The Swedish BioFINDER study was used for replicating the main results from ADNI and for the functional connectivity analyses. BioFINDER is a prospective study that focuses on identifying key mechanisms and improvement of diagnostics in AD and other neurodegenerative disorders. For details about study design, methods, and specific inclusion/exclusion criteria, see http://biofinder.se . The study was approved by the ethical review board in Lund, Sweden, and all participants gave their written informed consent. All non-demented individuals with CSF and Aβ PET data were selected for this study. This resulted in a cohort consisting of cognitively healthy elderly subjects (n = 138) and consecutively recruited patients who had been referred to memory clinics due to cognitive complaints (n = 268). The PET ligand 18F-flutemetamol was used for Aβ PET, and images were acquired 90–110 min post-injection. The PET scanning procedures have been described previously60 (link). CSF Aβ42 and P-tau were analyzed with INNOTEST ELISAs (Fujirebio Europe, Ghent, Belgium) and T-tau with EUROIMMUN ELISAs (EUROIMMUN AG, Lübeck, Germany) as previously described60 (link), 62 (link). Mixture modeling was performed in the sample to determine the cut-offs for abnormal CSF Aβ42 (<517 ng/L) and abnormal Aβ PET (>0.759 SUVR). Only baseline data was available for analysis in the BioFINDER cohort.
Imaging was performed on a 3 Tesla Siemens Tim Trio scanner (Siemens Medical Solutions, Erlangen, Germany). The high-resolution 3D T1-weighted volume used for segmentation and normalization was acquired using an MPRAGE sequence (in-plane resolution = 1 × 1 mm2, slice thickness = 1.2 mm, TR = 1950 ms, TE = 3.37 ms, flip-angle = 9°). Spontaneous BOLD oscillations in the absence of external stimuli were imaged with a gradient-echo planar sequence (eyes closed, in-plane resolution = 3 × 3 mm2, slice thickness = 3mm, TR = 2000 ms, TE = 30 ms, flip-angle = 90°, 180 dynamic scans, 6 min).
Resting-state data preprocessing was performed with a pipeline composed of FSL63 (link), AFNI64 (link), and ANTS65 (link). Anatomical processing involved skull stripping, segmentation of white matter (WM)/GM/CSF and normalization to MNI152-space53 . Dropping the first five frames in anticipation of steady state, functional data was bulk motion, and slice timing corrected, furthermore nuisance regressed using the WM/CSF average signal, 6 components of physiological noise66 (link), 24 motion parameters67 (link), and linear/quadratic trends. Finally, the functional data was transformed to MNI space. Frames causing outliers in total frame-to-frame signal variation were censored, on average constituting 4% of the fMRI series68 (link). The signal was band-pass filtered to 0.01–0.1 Hz, further discriminating against scanner drift and physiological noise. No spatial smoothing was applied.
Subjects with a mean/maximum frame-wise displacement29 (link), 69 (link) exceeding 0.6/3.0 mm were excluded. As an extra precautionary step, the voxel-to-voxel BOLD-signal correlations across the whole brain (including GM, WM, and CSF) was calculated and summed. Outliers in this measure likely originate in a motion-induced global signal confound capable of eluding conventional motion estimation70 (link) and were removed.
The processed fMRI data was resampled using trilinear interpolation to 6 × 6 × 6 mm3 resolution and masked with GM derived from a cortical resting-state network atlas71 (link) and Harvard-Oxford subcortical atlas72 (link). Fisher-z transformed Pearson correlation between the resulting 5071 GM voxel time series then yielded a measure of functional connectivity (FC), corresponding to a weighted graph with nodes (voxels) and links (voxel BOLD time series correlations).
Network components correlating with CSF Aβ42 were calculated using a method similar to the NBS algorithm73 (link). We calculated the largest network component C (defined as a connected set of links), for which rij > r0, where rij is the Spearman correlation over all subjects between CSF Aβ42 and zij for the link between voxel i and j (Fisher z-transformed voxel BOLD time-series correlation), and r0 controls the component size and significance level of constituent links. We chose r0 corresponding to approximately p = 0.001 given the number of subjects in the calculation. Component size was then defined as the sum of all multiplied by the Spearman correlation of the sum of all z-values in the component and CSF levels. This last step allows for reasonable and natural variations on the extracted component. The component size was compared to a permutation-generated null distribution of sizes, thus controlling for the family-wise error rate in the weak sense at α = 0.05 The result of the algorithm is a network component on which sum of z-scores correlates significantly higher than for randomized sets of subject FC-CSF pairs. Age, gender and APOE ε4 status was controlled for by partial correlation.
In order to simplify the analysis of network components, we grouped nodes using a resting-state network atlas71 (link) containing: default mode, dorsal and ventral attention, sensory motor, visual, fronto-parietal, and fronto-temporal (medial temporal lobe/orbitofrontal cortex). To this set of labels we added two anatomically defined subcortical structures from the Harvard-Oxford atlas72 (link): the Basal Ganglia (BG: thalamus, caudate, putamen and pallidum) and hippocampus/amygdala (HI). Note that the permutation-based approach generates p values for the network component as a whole, but since these are too large and complex to visualize, a network-based break up is necessary.
