Interpretability of Bayesian Decision Trees Induced from Trauma Data
Decision Tree (DT) technology with its sequential decision process composed of hard thresholds applied to individual features is often not an optimal choice. But DTs are typically credited with being predictive structures that a domain specialist can interpret as opposed to a ‘black-box’ neural network, for example. There are two anticipated classes of benefit when Artificial Intelligence technology is applied to medicine: automation of decision making and enhancement of medical knowledge. In this paper we present the use of Bayes-ian averaging as a principled approach to optimal classifier systems using DT technology in which a confidence rating can be associated with every predicted result. However, averaging over an ensemble of DTs causes the problem that such an ensemble becomes uninterpretable. Thus we also present a procedure for extracting interpretable ‘archetype’ DTs. We demonstrate these innovations by application to Trauma data.
KeywordsMarkov Chain Monte Carlo Glasgow Coma Score Markov Chain Monte Carlo Method Decision Tree Model Trauma Data
Unable to display preview. Download preview PDF.
- 2.Galant, S.: Neural Network Learning and Expert Systems. MIT Press, Cambridge (1993)Google Scholar
- 4.Quinlan, J.: C4.5. In: Programs for Machine Learning, Morgan Kaufmann, San Francisco (1993)Google Scholar
- 8.Denison, D., Holmes, C., Malick, B., Smith, A.: Bayesian Methods for Nonlinear Classifica-tion and Regression. Wiley, Chichester (2002)Google Scholar
- 11.Chipman, H., George, E., McCulloch, R.: Making Sense of a Forest of Trees. In: Weisberg, S. (ed.) Proceedings of the 30th Symposium on the Interface, Interface Foundation of North America, pp. 84–92 (1998)Google Scholar
- 13.Schetinin, V., Partridge, D., Krzanowski, W.J., Everson, R.M., Fieldsend, J.E., Bailey, T.C., Hernandez, A.: Experimental Comparison of Classification Uncertainty for Randomized and Bayesian Decision Tree Ensembles. J. Math. Modeling and Algorithms 4 (forthcoming, 2006)Google Scholar
- 14.Blake, C.L., Merz, C.J.: UCI Repository of Machine Learning Datasets. Irvine, University of California (1998), cs.uci.edu/mlearn/MLRepository