Diagnostic Decision Support with Bayesian Networks
In this webinar, we will illustrate how Bayesian networks can serve as a practical tool for optimizing the sequence of diagnostic steps with the objective of arriving at a medical diagnosis in a quick and cost-efficient manner. Bayesian networks allow us to precisely quantify the amount of information contributed from each to-be-observed variable, such as risk factors and symptoms. This capability is one of the key points whereby machine-learned Bayesian networks distinguish themselves from other predictive models, e.g. neural networks.
We will utilize the dataset published by Dr. Zahra Alizadeh Sani on Coronary Artery Disease to demonstrate a complete research workflow, from importing the raw data all the way through publishing a final model with a web interface.
Workflow with the BayesiaLab Software Platform:
- Data Import into BayesiaLab.
- Discretization of continuous variables.
- Definition of variable classes.
- Supervised Learning using the Markov Blanket and Augmented Markov Blanket algorithms.
- Structural Coefficient Analysis for Bayesian network model optimization.
- Network Performance Analysis with regard to one or multiple Target Nodes (Stenosis of LAD, LCX, or RCA).
- Introduction to information-theoretic concepts, such as Entropy and Mutual Information.
- 2D Mapping to illustrate Mutual Information between variables and Target Nodes.
- Computation of an interactive and dynamic Adaptive Questionnaire for optimized evidence-seeking with regard to the diagnosis.
- Introduction of the cost of diagnostic procedures for optimization, i.e., trading off the cost of information gain vs. the expected reduction of uncertainty.
- Computation of Target Interpretation Tree as a static decision support tool.
- Publication of the Adaptive Questionnaire to the BayesiaLab WebSimulator as a decision support tool for external users.
Cardiovascular Imaging Department, Rajaei Cardiovascular, Medical & Research Center, Iran University, Tehran, Iran
Updated (Extended) Dataset:
Original (Limited) Dataset: