Early diagnosis of Alzheimer disease (AD), while still at the stage known as Mild Cognitive Impairment (MCI), is important for the development of new treatments. However, this is a difficult task because the spatial pattern of brain degeneration in MCI is highly variable and changes as the disease progresses. Despite these difficulties, many machine learning techniques have already been used for the diagnosis of MCI and for predicting MCI to AD conversion, but the MCI group used in previous works is usually very heterogeneous containing subjects at different stages. The goal of this paper is to investigate how the disease stage impacts on the ability of machine learning methodologies to predict conversion. After identifying the converters and estimating the time of conversion (TC) (using neuropsychological test scores), we devised 5 subgroups of MCI converters (MCI-C) based on their temporal distance to the conversion instant (0, 6, 12, 18 and 24 months before conversion). Next, we used the FDG-PET images of these subgroups and trained classifiers to distinguish between the MCI-C at different stages and stable non-converters (MCI-NC). Our results show that MCI to AD conversion can be predicted as early as 24 months prior to conversion and that the discriminative power of the machine learning methods decreases with the increasing temporal distance to the TC, as expected. Our results also show that this decrease arises from a reduction in the information contained in the regions used for classification and by a decrease in the stability of the automatic selection procedure.
Published at: Computers in Biology and Medicine, Vol. 58, pp. 101-109, March, 2015