Seizure Prediction & Machine Learning

Abstract

Epileptic seizures are a common neurological disorder, affecting 1% of the world’s population.  A major source of disability among epileptic patients is uncertainty around when their next seizure will occur, which often leads to anxiety and self-imposed limitations on activities.  This talk describes the first medical device designed for long-term, ambulatory monitoring of brain EEG activity in epileptic patients.  The device includes a real-time advisory system that can predict increases in seizure likelihood up to hours in advance of a seizure.  The latter part of the talk focuses on the speaker’s work developing a second-generation algorithm for the advisory system, using techniques from signal processing and machine learning to create new predictive features and a simple but robust predictive model.  Long-standing issues with the statistical validation of such models will be discussed.

About Jeff Howbert

Jeff Howbert received a BA in English from Stanford Univ. in 1977 and a PhD in Synthetic Organic Chemistry from Harvard Univ. in 1983. Over the ensuing 25 years, he led medicinal chemistry and drug discovery efforts at a large pharmaceutical company and several small biotech companies. He holds 45 US patents and is responsible for the entry of 6 compounds into clinical development. After earning a MS in Computer Science from Univ. of Washington in 2008, he began a second career in computational biology, with an emphasis on machine learning. He has worked in several labs on building predictive models for diverse biomedical problems, including seizure risk, cardiovascular biomarker discovery, and proteomic analysis. He is also currently teaching a course at UW Bothell on machine learning.

Using Natural Language Processing and Machine Learning in Medical Information Systems

Abstract:

Atigeo (http://atigeo.com) is a Bellevue-based software company offering a big data analytics platform that includes a number of natural language processing and machine learning components used in a variety of domains and industries. One area of focus is medical information systems that require the extraction of information from unstructured and semi-structured corpora. 

The first portion of our presentation will describe some of the NLP and ML techniques we use in our text analytics framework to support a computer-aided coding product. Our CAC system identifies medically relevant terms in an electronic medical record, and uses a large suite of classifiers to assign diagnostic and procedure codes to that EMR. 

The second portion of our presentation will describe our participation in the 2012 Text REtrieval Conference (http://trec.nist.gov), an annual multi-track evaluation of large-scale information retrieval systems sponsored by the National Institute of Standards and Technology. The TREC 2012 Medical Records Track task involved the identification of EMRs that are relevant to a short search queries describing medical diagnoses, tests and/or treatments. Our system achieved some of the highest scores reported by the 24 participants, across several different metrics.

Our TREC success was largely the result of the work of two dedicated software engineering interns. We are currently recruiting interns to help us participate in the TREC 2014 Clinical Decision Support track, which involves the retrieval of biomedical articles from the PubMed relevant for answering generic clinical questions about diagnoses, tests and treatments of medical conditions.

Bios:

Bryan Tinsley is a former intern and current Software Engineer at Atigeo, with a BS in computer science and software engineering from UW Bothell and a BA in linguistics from Western Washington University. Alex Thomas is a former intern and current Software Engineer at Atigeo, with a BS in computer science and a BA in mathematics from UW Seattle. Joe McCarthy is Director, Analytics & Data Science at Atigeo, and a Senior Lecturer in the Computer and Software Systems program at UW Bothell (currently on leave), with a PhD in computer science from the University of Massachusetts.