prediction is difficult, since the assumptions of uncorrelated data common in machine learning do not hold--the situation in the second inning is strongly related to the situation in the first inning, for example. The aim of this project is to develop techniques robust to violations of this independence assumption, with baseball targeted as an application of particular interest.


Finding predictive features in soccer matches


Soccer (football) is the most popular sport on Earth, but our understanding of how matches turn out is far from data-driven. This goal of this project is to identify predictive features of soccer matches--such as patterns of passes or shots--that could be used to predict match outcomes. This would not only offer actionable insights for teams throughout the world, but also allow deeper analysis by commentators and broadcasters.


Visualizing motion


Motion magnification is a potent tool in medical monitoring, but using it currently requires expertise in computer vision and programming. This project seeks to eliminate this barrier by building a cross-platform visualization program that can be used by medical practitioners with little to no training.


Learning examples from weakly-labeled time series


Thanks to devices such as the Fitbit and Apple Watch, temporal records of human vital signs and motion are becoming increasingly common. However, learning from this data can be challenging, since events of interest, such as anomalous heartbeats or health behaviors, are hidden within constant streams of data. The objective of this project is to learn what such events look like with minimal human intervention and recognize them in the real world.