Earliest Amyloid and Tau Deposition Modulate the Influence of Limbic Networks During Closed-loop Hippocampal Downregulation

Stavros Skouras; Jordi Torner; Patrik Andersson; Yury Koush; Carles Falcon; Carolina Minguillon; Karine Fauria; Francesc Alpiste; Kaj Blenow; Henrik Zetterberg; Juan D. Gispert; José L. Molinuevo

Disclosures

Brain. 2020;143(3):976-992. 

In This Article

Materials and Methods

Participants

Participants comprised 48 adult volunteers with complete data [mean age = 62.705 years; standard deviation (SD) = 4.628] from the ongoing ALFA (Alzheimer's and Families) project and related substudies, many of them descendants of patients with Alzheimer's disease and APOE ε4 allele carriers, hence presenting increased risk for Alzheimer's disease (Molinuevo et al., 2016). All data were acquired at the Barcelonaβeta Brain Research Center, from 2016–18. Researchers were blind to biomarker levels and cognitive performance during data acquisition. All participants were highly functional and without neurological or psychiatric history at the time of scanning, as assessed in the medical baseline session of their ongoing ALFA-related study within the previous 12 months (time between baseline ALFA and rt-fMRI scanning: mean = 6.4 months, SD = 2.5). The cognitively unimpaired status of participants was also assessed as an inclusion criterion in the cognitive testing session of their ongoing ALFA-related study within the previous 12 months (mean = 6.4 months, SD = 2.5), according to their scores on the Clinical Dementia Rating scale (CDR = 0; Hughes et al., 1982; Morris, 1993) and the Mini-Mental State Examination (MMSE ≥ 27; Folstein et al., 1975; Blesa et al., 2001). During that cognitive testing session, participants also completed the Free and Cued Selective Reminding Test (FCSRT; Grober et al., 2009), the Subjective Cognitive Decline Questionnaire (SCD-Q; Rami et al., 2014) and the Visual Puzzles test included in the Wechsler Adult Intelligence Scale IV (WAIS IV; Weschler, 2013); all scores are summarized in Table 1. The FCSRT is an internationally standardized research instrument recommended for the assessment of episodic memory and diagnosis of prodromal Alzheimer's disease (Dubois et al., 2007, 2014). The SCD-Q is a useful tool to measure self-perceived cognitive decline, independently of a participant's objective neuropsychological assessment. According to their SCD-Q responses, four participants had a subjective impression of cognitive or memory decline, while their cognition was preserved based on objective cognitive testing (see 'Discussion' section). The Visual Puzzles test measures non-verbal reasoning and the ability to analyse and synthesize abstract visual stimuli. It is related to visual perception, general visuospatial intelligence, fluid intelligence, simultaneous processing, visualization and spatial manipulation, as well as the ability to anticipate relationships between constituent parts of a whole. Within a given time limit (either 20 or 30 s, depending on the complexity of a puzzle), the participant was instructed to select the three pieces necessary for the reconstruction of the presented puzzle. Each item is scored with 0 or 1 and the testing is suspended after three consecutive scores of 0. The test consists of 26 items and the estimated administration time is ~10 min. At inclusion of participants in the ALFA project, they had also completed the Cognitive Reserve Questionnaire (CRQ; Supplementary Table 1); a questionnaire comprising eight questions whose total score serves as a proxy for cognitive reserve (Rami et al., 2011). To assess core Alzheimer's disease biomarker levels in CSF, participants underwent lumbar puncture within the preceding 10 months (time between lumbar puncture and rt-fMRI scanning: mean = 5.23 months, SD = 2.49). Eighteen participants presented normal CSF biomarkers (amyloid-β42 levels > 1098 pg/ml and p-tau levels < 19.2 pg/ml), 22 participants presented CSF amyloid-β42 levels < 1098 pg/ml, 15 participants presented CSF p-tau levels > 19.2 pg/ml and seven of the latter met both criteria. Participants did not differ significantly with regards to any cognitive or neuroimaging variables, regardless of grouping based on CSF amyloid-β42 and p-tau levels (Table 1). Thorough quality control was applied in advance, to exclude five datasets with >10% invalid functional volumes due to movement, as well as seven datasets with rt-fMRI acquisitions during which technological complications occurred and six datasets with rt-fMRI acquisitions during which fatigue, discomfort or sleep had occurred. The local ethics committee 'CEIC-Parc de Salut Mar' reviewed and approved the study protocol and informed consent form, in accordance with current legislation.

