Neural network to identify glaucomatou viual field progreion

Purpoe To decribe a method to determine progreion of glaucoma baed on viual field threhold. Deign Obervational retropective longitudinal cohort tudy. Method A back propagation neural network with three hidden layer wa developed with commercial oftware. Viual field data from 80 patient who participated in the Advanced Glaucoma Intervention tudy (AGI) were ued. Glaucomatou viual field progreion wa defined a a change of 4 or more unit in the AGI core, confirmed by at leat two equential ubequent tet. Input to the neural network conited of threhold meaurement from 55 viual field location from the baeline examination and each follow-up examination. The data et wa randomized o the equence of examination would not influence the training or teting of the neural network. Two third of the randomized data were ued for training and the remaining one third for teting. Reult The mean age of 80 patient enrolled in AGI at initial examination wa 67.4 (± 7.3 tandard deviation [D]) year. The average follow-up period wa 7.2 (±2.3 D) year and the mean duration between examination wa 0.46 (± 0.39 D) year. The neural network etimated the probability of progreion for each baeline and follow-up comparion with an average enitivity of 86% and pecificity of 88%. The area under the receiver operating characteritic (ROC) curve wa 0.92, with a enitivity of 86% at the 80% pecificity level and a enitivity of 91% at the 90% pecificity level. Concluion From analyi of AGI data, progreion of glaucoma could be detected from viual field threhold with a neural network.