5% (Fig. 5B). This classifier performed significantly better than a random classifier (McNemar χ2 = 6.54, P < 0.05). Discussion The findings presented here constitute an initial attempt to apply fundamental concepts from IR to the AD problem set. Techniques borrowed from IR include (1) arrangement of PET scans in a vector space, with one dimension for each PET scan voxel, (2) refinement of queries by subtraction Inhibitors,research,lifescience,medical of orthogonal vectors (a technique used to implement a logical NOT operation for search
engines—see Widdows 2004; Widdows and Peters 2003), and (3) scoring of PET scan “relevance” to a diagnostic query by means of cosine similarity between vectors. Cosine similarity scores derived in this manner are useful for constructing classifiers that differentiate NC Pazopanib solubility subjects from AD subjects, as well as MCI patients who are destined to convert to AD within 2 years from those who are not. Furthermore, both types of cosine similarity scores derived here make independent contributions to Inhibitors,research,lifescience,medical variance in follow-up FAQ scores that supersede
the contribution of diagnostic group, suggesting that this method may be useful for making more precise prognostications regarding the functional status of selleck AZD9291 individuals. The validity of the method is given further support by the Inhibitors,research,lifescience,medical fact that the residual vectors bear a Inhibitors,research,lifescience,medical topographic resemblance to maps of the default mode network. The method is computationally simple, at least relative to many techniques commonly run on modern computers. Ordinary least squares regression (the first step for computing the residual vectors) is a common approach to finding approximate Inhibitors,research,lifescience,medical solutions to many problems in statistics and engineering. Accordingly, algorithms for regression are fast and implementations are convenient. In MATLAB®, the regression step takes only one line of code and usually runs in less than
1 sec, even with large matrices. Classifiers built from structural MRI data that discern between controls and AD patients have similar accuracy to the ones presented here, but are much more computationally intensive, sometimes requiring Entinostat more than 1 week to build the classifier and hours to test it (Cuingnet et al. 2010). The method presented here compares favorably with other methods. Classifiers built from structural MRI data alone perform well when differentiating between patients with AD and subjects with normal cognition (up to 81% sensitivity with 95% specificity for voxel-based methods) (Cuingnet et al. 2010). Some studies have reported comparable accuracy with MRI methods for predicting conversion from MCI to AD, but sample sizes have been small and lack of cross-validation may mean that the results will not generalize to other samples (Convit et al. 2000).