Tutorials

• Overview
• Evaluation of Prognostic Models - CANCELLED
• Evolutionary Computation Approaches to Mining Biomedical Data
• Causal Discovery from Biomedical Data
• Applied Data Mining in Clinical Research

Overview

Follow the link in each tutorial to see more details. Tutorials will be held on Saturday 23rd July 2005. It is possible to attend one morning and one afternoon tutorial (but not two in the morning or two in the afternoon).

Register for tutorials at the same time as you register for the main conference. It is possible to register for a tutorial without attending the main conference, but an additional charge will be levied as a contribution to conference overheads.

Tutorial 1: (morning)

Evaluation of Prognostic Models: Ameen Abu-Hanna and Niels Peek (University of Amsterdam)

We are sorry, but due to the low number of registrations for this tutorial, it has had to be cancelled.

Tutorial 2: (morning: 09:30 to 12:30)

Evolutionary Computation Approaches to Mining Biomedical Data: John Holmes (University of Pennsylvania)

Evolutionary computation (EC) encompasses a variety of Darwinian and genetics-inspired machine learning approaches, including the genetic algorithm, learning classifier systems, genetic programming, and evolutionary algorithms. Interest in using EC as a foundation for building data mining tools has been growing at a fast rate over the past several years, as evidenced by the burgeoning of papers and presentations related to data mining and knowledge discovery in EC-related literature. The advantages of EC-based data mining tools include easily understood algorithmic and knowledge representations, graceful learning in supervised as well as unsupervised environments, and demonstrated high levels of classification and prediction accuracy of evolved models. However, several problem areas remain, including scalability and reduction of model complexity, although these are currently under intense investigation.

After a discussion of the general characteristics of biomedical data, such as missing values and feature selection problems, and methods for preparing biomedical data for mining, this tutorial will introduce the various EC-based tools for mining biomedical data, including thorough algorithmic descriptions, functional examples, and live demonstrations of each on several real-world biomedical datasets. The applications will focus specifically on rule discovery, emergence of clinical prediction rules, and clustering, as appropriate to each EC method.

The tutorial will include a rigorous discussion of methods for evaluating the results obtained by applying EC tools to mining biomedical data, including classification and prediction accuracy and test characteristics such as sensitivity, specificity, area under the receiver operating characteristic curve, and predictive values. Finally, methods for validating the classification and prediction models discovered by the EC tools will be presented and discussed. These include the choice and use of suitable validation datasets, methods for comparing EC-derived models with those derived by other data mining tools such as decision tree induction and logistic regression, and the use of human expert panels in providing content for qualitative model validation.

Ample time will be provided throughout the tutorial for questions and discussion.

Tutorial 3: (afternoon: 14:00 to 17:00)

Causal Discovery from Biomedical Data: Subramani Mani (University of Pittsburg)

Discovering causal relationships is a useful pursuit. Causal knowledge is helpful for planning and decision making. For example, knowing the cause of a disease helps in prevention and treatment.

Well designed experimental studies, such as randomized controlled trials, are typically employed in assessing causal relationships. Here the value of the variable postulated to be causal is set randomly and its effects measured. These studies are appropriate in certain situations, for example, animal studies and studies involving human subjects that have undergone a thorough procedural and ethical review. Experimental studies may not, however, be feasible in many contexts due to ethical, logistical, or cost considerations. These practical limitations of experimental studies highlight the importance of exploring, evaluating and refining techniques to learn more about causal relationships from observational data, as for example data routinely collected in astronomy, earth sciences and healthcare. The aim is not to replace experimental studies, which are extremely valuable in science, but to complement experimental studies when feasible.

This tutorial will focus on data driven approaches to learning causal relationships as opposed to philosophical approaches to causality. The topic is closely related to machine learning, data mining and knowledge discovery that are key areas represented in the broad field of “Artificial Intelligence in Medicine”.

Tutorial 4: (afternoon: 14:00 to 17:00)

Applied Data Mining in Clinical Research: John Holmes (University of Pennsylvania)

Health care is becoming increasingly data-driven. The proliferation of databases in every quarter of health care practice and research is evident in the large number of claims databases, registries, electronic medical record data warehouses, disease surveillance systems, and ad hoc research database systems. The number of databases grows daily, but even more importantly, so does the amount of data within them. Pattern-identification tasks such as detecting associations between certain risk factors and outcomes, ascertaining trends in health care utilization, or discovering new models of disease in populations of individuals rapidly become daunting even to the most experienced health care researcher or manager. In many systems, data have become so large as to overwhelm traditional statistical approaches for performing these tasks. An alternative approach uses a variety of methods drawn from statistics and machine learning disciplines to mine databases for patterns that may be missed using traditional techniques.

Up until recently, much of the work of data mining has been the domain of a small number of computer scientists, programmers, data base administrators, and management information specialists. However, new tools have put the power to mine databases into the hands of end-users. This tutorial is intended to bring together AMIE attendees with an interest or experience in mining medical databases of all kinds, introducing attendees to the practical application of data mining. Using a well-known data mining life cycle as a conceptual framework, attendees will experience first-hand, thorough demonstration and direct participation, the techniques of mining clinical data. Specifically, the tutorial will cover: Introduction to Weka, an open source data mining software suite, and the demonstration databases; Data preparation and reduction; Data description and visualization; Association rule mining; Clustering; Classification and prediction rule mining; Interpreting and applying the results to analysis; and Summary and conclusion.

This tutorial has been designed as a “hands-on” practical experience. The Weka data mining software suite will be distributed on CD to all attendees for use in the tutorial. Thus, each attendee is encouraged to bring a Windows-based laptop computer with CD-ROM drive that supports the Java Runtime Environment (JRE), version 2, for installation and use of the software during the tutorial. The CD will include a copy of the appropriate JRE, should it be needed, as well as several biomedical datasets that will be used as examples in the session. Laptops are not required, however; those without laptops will still benefit from the tutorial by means of the live demonstrations and the opportunity to work with other attendees.

Ample time will be provided throughout the tutorial for questions and discussion.