Analysis of tear protein patterns by a neural network as a diagnostical tool for the detection of dry eyes

The electrophoretic patterns of tears from patients with dry-eye disease (n = 43) and from healthy subjects (n = 17) were analyzed by means of multivariate statistical methods and an artificial neural network (ANN), following sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE). From each electrophoretic pattern a data set was created, randomly divided into test (unknown samples) and training patterns (known samples), with ANN training by one of these sets. After training, the performance of the ANN was checked by presenting the test data set to the ANN. Furthermore, the data was classified using multivariate analysis of discriminance. The groups were significantly different from each other (P<0.05). The statistical procedure yielded 97% (known samples) and 71% (unknown samples) correct classifications. The ANN revealed 89% of correct classifications using the test set (unknown samples). The use of pruning algorithms (optimization procedure which automatically eliminates small weighted neurons) or genetic algorithms (optimization procedure which performs genetically induced changes of the neural net) resulted in a slight decrease of correct classifications compared to those of the nonoptimized neural network. The results reveal significant differences between the two groups. Using the ANN we were able to classify the electrophoretic tear protein pattern for diagnostic purposes.