Performance evaluation of artificial intelligence classifiers for the medical domain

The application of artificial intelligence systems is still not widespread in the medical field, however there is an increasing necessity for these to handle the surfeit of information available. One drawback to their implementation is the lack of criteria or guidelines for the evaluation of these systems. This is the primary issue in their acceptability to clinicians, who require them for decision support and therefore need evidence that these systems meet the special safety-critical requirements of the domain. This paper shows evidence that the most prevalent form of intelligent system, neural networks, is generally not being evaluated rigorously regarding classification precision. A taxonomy of the types of evaluation tests that can be carried out, to gauge inherent performance of the outputs of intelligent systems has been assembled, and the results of this presented in a clear and concise form, which should be applicable to all intelligent classifiers for medicine.