Incremental learning in a multilayer neural network as an aid to Alzheimer´s disease diagnosis

In medical diagnostic expert systems, despite some interesting and successful results, significant shortcomings were encountered, due to an inherent risk of errors, the complexity of the problems considered, and the difficulty of accessing the expert knowledge held by physicians. Many techniques have been proposed in the field of artificial neural networks (ANNs) allowing knowledge to be directly learned from experimental results and medical practice. In this paper an aid to Alzheimer´s disease diagnosis is suggested using cerebral blood flow (CBF) measurements from single photon emission computed tomography (SPECT) which facilitate ANN training. An incremental learning technique is presented to memorize new input-output associations without necessarily starting the learning phase all over again. Two steps are proposed: (1) Off-line learning of the overall features that remain unchanged during any implementation, based on the patient´s CBF measurements studied in the laboratory. (2) In-situ learning of each site´s features: The physician can enter new input-output associations. In case of failure, a constructing procedure adds and trains new unit designed to represent new associations. Thus, learning is being added incrementally. This approach turns out to be more particularly useful for medical diagnosis in which the system designer does not possess the complete learning set right from the start. Indeed pieces of information are added as they become available, thus allowing for a man-machine dialogue to take place.