Risk Management Collaboratory - Risk Models from Text


Bayesian networks are commonly used to capture domain knowledge for probabilistic models especially in medical decision support systems.

 

Three approaches exist for building such models: (1) learning information from data, (2) eliciting from domain expert, and (3) having a knowledge engineer extract information from literature. While the first approach can be automated, the other two require a significant amount of manual work which can make them impractical on a large scale.

With this research, our intention is to provide decision makers with a method to automatically generate the skeleton of a Bayesian network and develop an authoring tool for such networks.

We focus initially on networks automatically extracted from medical articles. Our goal is to jump start the creation of models related to customized decision support models in healthcare settings.

What if we could

  • Automatically pre-process medical articles to extract relevant information
  • Facilitate aggregation
  • Provide an editing interface for the automatically generated model

While keeping track of the supporting evidence behind modeling elements?

 

diagram showing texts being processed into Bayesian networks.

References

Text Mining for Causal Relations. Girju R and Moldovan D I. (2002)

Acquiring Bayesian Networks from Text. Sanchez-Graillet O and Poesio M. (2004)

Bridging Text Mining and Bayesian Networks. Raghuram S, Xia Y, Palakal M, Jones J, Pecenka D, Tinsley E, Bandos J and Geesaman J (2009)