This study investigates the possibility of using facial images to determine if a non-depressed person is likely to develop clinical depression within the next 1-2 years. The prediction method uses a typical classification approach of training and testing. Class models were determined using image data from adolescents who were either "at risk" or "not at risk" of depression. The risk factor was confirmed through 2 years of follow-up data collection. Two feature extraction approaches were compared; the eigenface (PCA) features and the fisherface (PCA+LDA) feature. The nearest neighbor (NN) classification was implemented using person independent and person dependent approaches. Best results were with the fisherface (PCA+LDA) method, providing prediction accuracy of 51% with the person independent approach and 61% when using a person dependent approach.