CSF Sampling

CSF was collected by lumbar puncture between 9 and 12 am in polypropylene tubes. Samples were processed within 1 h and centrifuged at 4°C for 10 min at 2000g, stored in polypropylene tubes and frozen at −80°C. Core Alzheimer's disease CSF biomarkers (namely amyloid-β42 and p-tau) were determined using cobas Elecsys® assays (Hansson et al., 2018).

APOE Genotyping

Proteinase K digestion and subsequent alcohol precipitation was used to obtain DNA from the blood cellular fraction. Samples were genotyped for two single nucleotide polymorphisms (rs429358 and rs7412) and the number of APOE ε4 alleles was determined for each participant (Molinuevo et al., 2016). Results are displayed in Table 2.

Image Acquisition

All scanning was performed in a single 3 T Philips Ingenia CX MRI scanner (2015 model). Pre-neurofeedback scanning comprised a 3D T1-weighted sequence of 240 sagittal slices with voxel resolution = 0.75 × 0.75 × 0.75 mm3, repetition time = 9.90 ms, echo time = 4.60 ms, flip angle = 8; and a diffusion-weighted sequence of 66 axial slices with voxel resolution = 2.05 × 2.05 × 2.20 mm3, repetition time = 9000 ms, echo time = 90 ms, flip angle = 90, featuring 72 non-collinear directions (b = 1300 s/mm2) and one non-gradient volume (b = 0 s/mm2). During neurofeedback, echo planar imaging was used with voxel resolution = 3 × 3 × 3 mm3, repetition time = 3000 ms, echo time = 35 ms, matrix size = 80 × 80 voxels, field of view = 240 mm and interleaved slice acquisition with an interslice gap of 0.2 mm (45 slices, whole brain coverage). In total, 610 functional volumes were acquired.

Structural Image Processing

The standard FreeSurfer pipeline was applied on T1-weighted images to estimate each participant's normalized hippocampal brain volume, as a proxy of brain reserve, by dividing total hippocampal volume by total intracranial volume (Cavedo et al., 2012). Using the Advanced Normalization Tools (Avants et al., 2009) (ANTs v2.x; RRID: SCR_004757), the N4 non-parametric non-uniform intensity normalization bias correction function (Tustison et al., 2010; Tustison and Avants, 2013) was applied on all T1 images, followed by an optimized blockwise non-local means denoising filter (Coupé et al., 2008). Multi-atlas segmentation with joint label fusion (Wang et al., 2013) segmented hippocampal subfields to derive probabilistic maps for CA1, which were thresholded at P = 0.9 to create the neurofeedback target region of interest masks.

Experimental Paradigm

The design of the VR paradigm was guided by the following objectives: (i) we used VR to make the task immersive, engaging and entertaining, which in turn; (ii) enabled us to extend the task to 30 min to maximize first-level statistical power; (iii) we used sliding-window closed-loop neurofeedback to achieve an optimal design for analysis of functional connectomics; and (iv) we narrowed the neurofeedback target region of interest to hippocampal subfield CA1, which presents atrophy in Alzheimer's disease but not in healthy ageing (Wilson et al., 2004; Frisoni et al., 2008; Yushkevich et al., 2015), while also being consistently implicated in memory encoding (Kim, 2011) and differing in activity in patients (Schwindt and Black, 2009). Prior to the acquisition of CSF biomarkers, the task was validated in a larger partly independent sample (77% overlap), by showing that hippocampal downregulation was associated with genetic predisposition to Alzheimer's disease, neurodevelopmental processes and bilateral cohesion of hippocampal function (Skouras et al., 2019b).

