The present investigation aims at the construction of a sequential decision procedure based on a partially observable Markov decision process (POMDP) model. An optimal clinical management strategy based on a risk assessment of patients is to be found. A decision theoretic cost model different from the general approach has been selected for this clinical management (and classification) task: costs were determined by specifying a minimum acceptable sensitivity and specificity of the overall procedure. The aim is to find the earliest possible decision epoch where a final decision can be made under these quality restrictions. Solution method is non-linear optimisation combined with a robust partial classification method. The probabilities necessary for the model are estimated from data of a clinical study in liver transplantation patients. Decision epochs were at donor organ assessment, immediately before surgery and postoperatively in the intensive care unit. Parameters obtained within decision epochs were combined to scores by artificial neural networks (ANNs). The encouraging results show the applicability of the model in the clinical setting.
Keywords: Diagnosis, Planning, Probabilistic Networks, Neural Networks, Machine Learning, Uncertainty in AI
Citation: Guenter Tusch: Optimal Sequential Decisions in Liver Transplantation Based on a POMDP Model. In W.Horn (ed.): ECAI2000, Proceedings of the 14th European Conference on Artificial Intelligence, IOS Press, Amsterdam, 2000, pp.186-190.