Alzheimer’s disease (AD) is the most common form of dementia in the elderly characterized by memory loss with pathological changes originated in the medial temporal lobes (MTL). Since the treatment of AD is only symptomatic, early intervention may be more effective to prevent or slow the disease progression. Amnestic mild cognitive impairment (aMCI) is a clinical predictor of AD [1], however its diagnosis is complicated and time-consuming. With the availability of novel imaging tools, such as fMRI, image markers for a quick and objective diagnosis are now on the horizon. In this study we investigate the classification of aMCI and normal control subjects using memory-related fMRI activation patterns in subregions of the MTL [2-3], which are CA1, CA23DG (CA2/CA3/dentate gyrus), Subiculum (SUB), Entorhinal cortex (ERC), Perirhinal cortex (PRC), Fusiform gyrus (FUS), and parahippocampal cortex (PHC) in both hemispheres of the brain (Fig.1).

Comprehensive neuropsychological and clinical tests were administered to recruit and screen subjects. Sixteen right-handed subjects (8 aMCIs and 8 normal controls, matched in gender, age and education) participated in this study, which was approved by the institutional review board. Three memory paradigms involving encoding and recognition tasks were performed. These memory paradigms are as follows: 1) outdoor pictures (“Pict”), 2) faces-and-occupations (“Face”), and 3) unrelated noun word pairs (“Word”). Echo planar imaging and high-resolution coplanar T2-weighted structural imaging were performed in a 3.0T GE MRI scanner (slices perpendicular to the long axis of the hippocampus) for each subject for identifying fMRI activations in each subregion. Activation maps were generated for contrasts Encoding-Control (E-C), Recognition-Control (R-C), Encoding-Recognition (E-R), New-Control (N-C), Old-Control (O-C), New-Old (N-O). Each contrast could be positive (+) or negative (-) that resulted in 12 contrasts for each paradigm. The subregions of MTL were manually segmented using the T2 image (Fig. 1) and activation volumes above threshold (uncorrected p=0.001) were calculated for each subregion, contrast and paradigm. More details of subjects, paradigms and imaging protocols can be found in [2].

Since there were 504 features (14 subregions x 36 contrasts), the feature selection was conducted using both principle component analysis (PCA) and least absolute shrinkage and selection operator (LASSO). PCA transforms the original feature vectors into linearly uncorrelated vectors in a descending order of variances. LASSO is a linear regression method with L1 norm constraint (penalty). The feature selection is achieved by shrinking the coefficient for some features to zero with large constraint weighting. LASSO does not transform the original feature space, thus it is better suited for interrogation of contrasts and subregions, which are more powerful for the classification of aMCI and normal control subjects. Following the feature selection, four classification methods [4] were used for the classification of aMCI subjects: linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), support vector machine (SVM) and radial basis function network (RBFN). For interpretability of the results, only two most dominant features were used for classification.

The decision boundaries of four classification methods using the first two eigenvectors of PCA are shown in Fig. 2. As the two groups are not well separated in the first two eigenvector space, the classification accuracy is only for 62.50% LDA and 81.25% for QDA, SVM and RBFN. These two eigenvectors only account for 44% of the total variance and at least nine principal components have to be included to account for more than 90% total variance. In Fig. 3, two features (rFus for Face E-C+ and lCA23DG for Face E-C+) are selected using LASSO with a stringent constraint. The two groups are better separated using LASSO features than PCA components. Both SVM and RBFN achieved the perfect classification, while LDA and QDA missed one aMCI and one normal control. To further identify contrasts and subregions that are potentially important to discriminate two groups, the LASSO constraint was relaxed and resulted in ten features as shown in Table 1. The last column is the relative weights for each feature. (Note that lCA23DG for Face E-C+ was selected as the second feature with a large penalty, but is only ranked in the third place in this small penalty case.) In the table, the Face paradigm (6 out 10) is more dominant than the Pict (3 out of 10) and Word (1 out of 10) paradigms. Future studies for prediction of aMCI using memory activations will likely be more effective using the Face paradigm, which combines two memory domains. In summary, LASSO feature selection combined with non-linear classifier (SVM and RBFN) shows good classification performance in this study.