"Simplicity, Predictive Performance, and Comprehensive Decision Support: Three Challenges for 'Intelligent' Clinical Systems"

11/30/07: 10:30 AM, webcast
11th Floor Large Conference Room
Host: Clayton Morrison, schedule

Abstract: The move of the medical profession towards evidence-based medicine combined with a proliferation of clinical information creates research opportunities to analyze clinical data for disease patterns or treatment characteristics. The medical informatics community has seized this opportunity by working on research ranging from automatic generation of clinical practice guidelines to data-driven decision support. All this research has resulted to some extent, in a number of clinical decision support systems created using less or more sophisticated data analysis methodologies. However, given demands such as interpretability and simplicity on one hand and elevated level of predictive performance demanded by the physicians on the other, many such efforts have not readily transferred to clinical practice. This talk will look at this issue by discussing research on the MET clinical system (www.mobiledss.uottawa.ca). It will present a process of applying machine learning to discover clinical knowledge about assessing severity of pediatric asthma exacerbation from data collected in the emergency department of a hospital. It will discuss data pre-processing, including contextual normalization and will talk about development of a decision model (treebased and rule-based) created using C4.5 and RIPPER. Moreover, we will explain why other models (KNN, Naive Bayes, SVM) were found to be less appropriate. We will describe our research on an evidence-based software agent to retrieve clinical evidence from the Cochrane Library and how such agent uses results from the prediction model. At the end we will show why broad research focus is necessary in order to create a MET multi-agent system for point of care support.

Last updated: Wed Jan 27 11:30:01 2010