Do Alzheimer's Patients Appear Younger than Their Age? A Study with Automatic Facial Age Estimation

Abstract

Facial age estimation from images is a challenging task, especially if the subjects are older, since idiosyncratic variations increase with age, and lifestyle factors have an impact on the appearance. In this paper, we test the hypothesis that the Alzheimer's patients appear younger than their real age. To do this objectively, and to be able to analyze the factors in case the hypothesis holds true, we use automatic age estimation methods for this task. We first propose training and normalization regimes to improve deep learning based facial age estimation. We fine-tune a pre-trained ImageNet model using first the APPA-REAL database and then the UTKFACE database. The experimental results show that the proposed approach predicts older faces more accurately compared to other studies, and improves the mean absolute error for the FG-NET database to 8.14 for the age group 60-69. We then run our approach on a special database collected from Alzheimer's patients and healthy controls, to test our main hypothesis. The database we collected for this purpose contains video recordings of 96 subjects with an age range between 64 and 87. Our findings show that automatic age estimation indeed underestimates the age of AD patients significantly more than the healthy subjects.