MIT EECS Undergraduate Research and Innovation Scholar
Feature Representation for Predicting ICU Mortality
Collin M. Stultz
Current representations of features in clinical data may underrepresent differences, in the case of features reduced to binary values, or over-represent them, in the case of features which are kept at their exact value. I aim to critically and quantitatively evaluate different feature representations. In addition to the representations that are already widely used, I introduce two new representations and compare all of these with the aim of defining a methodology for choosing the best set of features, representations, and model for building predictive models in a clinical setting. I evaluate the models based on their accuracy in predicting the outcomes of ICU patients from the MIMIC II & III databases.
I’ve been interested in machine learning ever since I took 6.034, and I think it can be very impactful in changing the way that healthcare is administered. I’ve spent the last two summers doing data science internships at health-focused companies, and I am excited to delve deeper into a large medical database and use a variety of machine learning techniques in this project.