The VR environment had been developed in the game-engine Unity (Unity Technologies ApS, San Francisco, USA) (Figure 1). During a 30-min long scanning session, participants were immersed in the VR environment and could run in a fixed, circular path. The experimental task was to explore different mental strategies, aiming to achieve the maximum velocity and to traverse the maximum distance possible, while at the same time attending to features of the VR environment, remembering them and considering whether they changed between each lap of the environment. Debriefing interviews revealed that most participants tried similar simple strategies, such as internal self-motivational speech, imagining that they were on a speeding vehicle, or focusing their sight on a specific point in the field of view. The first 30 functional volumes of the session served to establish a baseline of hippocampal activity, for each participant. Subsequently, with every repetition time, the most recent shift in hippocampal CA1 activity was compared to reference measures of change derived from the preceding 90 s (30 functional volumes; see 'Online functional image processing and neurofeedback' section). A 5% decrease of VR velocity was triggered by increases in hippocampal activity and a 5% increase of velocity was triggered by decreases in hippocampal activity. At any given moment, the current velocity was displayed as a percentage of the maximum velocity possible and a green or red signal was superimposed on a coronal brain view, reflecting the direction of the most recent change (Figure 1). Every 30 volumes, the velocity was reset to 50%, resulting in 19 trials per session. The trials were not related to the experimental design from a statistical perspective, because computing EC requires hundreds of continuous functional volumes; however, having distinct trials served three important purposes. First, by resetting the speed every 90 s, all participants had a similar gamified experience, regardless of their actual performance. This was important from an ethical perspective because participants were aware that the gamified task, even though experimental, could be reflecting aspects of their brain function and health. It was therefore important that all participants would exit from the scanning session with an overall positive experience (and without any reason to worry) and this was also a secondary motivation to implement an online analysis pipeline with adaptive difficulty, based on a sliding-window baseline (see 'Online functional image processing and neurofeedback' section). Second, by resetting the speed and adapting the difficulty of the task, all participants had a similar overall sensory and perceptual experience. This was important to control for the effects of potential confounding factors such as individual differences in visual cortex activity during the scanning sessions. Third, resetting the speed enabled the developed paradigm to utilize effectively the basic principles of operant conditioning and sensory-aided learning. Because in the context of the gamified goal accelerations are perceived to be of positive valence and decelerations are perceived to be of negative valence, the 570 neurofeedback signals that occurred during each 30-min scanning session acted as positive and negative reinforcers, respectively. Thus, resetting the speed at regular intervals was necessary to avoid ceiling or floor effects that could compromise the number of reinforcers produced in a particular session.

Figure 1.

VR environment and neuroimaging pipeline. The VR neurofeedback paradigm developed for the study used principles of passive, sensory-aided, operant conditioning and featured 570 neurofeedback signals per session. To maintain a balanced perceptual experience across participants, task difficulty adapted to individual performance dynamically, aiming to drive CA1 activity in each participant to the minimum possible. Using previously acquired anatomical images, multi-atlas hippocampal subfield segmentation localized and segmented hippocampal subfield CA1, prior to real-time scanning. With every real-time functional volume, moment-to-moment changes in hippocampal CA1 activation effected inverse changes of velocity in VR. Offline statistical modelling was used to derive a measure of neurofeedback regulation performance and to perform EC mapping (i.e. to estimate how much influence each voxel exerts during hippocampal CA1 downregulation with neurofeedback). APOE = apolipoprotein genotype; Aβ42 = amyloid-β42; BrainRes = brain reserve; CA1 = cornu ammonis 1; CogRes = cognitive reserve; fMRI = functional MRI; GLM = general linear model; NF = neurofeedback; p-tau = phosphorylated tau; ROI = region of interest; TIV = total intracranial volume.

