Detection of type-specific herpes virus antibodies by neural network classification of Western blot densitometer scans

A fully connected, feedforward, three-layer (350-3-3) neural network (NN) for Western blot densitometer scan (WBDS) pattern classification for diagnosing B virus infections in humans is developed. NN supervised training used backpropagation. The training set consists of average WBDSs for three classification groups. Optimum NN parameters for this NN topology and application are determined. The NN achieves a correct classification rate of 85% and an incorrect classification rate of 2%. The average NN true positive rate is 0.87, and false positive rate is 0.02. Receiver operating characteristic curve analysis of NN classification performance demonstrated excellent results. A NN can be successfully trained to classify WBDSs with performance superior to that of humans.