Gray matter volume and estimated brain age gap are not linked with sleep‐disordered breathing

2.1 Participants

Subjects were drawn from the ADNI database (adni.loni.usc.edu) (Petersen et al., 2010) based on their cognitive status and the medical history regarding SDB, as suggested previously (Osorio et al., 2015). Diagnoses of MCI and AD were based on the ADNI criteria. Subjects with self‐reported “sleep apnea” or “obstructive sleep apnea” (or “OSA”) symptoms or receiving treatment with “Continuous Positive Airway Pressure” (or “CPAP”) or “bilevel positive airway pressure” (or “BiPAP”/“BPAP”) were defined as “positive SDB.” Two independent physicians reviewed medical history to confirm diagnosis and grouping the subjects. Demographic and clinical variables were extracted for all individuals, missing covariate data were assessed and multiple imputation method including variables of sex, age, cognitive status, body‐mass index (BMI), and education years (Zhang, 2016) was used for five participants with missing data‐points on BMI (three subjects), and education years (two subjects). Using the 1:1 propensity score matching method, we assembled six distinct sub‐groups according to their cognitive (HC, MCI, AD) and SDB (positive or negative) status. Covariates included in the matching were age, sex, years of education, BMI, cognitive status (HC, MCI, AD), presence of the Apolipoprotein E4 (APOE4) allele, history of SDB treatment (only when matching between those with SDB subjects), T1 imaging protocol, and field strength (Table 1). To efficiently match the study diagnosis groups for age, we have included both baseline and follow‐up ADNI images in the propensity‐score matching model. Only subjects that passed the quality assessment of the CAT (Computational Anatomy Toolbox), including weighted image quality rating based on the basic image properties, noise and geometric distortions, as well as checking homogeneity through the sample, were included for our analyses.

Open multimedia modal

2.2 Imaging acquisition and preprocessing

Participants had undergone a standardized protocol for high‐resolution MRI T1 scans of the brain as previously described (Clifford R. Jack et al., 2008). T1 imaging acquisition parameters were: TR = 2,400 ms, minimum full TE, TI = 1,000 ms, flip angle = 8°, 24 cm field of view, acquisition matrix of 192 × 192 × 166 and with 1.25 × 1.25 × 1.2 mm3 slice size. We used CAT12 (Computational Anatomy Toolbox) (Gaser & Dahnke, 2016) and SPM12 (Statistical Parametric Mapping,www.fil.ac.uk/spm) to perform voxel‐based morphometry (VBM). This included correcting the bias‐field distortions and noise removal, skull stripping, normalization to standard space and brain tissue segmentation into gray matter, white matter, and cerebrospinal fluid. Gray matter segments were modulated to represent actual gray matter volume. We then performed a biologically informed compression of the VBM data to the 673 gray matter parcels based existing in‐vivo brain parcellation (600 cortical gray‐matter parcels from Schäfer, 36 subcortical gray‐matter parcels from Brainnetome, and 37 cerebellar parcels from Buckner (Fan et al., 2016; Schaefer et al., 2018; Yeo et al., 2011). Thus, gray matter volume of each participant was represented by 673 features, each representing an individual parcel volume of that subject. The following analyses were performed on this data.

2.3 Statistical analysis of gray matter volume

Statistical analysis of gray matter volume of 673 parcels included three consecutive parts, suggested by Bludau and colleagues (Bludau et al., 2018); generating reference statistics, permuted statistics, and a family‐wise error (FWE) correction for multiple comparisons. Here we used an n‐way analysis of variance (ANOVA), to test the effect of age, cognitive status (HC, MCI, AD), SDB status, and SDB‐by‐cognitive status interaction, separately as independent variables (factors), on gray matter volume of each parcel as the dependent variable. The F values (per parcel) of this ANOVA were considered as the reference statistics. In the subsequent permutation statistics for each factor, we randomly shuffled the labels for that factor 10,000 times, replicated the analysis and recorded the F values to build a null‐distribution. The comparison of the reference statistic with this distribution then allows nonparametric inference per parcel and factor, yielding uncorrected p‐values. Importantly, however, we also recorded, per replication of the permutation, the highest statistics in the random data across the entire set (i.e., 673 brain regions), building a null‐distribution for FWE correction. The threshold corresponding to P_FWE < .05 was then provided by the (set‐wise maximum) value exceeded only in 5% of the replications.