Only Aβ PET negative subjects were used in the connectivity analysis. In addition to the previously described group of early Aβ accumulators (CSF+/PET−, n = 23 after fMRI quality control) we also defined a group of biomarker negative subjects with indications of very early Aβ accumulation (n = 80). Those with low levels of CSF Aβ42, but still within the normal range, have a high risk of becoming abnormal within the next couple of years13 (link), which suggests that sub-threshold CSF Aβ42 levels indicate very early Aβ accumulation. This group was characterized as Aβ PET negative with CSF Aβ42 between 517–750 μg/mL (CSF−low/PET−).
Imaging was performed on a 3 Tesla Siemens Tim Trio scanner (Siemens Medical Solutions, Erlangen, Germany). The high-resolution 3D T1-weighted volume used for segmentation and normalization was acquired using an MPRAGE sequence (in-plane resolution = 1 × 1 mm2, slice thickness = 1.2 mm, TR = 1950 ms, TE = 3.37 ms, flip-angle = 9°). Spontaneous BOLD oscillations in the absence of external stimuli were imaged with a gradient-echo planar sequence (eyes closed, in-plane resolution = 3 × 3 mm2, slice thickness = 3mm, TR = 2000 ms, TE = 30 ms, flip-angle = 90°, 180 dynamic scans, 6 min).
Resting-state data preprocessing was performed with a pipeline composed of FSL63 (link), AFNI64 (link), and ANTS65 (link). Anatomical processing involved skull stripping, segmentation of white matter (WM)/GM/CSF and normalization to MNI152-space53 . Dropping the first five frames in anticipation of steady state, functional data was bulk motion, and slice timing corrected, furthermore nuisance regressed using the WM/CSF average signal, 6 components of physiological noise66 (link), 24 motion parameters67 (link), and linear/quadratic trends. Finally, the functional data was transformed to MNI space. Frames causing outliers in total frame-to-frame signal variation were censored, on average constituting 4% of the fMRI series68 (link). The signal was band-pass filtered to 0.01–0.1 Hz, further discriminating against scanner drift and physiological noise. No spatial smoothing was applied.
Subjects with a mean/maximum frame-wise displacement29 (link), 69 (link) exceeding 0.6/3.0 mm were excluded. As an extra precautionary step, the voxel-to-voxel BOLD-signal correlations across the whole brain (including GM, WM, and CSF) was calculated and summed. Outliers in this measure likely originate in a motion-induced global signal confound capable of eluding conventional motion estimation70 (link) and were removed.
The processed fMRI data was resampled using trilinear interpolation to 6 × 6 × 6 mm3 resolution and masked with GM derived from a cortical resting-state network atlas71 (link) and Harvard-Oxford subcortical atlas72 (link). Fisher-z transformed Pearson correlation between the resulting 5071 GM voxel time series then yielded a measure of functional connectivity (FC), corresponding to a weighted graph with nodes (voxels) and links (voxel BOLD time series correlations).
Network components correlating with CSF Aβ42 were calculated using a method similar to the NBS algorithm73 (link). We calculated the largest network component C (defined as a connected set of links), for which rij > r0, where rij is the Spearman correlation over all subjects between CSF Aβ42 and zij for the link between voxel i and j (Fisher z-transformed voxel BOLD time-series correlation), and r0 controls the component size and significance level of constituent links. We chose r0 corresponding to approximately p = 0.001 given the number of subjects in the calculation. Component size was then defined as the sum of all multiplied by the Spearman correlation of the sum of all z-values in the component and CSF levels. This last step allows for reasonable and natural variations on the extracted component. The component size was compared to a permutation-generated null distribution of sizes, thus controlling for the family-wise error rate in the weak sense at α = 0.05 The result of the algorithm is a network component on which sum of z-scores correlates significantly higher than for randomized sets of subject FC-CSF pairs. Age, gender and APOE ε4 status was controlled for by partial correlation.
In order to simplify the analysis of network components, we grouped nodes using a resting-state network atlas71 (link) containing: default mode, dorsal and ventral attention, sensory motor, visual, fronto-parietal, and fronto-temporal (medial temporal lobe/orbitofrontal cortex). To this set of labels we added two anatomically defined subcortical structures from the Harvard-Oxford atlas72 (link): the Basal Ganglia (BG: thalamus, caudate, putamen and pallidum) and hippocampus/amygdala (HI). Note that the permutation-based approach generates p values for the network component as a whole, but since these are too large and complex to visualize, a network-based break up is necessary.
Only Aβ PET negative subjects were used in the connectivity analysis. In addition to the previously described group of early Aβ accumulators (CSF+/PET−, n = 23 after fMRI quality control) we also defined a group of biomarker negative subjects with indications of very early Aβ accumulation (n = 80). Those with low levels of CSF Aβ42, but still within the normal range, have a high risk of becoming abnormal within the next couple of years13 (link), which suggests that sub-threshold CSF Aβ42 levels indicate very early Aβ accumulation. This group was characterized as Aβ PET negative with CSF Aβ42 between 517–750 μg/mL (CSF−low/PET−).
Free full text: Click here