Online Functional Image Processing and Neurofeedback

In each rt-fMRI session, EPI (echo planar imaging) images of whole-brain activity were acquired, reconstructed and exported every 3 s. The first 10 functional volumes were discarded to allow for gradient and tissue excitation levels to stabilize (Soares et al., 2016). Online movement correction through rigid-body registraperformedtion was relative to an initial reference volume. To remove low frequency drifts in the functional MRI time series, temporal high-pass filltering with a cut-off frequency of 1/200 Hz (Tarvainen et al., 2002) was applied in real-time. An initial baseline mean and standard deviation were computed based on the first 30 processed functional volumes, during which the VR movement velocity remained constant at 0.5. Subsequently, voxel-efficiency weighting (Stoeckel et al., 2014) was performed by normalizing new images based on the mean and standard deviation of the preceding 30 observations in each voxel, according to Equation 1:

where, at time t, X t represents the filtered value of a single voxel, represents the mean of that voxel's time series within a time-window of 30 volumes up to time (t − 1), σ(t−30−1).(t−1) represents the corresponding standard deviation and ztzt represents the efficiency-weighted value of the voxel. The length of the sliding window was set to 30 as a trade-off between the speed of adaptation and the reliability of the mean estimates, in accordance with generic statistical guidelines and the central limit theorem (Hogg et al., 1977). For each acquired functional volume, an outcome variable was instantiated for the activity measured within each participant's hippocampal CA1 region of interest, in native space, based on the following computations. At any given time point t0, a voxel-wise expected data vector Y was computed according to Equation 2.

where Z(t) represents the voxel-wise observed data vector composed of the output of Equation 1 for all voxels within the region of interest at time t. A non-linear metric neurofeedback metric (NF) was computed for each new volume, in real-time, according to Equation 3.

With every repetition time, if the average expected signal surpassed the average observed signal in the hippocampal CA1 region of interest, the velocity v in the VR environment increased by 5% according to Equation 4; in the opposite case, the velocity decreased by 5% according to Equation 5, where v represents the VR movement velocity in the range {0.1, 1}.

Offline Functional Image Processing

Raw images were processed and examined using a well-controlled combination of highly specialized functions from open-source neuroimaging software. To maximize cortical segmentation accuracy, T1 images were subjected to the N4 non-parametric non-uniform intensity normalization bias correction function (Tustison et al., 2010, 2013) of the Advanced Normalization Tools (ANTs; version 2.x, committed in January 2016; Avants et al., 2009) and to an optimized blockwise non-local means denoising filter (Coupé et al., 2008). SPM12 (Statistical Parametric Mapping, RRID: SCR_007037) together with VBM8 (VBM toolbox, RRID: SCR_014196), were used to segment each anatomical image into grey matter, white matter and CSF. Whole-brain images with removed cranium were also derived and used for normalization. Neuroanatomically-plausible symmetric diffeomorphic matrices were computed to transform each subject's anatomical data to a custom template derived from the same population (Skouras et al., 2019a) and subsequently to MNI space (Avants et al., 2011; Tustison and Avants, 2013) as defined by the robust, high-definition ICBM (International Consortium for Brain Mapping) brain template (Fonov et al., 2011), also with removed cranium. All transformation matrices were concatenated and applied to the CA1 region of interest masks and functional MRI datasets in a single step to ensure optimal normalization while avoiding multiple interpolations, in accordance with best image processing practices.

Using SPM12 and the MIT connectivity toolbox (Connectivity Toolbox, RRID: SCR_009550), functional data were subjected to standard preprocessing, consisting of slice-time correction, realignment and reslicing of functional volumes, denoising via regression of average white matter timeseries, average CSF timeseries, 24 Volterra expansion movement parameters and scan-nulling regressors (Lemieux et al., 2007) produced by the Artifact Detection Tools (ART; RRID: SCR_005994). Each subject's functional data were masked by the grey matter of their equivalent anatomical datasets using FSL and then normalized to MNI space in accordance with their respective diffeomorphic matrices using ANTs. Temporal high-pass filtering with a cut-off frequency of 1/90 Hz and spatial smoothing using a 3D Gaussian kernel and a filter size of 6 mm full-width at half-maximum, were performed using the Leipzig Image Processing and Statistical Inference Algorithms (LIPSIA v2.2.7, 2011, RRID: SCR_009595; Lohmann et al., 2000).

