Parkinson's disease-related network topographies characterized with resting state functional MRI

Abstract

Spatial covariance mapping can be used to identify and measure the activity of disease-related functional brain networks. While this approach has been widely used in the analysis of cerebral blood flow and metabolic PET scans, it is not clear whether it can be reliably applied to resting state functional MRI (rs-fMRI) data. In this study, we present a novel method based on independent component analysis (ICA) to characterize specific network topographies associated with Parkinson's disease (PD). Using rs-fMRI data from PD and healthy subjects, we used ICA with bootstrap resampling to identify a PD-related pattern that reliably discriminated the two groups. This topography, termed rs-MRI PD-related pattern (fPDRP), was similar to previously characterized disease-related patterns identified using metabolic PET imaging. Following pattern identification, we validated the fPDRP by computing its expression in rs-fMRI testing data on a prospective case basis. Indeed, significant increases in fPDRP expression were found in separate sets of PD and control subjects. In addition to providing a similar degree of group separation as PET, fPDRP values correlated with motor disability and declined toward normal with levodopa administration. Finally, we used this approach in conjunction with neuropsychological performance measures to identify a separate PD cognition-related pattern in the patients. This pattern, termed rs-fMRI PD cognition-related pattern (fPDCP), was topographically similar to its PET-derived counterpart. Subject scores for the fPDCP correlated with executive function in both training and testing data. These findings suggest that ICA can be used in conjunction with bootstrap resampling to identify and validate stable disease-related network topographies in rs-fMRI. Hum Brain Mapp 38:617-630, 2017. © 2016 Wiley Periodicals, Inc.

Keywords: Parkinson's disease; biomarker; brain network; independent component analysis; resting state functional MRI; spatial covariance analysis.

Conflict of interest statement

All other authors declare no competing financial interests.

© 2016 Wiley Periodicals, Inc.

Figures

Figure 1

Identification and forward application of disease‐related patterns using resting state fMRI. (A) rs‐fMRI scan data are analyzed with spatial group ICA to generate group IC maps. Dual regression is then used to estimate spatial maps and temporal dynamics for individual subjects. Based on the group and subject maps, subject scores are computed, which represent the expression of each IC in a given subject's scan data. The ICs that best separate control and disease groups are selected by bootstrap resampling and logistic regression analysis. The disease‐related pattern is determined as a linear combination of the selected group ICs; coefficients (region weights) on each IC are estimated through a second bootstrap procedure applied to the corresponding subject scores (see text). (B) For forward application in rs‐fMRI scan data from prospective single cases, dual regression is used to estimate individual spatial maps for new subjects using the group IC maps identified in the pattern derivation procedure. Subject scores for the disease‐related pattern are computed by linear combination of expression values for the selected ICs, using the coefficient defined in the derivation step.

Figure 2

Network selection and parameter estimation. (A) Frequency histogram of ICs that best discriminate between the patient and control groups according to the bootstrap resampling procedure (1,000 iterations). ICs that are not selected or with frequency = 0 are not shown. (B) The plots show centered finite difference approximations to the first (dark blue) and second (red) derivative of the frequency histogram. The first derivative is minimal at IC3 and the second derivative changes sign between IC3 and IC28. Based on this inflection point, we selected three ICs (IC9, IC14, and IC3) for further analysis. (C) Frequency histograms of the estimated coefficients for IC9, IC14, and IC3 according to bootstrap resampling (1,000 iterations). For each iteration, model coefficient was estimated for which the associated subject scores best discriminated between patients and control subjects. The histogram provides an estimate of distribution of the model coefficients. In this case, the estimated mean values for the relevant ICs (0.7432 [IC9], 0.5393 [IC14], and 0.3959 [IC3]) were used to define the disease‐related pattern. The same coefficients were used prospectively to compute corresponding expression values (scores) for the disease‐related pattern in individual subjects.

Figure 3

PDRP identified with rs‐fMRI. (A) PDRP identified in rs‐fMRI (fPDRP, left) and PET (pPDRP, right) are shown on the MNI 152 template. fPDRP, derived from 20 normal controls and 20 PD patients, is characterized by increased activity in the basal ganglia, thalamus, cerebellum/pons, anterior cingulate cortex (ACC), and supplementary motor area (SMA). The major network regions that defined the fPDRP corresponded closely to the metabolically active (red areas) regional counterparts of the pPDRP topography. [The color stripes show Z‐values thresholded at ±0.5. Activity increases (fPDRP) or relative metabolic increases (pPDRP) are displayed in red; relative metabolic decreases (pPDRP) are displayed in blue.] (B) Expression scores for fPDRP and pPDRP are increased in the PD patients compared to normal controls (NL) (P < 0.001; Student's t‐test). [Error bars represent standard errors of the means.] (C) fPDRP subject scores correlated with UPDRS ratings for akinesia‐rigidity (r = 0.61, P < 0.005, circles) in the PD subjects scanned at North Shore University Hospital; tremor ratings measured in the same subjects exhibited only a marginal relationship with network expression values (r = 0.39, P = 0.09, triangles).

Figure 4

Discrimination of PD from healthy subjects based on fPDRP expression values. fPDRP expression was increased in PD patients relative to normal (NL) subjects in the North Shore University Hospital (NS) training sample (left, P < 0.001). The difference is also significant in the NS (middle, P = 0.008) and UF‐G (right, P = 0.026) testing samples. [Error bars represent standard errors of the means.]

Figure 5

Effect of levodopa administration on fPDRP network expression. (A) fPDRP expression was abnormally elevated in rs‐fMRI scans from eight PD subjects studied in the unmedicated off‐state (OFF). Network activity declined (P = 0.017; paired Student's t‐test) when these subjects were re‐scanned in the on‐state (ON) approximately 1 h following their usual dose of dopaminergic medication. [Error bars represent standard errors of the means.] (B) The baseline fPDRP expression scores off medication significantly correlated with the levodopa‐mediated changes in fPDRP expression (ΔfPDRP[OFF – ON]) (r = 0.66, P < 0.05; Pearson's correlation).

Figure 6

PDCP identified with rs‐fMRI. (A) PDCP topographies identified in rs‐fMRI (fPDCP, top) and PET (pPDCP, bottom) data are shown on the MNI 152 template. fPDCP, derived from 19 PD patients, is characterized by a negative correlation between pattern expression and performance on the California Verbal Learning Test (CVLTsum). [The color stripes show Z‐values thresholded at ±1.0. Relative metabolic increases (pPDCP) are displayed in red; negative correlation (fPDCP) and relative metabolic decreases (pPDCP) are displayed in blue.] (B) fPDCP expression values correlated significantly with performance on CVLTsum, a measure of memory and executive function, in the training data (open circles, solid line; r = −0.69, P < 0.001) and in the within‐subject testing data (filled circles, dashed line; r = −0.52, P < 0.02; Pearson's correlations).