2.4 Age prediction

Brain age was estimated from the atlas‐based representations of individual brain anatomy using a support vector machine (SVM) ensemble model. An independent (reference) large dataset consisting of 2089 (Figure 1a1) subjects (between 55 and 85 years old) was compiled from several large public and private datasets including 1000Brains (Caspers et al., 2014), Cambridge Centre for Aging and Neuroscience or Cam‐CAN (Shafto et al., 2014), OpenfMRI (Poldrack et al., 2013), Dallas Lifespan Brain Study or DLBS, Consortium for Reliability and Reproducibility or CoRR (Zuo et al., 2014), IXI, and Enhanced Nathan Kline Institute‐Rockland Sample or eNKI‐RS (Nooner et al., 2012). Given the imbalance between age brackets, sites, and sex, we performed a stratified subsampling, choosing the same number of men and women, as well as similar numbers across age‐brackets and a maximum of 30 subjects per age‐bracket and sex per site. The actual subjects sampled in each replication from the overall database were drawn from the pool independently at random without replacement. Each of these sampled sets was then used to fit an individual SVM providing a weak learner for the ensemble which was applied to the test data, that is, the ADNI sample. The process was repeated 10,000 times, yielding 10,000 age predictions based on models trained on (different) balanced subsamples of the multi‐cohort reference data. These predictions were then averaged (“bagging”) to yield the final age prediction based on the 673‐parcel representation of the voxel‐based morphometry data (Becker, Mahlke, Reckert, Eickhoff, & Ritz‐Timme, 2019). Each subject's BrainAGE score was finally calculated as bagged predicted age minus chronological age for each subject (Figure 1).

Open multimedia modal

2.5 Data availability

The original data used in this manuscript are publicly available in the online address of the ADNI database at http://adni.loni.usc.edu/data-samples/access-data/.

3.1 Effects on gray matter volume

There were strong (P_FWE < .001) and widespread negative associations of regional gray matter volume with “age,” in particular in the bilateral temporal lobes, bilateral prefrontal, middle and superior frontal areas, bilateral medial and lateral occipital areas, cerebellum and thalamus, caudate and putamen in the subcortical gray matter (Figure 2a). The “cognitive status” was significantly associated with reduced gray matter volume in many bilateral parcels with dominancy in the left hemisphere (P_FWE < .001). Bilateral temporal lobes including fusiform gyri, medial temporal lobes, and hippocampal formations, and inferior and middle temporal lobes, as well as bilateral insula, middle frontal, and cingulate cortices, as well as left superior frontal cortex had significantly lower volume in participants with MCI and particularly AD (Figure 2b). In turn, when testing for effects of SDB status and SDB‐by‐cognitive status interaction, we found no significant region anywhere in the brain (all P_FWE > .05).

Open multimedia modal

3.2 Effects on estimated brain age

The mean absolute error (MAE) between predicted and chronological age in the HC group was 3.59 years, indicative of the very good performance of the ensemble prediction model. We then calculated the BrainAGE score as the per‐subject difference between predicted and chronological age and tested for its association with cognitive status, SDB status, and the SDB‐by‐cognitive status interaction. Participants with MCI and in particular AD showed an advanced brain age (on average 4.0 (95% confidence interval or CI: 2.6–5.4) and 9.1 (95%CI: 5.8–12.4) years, respectively) (Figure 3), in line with previous studies. However, there was no significant effect on BrainAGE scores associated with SDB status, nor was there a positive SDB‐by‐cognitive status interaction, suggesting that SDB may not lead to advanced brain aging (Figure 3c).

