|
The aim of the proposed work is to provide methodological and software tools for the establishment of neural-network-based decision support systems which can learn autonomously. The main objectives are:(i) To develop robust computational models of Adaptive Resonance Theory networks.(ii) To devise a novel, hierarchical structure to allow for problem decomposition.(iii) To implement a prototype decision support development tool.(iv) To prove the concept by developing a medical decision aid.Progress:Work to date has concentrated on objective (i) with some attention to objectives (ii) and (iv). From January 1995 attention has shifted towards objective (ii). This is expected to continue throughout 1995. Some significant results have been obtained and six items related to the project have been accepted for publication.Objective (i):The main thrust here has been to overcome certain shortcomings in the chosen neural network. This has been done successfully so that a major difficulty associated with a sensitivity of the networks performance to the order in which data are presented has been removed. This is a by-product of a modification which enables the system to converge to the theoretically optimal classification rate in a non-stationary environment. [1], [3], [5].Objective (ii):A hierarchical structure comprising layers which classify highly sensitively, high specifically or highly accurately has been developed. [2], [4], [6]. This improves overall classification rates in samples with disparate prior distributions, a notoriously difficult task in statistical pattern recognition. Work continues on a hierarchical structure to fuse disparate data sources and to assist with the problem of missing data.Objective (iii):The results above have been applied to medical diagnostic [4], [6] and prognostic [2] data. These have been successful but have only demonstrated the applicability of the techniques off-line. A new data set to be provided by the co-investigator in late spring 1995 should be adequate to demonstrate their on-line, autonomous capacity.Objective (iv):Work here is at the planning stage.1. C.P.Lim and R.F. Harrison, 1995, Neural Networks, Modified fuzzy ARTMAP approaches Bayes optimal classification rates: an empirical demonstration.2. J. Downs, R.F. Harrison, R.L. Kennedy and K. Woods, 1995, Proceedings of the International Conference on Neural Networks and Genetic Algorithms, Al+s, The use of fuzzy ARTMAP to identify low risk patients hospitalised with acute chest pain.3. C.P. Lim and R.F. Harrison, 1995, Proceedings of the International Conference on Neural Networks and Genetic Algorithms, Ales, Minimal error rate classification in a non-stationary environment via a modified fuzzy ARTMAP network.4. J. Downs, R.F. Harrison and S.S. Cross, 1995, Proceedings of the 10th Biennial Conference on AI and Cognitive Science, Sheffield, A neural network decision support tool for the diagnosis of breast cancer.5. C.P. Lim and R F Harrison, 1995, Proceedings of the Institution of Electrical Engineers Fourth International Conference on Artificial Netral Networks, Cambridge, Probabilistic Fuzzy ARTMAP: An autonomous neural network architecture for Bayesian probability estimation.6. J. Downs, S.S. Cross, R.F. Harrison and T.J. Stephenson, 1995, Journal of Pathaology. An adaptive resonance theory mapping neural network provides confirmation of criteria cited by human experts in the cytodiagnosis of breast fine needle aspirated.
|