Artificial neural networks for the classification of electrophoretic patterns

The analysis of serum protein electrophoretic patterns is routinely performed in clinical settings to help in the diagnosis of many pathological states. Using an artificial neural network (ANN) approach we have developed a system which can classify the patterns, Four networks were developed, which used supervised learning with either the Linear Vector Quantization (LVQ) or the Back Propagation (BP) algorithm and the classification was made either among 3 classes (physiological, suspect and pathological) (3C) or among 6 classes (6C), with differentiation of pathological states. After training, the networks were tested on a validation set of 153 routinely collected patterns of different classes. In the 3C case the LVQ approach exhibited better performances, with a rate of 91% correct diagnoses, while the BP network was better in the 6C case, with 80% of correct classifications, These results were judged quite satisfactory, also considering that the aim of the ANN classifiers was mainly to screen out the suspect or pathological cases, which had then to be submitted to further and more sophisticated investigations