Using Belief Functions for Uncertainty Management and Knowledge Acquisition: An Expert Application

TL;DR

This paper explores the use of Belief Functions (Dempster-Shafer theory) in medical diagnosis to improve uncertainty management and knowledge acquisition from data and expert opinions.

Contribution

It introduces methods for integrating data and expert opinions using Belief Functions and compares their effectiveness in a medical diagnosis context.

Findings

One formulation statistically outperforms others including Shafer's method.

Expert opinions and data-derived dependencies are effectively combined.

Uncertainty management techniques enhance knowledge acquisition from medical data.

Abstract

This paper describes recent work on an ongoing project in medical diagnosis at the University of Guelph. A domain on which experts are not very good at pinpointing a single disease outcome is explored. On-line medical data is available over a relatively short period of time. Belief Functions (Dempster-Shafer theory) are first extracted from data and then modified with expert opinions. Several methods for doing this are compared and results show that one formulation statistically outperforms the others, including a method suggested by Shafer. Expert opinions and statistically derived information about dependencies among symptoms are also compared. The benefits of using uncertainty management techniques as methods for knowledge acquisition from data are discussed.