Electrical Engineering and Computer Science

AI Seminar & Defense Event

Learning Better Clinical Risk Models

Alexander Van Esbroeck

Monday, May 04, 2015
1:00pm - 3:00pm
3725 Beyster Bldg.

Add to Google Calendar

About the Event

Risk models are used to estimate a patient’s risk of suffering particular outcomes throughout clinical practice. These models are important for matching patients to the appropriate level of treatment, for effective allocation of resources, and for fairly evaluating the performance of healthcare providers. The application and development of methods from the field of machine learning has the potential to improve patient outcomes and reduce healthcare spending with more accurate estimates of patient risk. This dissertation addresses several limitations of currently used clinical risk models, through the identification of novel risk factors and through the training of more effective models. As wearable monitors become more effective and less costly, the previously untapped predictive information in a patient’s physiology over time has the potential to greatly improve clinical practice. However translating these technological advances into real-world clinical impacts will require computational methods to identify high-risk structure in the data. This dissertation presents several approaches to learning risk factors from physiological recordings, through the discovery of latent states using topic models, and through the identification of predictive features using convolutional neural networks. We evaluate these approaches on patients from a large clinical trial and find that these methods not only outperform prior approaches to leveraging heart rate for cardiac risk stratification, but that they improve overall prediction of cardiac death when considered alongside standard clinical risk factors. We also demonstrate the utility of this work for learning a richer description of sleep recordings. Additionally, we consider the development of risk models in the presence of missing data, which is ubiquitous in real-world medical settings. We present a novel method for jointly learning risk and imputation models in the presence of missing data, and find significant improvements relative to standard approaches when evaluated on a large national registry of trauma patients.

Additional Information

Sponsor(s): Professors Zeeshan H. Syed and Satinder Singh Baveja

Open to: Public