Neural network approaches and their reproducibility in the study of verbal working memory and Alzheimer's disease

Abstract

As clinical and cognitive neurosciences mature, the need for sophisticated neuroimaging analysis becomes more apparent. Multivariate analysis techniques have recently received increasing attention because they have attractive features that cannot be easily realized by the more commonly used univariate, voxel-wise, techniques. Multivariate approaches evaluate correlation/covariance of activation across brain regions, rather than proceeding on a voxel-by-voxel basis. Thus, their results can be more easily interpreted as a signature of neural networks. Univariate approaches, in contrast, cannot directly address functional connectivity in the brain. Apart from this conceptual difference, the covariance approach can also result in greater statistical power when compared with univariate techniques, which are forced to employ very stringent, and often overly conservative, corrections for voxel-wise multiple comparisons. Multivariate techniques also lend themselves much better to prospective application of results from the analysis of one dataset to entirely new datasets. We provide two examples that illustrate different uses of multivariate techniques in cognitive and clinical neuroscience. We hope this contribution helps facilitate wider dissemination of these techniques in the research community.

Figure 1

Schematic sketch of delayed-match-to-sample paradigm of the Sternberg variant.

Figure 2

Summary of the characteristics relating to subject expression across memory-load levels (1/3/6/letters) and topographic composition of the load-related covariance pattern, isolated from the maintenance phase of the DMS task in young people. Left: subject expression of the pattern in the elderly replication sample. Monotonically increasing expression is apparent on a subject-by-subject basis (rm-F=24.52, p<0.0001). Middle: subject expression of the covariance pattern in the original derivation sample of young subjects. The ordinal trend is also clearly visible here (rm-F=38.66, p<0.001). Right: Surface rendering of brain regions in the activation pattern, as ascertained by the bootstrap resampling procedure (p<0.01). Red color denotes brain regions whose activation associated with the pattern increases as a function of memory load; green color denotes regions whose activation decreases.

Figure 3

Voxel-wise univariate results from the data of the maintenance phase of the derivation sample of young subjects, conducted in parallel to the multivariate findings. Left: mean trend of load-related activation across memory load. Middle: correlation of the 6–1 differences in activation with NARTIQ. In both maps, red color denotes positive values and green color denotes negative values. Right: correlation of 6-1 difference in activation with corresponding difference in RT, only a negative correlation was found. All maps were thresholded at uncorrected p<0.001. There is no overlap between any two maps.

Figure 4

Summary of the characteristics relating to subject expression and topographic composition of the AD-related covariance pattern, obtained from the cross-sectional comparison of AD subjects and healthy controls. Left: subject expression of the pattern in all subjects, both from the replication as well as derivation samples. Middle: subject expression of the covariance pattern in the original derivation sample of young subjects. One can discern a mean increase in subject expression across disease severity, from healthy controls (CONTROLS), over minimally cognitively impaired (CI: CDR=0), mildly cognitively impaired (CI: CDR=0.5), to AD (CDR=1). Right: Surface rendering of brain regions in the covariance pattern, as ascertained by the bootstrap resampling procedure (p<0.01). Red color denotes brain regions whose associated relative rCBF is increased as a function of disease status; green color denotes regions whose relative rCBF is decreased.