Data extraction for improved prediction outcomes in organ transplantation - A hybrid approach

Neural network ensembles have made an impressive contribution in a number of different medical domains. Like simple neural network models, the neural network ensembles are known as 'black boxes' since how the outputs are produced is not obvious. Due to this limitation these techniques are not widely used by medical professionals. This paper first provides a short review of the different neural network rule extraction techniques. Then it describes a novel approach, namely "RDC-ANNE" that is designed to extract useful explanations from several combined neural network classifiers. The methodology employed utilises a dataset made available to us from a kidney transplant database. The dataset embodies a number of important properties, which make it a good starting point for the purpose of this research. Results reveal that this approach can be used to identify and extract the regions in the data space that have positive impact on the system performance, provide useful explanations from several combined neural networks and enhance the overall utility of current neural network models.