EC was computed for every voxel, based on the LIPSIA implementation of ECM (vecm command), with previously published detailed mathematical specification (Lohmann et al., 2010). In effect, the ECM procedure begins by computing a correlation coefficient between every possible pair of voxels within the computational volume (e.g. grey matter mask), based on their time series. It then assigns to each voxel an initial centrality value that is computed as the sum of all correlation coefficients that exist between that voxel and all other voxels (a value equivalent to degree centrality). Subsequently, ECM updates the centrality value in each voxel iteratively, each time by a weighted sum of its correlation coefficients with other voxels, using weights that correspond to each paired voxel's centrality value from the previous iteration. The iterations repeat until the values stabilize and there is no further change from one iteration to the next. Importantly, we limited the similarity matrix to positive correlations between grey matter voxels, similarly to our previous Alzheimer's disease study with EC (Skouras et al., 2019a). EC values were Gaussianized voxel-wise across all subjects to enable second-level parametric inference, using a general transformation designed to optimize the reliability of arbitrary distributions (Albada and Robinson, 2007). The resulting whole-brain images entered second level statistical modelling and voxel-wise multiple regression analyses (see 'Statistical analysis' section).

Statistical Analysis

For standard type I and II error rates (α = 0.05, β = 0.20), a priori computations of required sample size were performed using statistical power estimation software (G*Power v 3.1; RRID: SCR_013726; Faul et al., 2007, 2009). The computations suggested that to enable sufficient statistical power for effects of medium size (f 2 > 0.15), using multiple regression with up to seven predictor variables, a minimum sample size of 43 subjects would be required. Pearson's r was used as a metric of similarity in first-level general linear model (GLM) to quantify neurofeedback performance per participant. A linear vector coding continuous downregulation, as the target performance for neurofeedback, was compared to neurofeedback moment-to-moment regulation, measured by the realigned average CA1 time series, to produce a measure of neurofeedback performance for each participant similar to a previous study (Skouras and Scharnowski, 2019). First, we investigated whether the unique variance associated with CSF amyloid-β42 levels was reflected in EC during the neurofeedback task. For consistency, CSF amyloid-β42 values were multiplied by −1 during modelling, to align results and colour maps with the direction of the pathophysiological continuum of Alzheimer's disease. CSF amyloid-β42 values were orthogonalized to potential confounding variables; specifically sex, age, number of APOE ε4 alleles, hippocampal volume, cognitive reserve and neurofeedback performance. This resulted in an orthogonal second-level design matrix that was used for GLM of the unique effect of amyloid-β42 on EC during VR neurofeedback. The resulting whole-brain z-map was corrected for multiple comparisons using 10 000 iterations of Monte Carlo simulations, with a pre-threshold of z > 2.326 (P < 0.01) and a corrected significance level of P < 0.05. Similarly, we investigated for a possible correlation between p-tau and EC, controlling for the same confounding variables. Finally, we investigated the effect of healthy ageing on EC during VR neurofeedback, by orthogonalizing age to all potential confounding factors, including the CSF p-tau by amyloid-β42 ratio (Maddalena et al., 2003). All second-level neuroimaging analyses featured data from the entire sample. All scale variables were normalized prior to orthogonalization via the recursive Gram-Schmidt orthogonalization of SPM12 (SPM, RRID: SCR_007037). For quality control, we confirmed that each variable of interest was completely orthogonal to all covariates in its respective design matrix, as well as to all original vector data, while remaining in high correlation with its own original vector. This ensured investigating meaningful observations and effects due to the unique variance in each variable of interest.

Code and Data Availability

The VR environment and the analysis software developed for the experiment, are publicly available as open-source software via the 'VR_multipurpose_v1.0' and the 'HippDownReg_exp1' repositories (GitHub; RRID: SCR_002630), respectively. Data used can been made available to researchers for non-commercial purposes, following agreement and approval by the Barcelonaβeta Brain Research Center's Data and Publications Committee.

processing....