Design of artificial neural network and its applications to the analysis of alcoholism data

Artificial neural networks were applied to the alcoholism data to reveal nonlinear relationships between intermediate phenotypes, marker identity-by-descent sharing, and the affection status. A variable number of hidden units were considered to achieve a balance between the minimal mean-squared error and over-fitting of the data. The predictability of the affection status based on intermediate phenotype information (event-related potential 300, monoamine oxidase, and gender) was 65% to 75%, and sensitivity/specificity ranged around 50% to 80%. The IBD approach succeeded in identifying the same marker as previous studies, but also found additional peaks.