Tutorial AIME 2017: Medical Decision Analysis with Probabilistic Graphical Models

Medical Decision Analysis with Probabilistic Graphical Models

Tutorial at the 16th Conference on Artificial Intelligence in Medicine (AIME-2017). Vienna, Austria, June 21-24, 2017.

Teachers

Abstract

Probabilistic graphical models offer important advantages over the methods traditionally used in medicine:

Bayesian networks are much more accurate than the naïve Bayes method, which consists of multiplying the likelihood ratios of the findings.
Influence diagrams and decision analysis networks can be used to build models equivalent to decision trees containing thousands of branches, and to perform cost-effectiveness analysis for problems involving several decisions.
Markov influence diagrams can solve problems much more easily than when using spreadsheets, Markov decision trees, or a programming language such as R or C++. They can model various features of the patient without multiplying the number of states; in particular, they can represent the patient history without using tunnel states.
Factored Markov decision processes are especially suited for follow-up problems; for example, to determine the optimal time for an intervention.

This tutorial will present probabilistic graphical models and illustrate their application to medical decision analysis.

Motivation

Probabilistic graphical models have been extensively used in artificial intelligence, but in spite of their advantages most health decision analysts and health economists prefer to use traditional methods, such as the naïve Bayes, decision trees and spreadsheet Markov models. In this tutorial we will explain how probabilistic graphical models can be applied to medical diagnosis and to unicriterion and cost-effectiveness analysis. The tutorial can also serve as an introduction to health technology assessment, one of the branches of health economics.

Results of learning

Participants will learn how to:

build Bayesian networks for medical diagnosis and compare them with naïve Bayes models;
build decision analysis networks for unicriterion and cost-effectiveness analysis and compare them with decision trees;
build Markov influence diagrams for cost-effectiveness analysis and compare them with spreadsheets, Markov decision trees, and discrete event simulation.

The course will also present other temporal models, such as Markov decision processes (fully and partially observable) and dynamic limited-memory influence diagrams.

Format

The tutorial will introduce the basic concepts about probabilistic graphical models, medical diagnosis, health decision making and cost-effectiveness analysis, and explain them with real medical examples.

The construction and evaluation of these models will be illustrated with OpenMarkov, an open-source software package for probabilistic graphical models.

Attendants will obtain more benefit if they previously read OpenMarkov's tutorial and reproduce the examples it contains.

Audience

The tutorial can be followed by beginners because it does not require any knowledge of probabilistic graphical models or health economics. However, it will also address topics com-bining techniques from these two fields, which may make the course interesting for researchers proficient in any of them.