Human influenza viruses cause annual epidemics due to antigenic drifts in the hemagglutinin protein. Five antigenic sites in the influenza H3 hemagglutinin protein have been proposed and 131 amino acid positions have been identified in the five antigenic sites. A previous study had documented that a model based on the 131 positions in the five antigenic sites could moderately predict antigenic variants of influenza A/H3N2 viruses (agreement=83%). In this study, prediction models combining serology, bioinformatics and statistics were developed to predict antigenic variants of influenza A/H3N2 viruses. Amino acid sequences of hemagglutinin protein of 45 A/H3N2 viruses isolated during 1971-2002 and 181 pairwise antigenic distances determined by antibody cross-reactivity among the 45 viruses were analyzed as training dataset. In addition, 57 pairwise antigenic distances from 12 A/H3N2 viruses isolated during 1999-2004 were used as validation dataset. Multivariate regression models were employed to identify potential immunodominant positions and predict antigenic variants. Seventeen amino acid positions were identified as potential immunodominant positions in the training dataset. Prediction models based on the potential immunodominant positions have improved performance on predicting antigenic variants in the training (agreement=91%) and validation (agreement=93%) datasets. The model could be readily integrated to the global influenza surveillance system.