Resting fMRI-based classification of amyloid positive and negative aMCI patients
Presented During: Poster Session
Monday, June 27, 2016: 12:45 PM - 02:45 PM
Monday, June 27, 2016: 12:45 PM - 2:45 PM
Monday, June 27 & Tuesday, June 28
Kwangsun Yoo1, Peter Lee1, Young-Beom Lee1, Duk L. Na2, Sang Won Seo2, Yong Jeong1
1KAIST, Daejeon, Korea, Republic of, 2Samsung Medical Center, Seoul, Korea, Republic of
There have been many studies investigating the effect of amyloid burden on cognitive functions and structural or functional MRIs in healthy or patients with AD. It would be the next step to use these findings to classify or predict the existence of amyloid deposition in the brain without using radioligands under the circumstances with limited availability of amyloid PET images. In addition, recent studies have performed the multi-spectrum or sub-frequency analysis of resting fMRI and showed the distinct characteristics among sub-frequency bands.
In this study, we investigated the feasibility of resting fMRI-based classification of aMCI patients as amyloid positive or negative using sub-frequency functional connectivity.
Forty-six patients with aMCI were recruited consecutively at the memory disorder clinic in the Department of Neurology at Samsung Medical Center in Seoul, South Korea between May 2009 and July 2011. Each participant underwent MR and PiB-PET scans, clinical interviews, neurological examinations, and comprehensive neuropsychological assessments. The same diagnosis criteria and MRI protocols were applied as our previous publications . Subjects with excessive head motion (>3mm translation or >3° rotation) or noise in MRI were excluded for the analysis. Finally, 26 PiB+ and 16 PiB- aMCI patients were included.
The first three fMRI volumes were discarded for the further preprocessing, then resting fMRI were preprocessed (including slice-timing correction, realignment, coregistration, normalization to MNI space, and smoothing with a 4mm FWHM isotropic Gaussian Kernel) using SPM12 and DPARSF4.0 in MATLAB R2014b. Smoothed fMRI were detrended and then filtered into the traditional resting low frequency band (0.008~0.08 Hz) as well as sub-frequency bands of 0.01~0.025, 0.025~0.04, 0.04~0.055 and 0.055~0.07 Hz based on previous studies.
Each frequency band connectivity was separately processed in this step. We first constructed connectivity matrix using Pearson's correlation among grey matter-masked 116 AAL regions. We then applied PCA on connectivity features to reduce the number of features. We trained linear support vector machine with the reduced features. Accuracy (test) for each frequency band connectivity was estimated in leave-one-out cross-validation way.
With 10 features from PCA, 61.9% classification accuracy was achieved using the whole 0.008~0.08 Hz band of connectivity. Classification accuracy based on sub-frequency band connectivities (0.01~0.025, 0.025~0.04, 0.04~0.055 and 0.055~0.07 Hz) were 71.4, 52.4, 76.2, and 57.1%, respectively.
With 15 features from PCA, the same 61.9% classification accuracy was achieved using the 0.008~0.08 Hz band of connectivity. Classification accuracy based on sub-frequency band connectivity (0.01~0.025, 0.025~0.04, 0.04~0.055 and 0.055~0.07 Hz) were 66.7, 71.4, 73.8, and 52.4%, respectively.
We demonstrated the feasibility of resting fMRI connectivity to classify the amyloid positive and negative aMCI patients. Our results suggest that resting functional connectivity from sub-frequency band may contain better information about existence of amyloid deposition in aMCI patients than the traditional entire low frequency band.
Disorders of the Nervous System:
Alzheimer's Disease and Other Dementias 1
Imaging Methods: BOLD fMRI 2
Modeling and Analysis Methods: fMR