Validating a Topic Model that Predicts Pancreatic Cancer from Latent Structures in the Electronic Medical Record

Project Summary

Furthering the work of a 2016 Data+ team in predictive modeling of pancreatic cancer from electronic medical record (EMR) data, students Siwei Zhang (Masters Biostatistics) and Jake Ukleja (Computer Science) spent ten weeks building a model to predict pancreatic cancer from Electronic Medical Records (EMR) data. They worked with nine years worth of EMR data, including ICD9 diagnostic codes, that contained records from over 200,000 patients.

Themes and Categories
Year
2017
Contact
Paul Benich
bendich@math.duke.edu

Project Results: The team began with exploratory data analysis that illustrated median times of appearance and frequency of specific ICD9 codes, with an eye toward understanding the relation between these statistics and pancreatic cancer diagnosis. They then trained a topic model which predicted past pancreatic cancer diagnosis with high accuracy (93 percent AUC) from ICD9 codes. Finally, they used the topic model outcomes to identify a pool of high-risk patients for potential future study.

Click here for the Executive Summary

Project Leads:

Lisa Satterwhite, PhD

James Abbruzzese, MD

Joseph Lucas, PhD

Project Manager: Tyler Massaro

Related People

Related Projects

Social and environmental contexts are increasingly recognized as factors that impact health outcomes of patients. This team will have the opportunity to collaborate directly with clinicians and medical data in a real-world setting. They will examine the association between social determinants with risk prediction for hospital admissions, and to assess whether social determinants bias that risk in a systematic way. Applied methods will include machine learning, risk prediction, and assessment of bias. This Data+ project is sponsored by the Forge, Duke's center for actionable data science.

Project Leads: Shelly Rusincovitch, Ricardo Henao, Azalea Kim

Project Manager: Austin Talbot

Producing oil and gas in the North Sea, off the coast of the United Kingdom, requires a lease to extract resources from beneath the ocean floor and companies bid for those rights. This team will consult with professionals at ExxonMobil to understand why these leases are acquired and who benefits. This requires historical data on bid history to investigate what leads to an increase in the number of (a) leases acquired and (b) companies participating in auctions. The goal of this team is to create a well-structured dataset based on company bid history from the U.K. Oil and Gas Authority; data which will come from many different file structures and formats (tabular, pdf, etc.). The team will curate these data to create a single, tabular database of U.K. bid history and work programs.

Project Lead: Kyle Bradbury

Project Manager: Artem Streltsov

Producing oil and gas in the Gulf of Mexico requires rights to extract these resources from beneath the ocean floor and companies bid into the market for those rights. The top bids are sometimes significantly larger than the next highest bids, but it’s not always clear why this differential exists and some companies seemingly overbid by large margins. This team will consult with professionals at ExxonMobil to curate and analyze historical bid data from the Bureau of Ocean Energy Management that contains information on company bid history, infrastructure, wells, and seismic survey data as well as data from the companies themselves and geopolitical events. The stretch goal of the team will be to see if they can uncover the rationale behind historic bidding patterns. What do the highest bidders know that other bidders do not (if anything)? What characteristics might incentivize overbidding to minimize the risk of losing the right to produce (i.e. ambiguity aversion)?

Project Lead: Kyle Bradbury

Project Manager: Hyeongyul Roh