Open multimedia modal

4.1 Gray matter volume alterations in AD and SDB

One of the main characteristics of MCI and AD is generalized gray matter loss in the brain, which mostly starts in the medial temporal lobe and multimodal association areas (Fox & Schott, 2004; C. R. Jack et al., 2004; Karas et al., 2004; Pasquini et al., 2019). Neuroimaging meta‐analyses in AD have demonstrated atrophy in the medial temporal lobe, limbic regions (left parahippocampal gyrus, left posterior cingulate gyrus, amygdala, and uncus), thalamus, temporal, parietal, frontal and cingulate cortices (W.‐Y. Wang et al., 2015; Yang et al., 2012). A similar, but milder distribution of gray matter atrophy is evident in the brain of patients with MCI (Nickl‐Jockschat et al., 2012; Yang et al., 2012). In accordance with the previous brain volumetric studies, we found diffuse gray matter loss in MCI and AD. The atrophy was mainly located in the bilateral temporal lobe and medial temporal areas with higher intensity in AD compared to MCI.

Assessing the volumetric changes due to SDB, we did not observe any significant alteration in gray matter volume, neither in HC subjects, nor in patients with MCI or AD. Furthermore, self‐reported SDB interaction with cognitive status (HC, MCI, AD) revealed no associations with gray matter volume. Historically, there has been an inability to replicate results among the brain imaging studies of SDB in non‐demented populations. While several studies have reported gray matter atrophy in the insula, amygdala, middle and lateral temporal regions, and cerebellum in non‐demented populations with SDB (Joo et al., 2013^, 2010; Morrell et al., 2010; Torelli et al., 2011; Weng et al., 2014), others have either shown no associations (O'Donoghue et al., 2005; Yun et al., 2017) or even enhancement in the gray matter volume in the motor cortices, prefrontal cortex, thalamus, putamen, and the hippocampus (Baril et al., 2017; Kumar et al., 2014; Lin et al., 2016; Lundblad et al., 2014; Rosenzweig et al., 2013; Taylor et al., 2018). In addition, there is a general lack of longitudinal studies, which would enable the study of nonlinear associations between SDB and cortical atrophy, suggested by the present cross‐sectional findings. Despite these important gaps in the literature, three neuroimaging meta‐analyses have demonstrated that OSA is associated with gray matter atrophy in a few selected regions including the amygdala and hippocampus (Tahmasian et al., 2016), as well as cingulate, right central insula, right middle temporal gyrus, right premotor cortex, and cerebellum (Shi et al., 2017; Weng et al., 2014).

The observed null association between SDB and gray matter volume should, however, be interpreted with caution. Firstly, it has been suggested that aging may have partially protective mechanisms against SDB, such as reduced production of oxidative stress after apneas and decreased blood pressure and heart rate responses after arousals (Baril et al., 2017). The average old age of ADNI subjects (~75 years‐old) could, therefore, explain this nonsignificant association between SDB and brain morphometry. Despite numerous individual studies and meta‐analyses focused on the changes in gray matter in middle‐aged patients with OSA, there are few studies on gray matter changes in older adults with SDB and, to our knowledge, neither have found decreases in thickness or volume in cortical gray matter (Sébastien Celle et al., 2009; Cross et al., 2018; Lutsey et al., 2016). Secondly, it is possible that SDB impairs selective brain functions (Canessa et al., 2018) or amyloid burden (Yun et al., 2017) before gray matter volume. Furthermore, differential diagnosis between SDB‐related and age‐related brain atrophy is difficult in single‐point observational studies, particularly in those cases in which groups are matched by age and cognitive status. Thirdly, this could also be a sign of a) survival bias, as most patients with SDB may have transitioned to AD and only those with very low cortical atrophy or high in cognitive reserve at disease onset would remain as HC or MCI at cross‐section; or b) selection bias due to matching by the APOE4 allele, as it has been reported that the APOE4 allele interacts with brain aging scores measured by the BrainAGE method, revealing potential neuronal compensation in healthy APOE4+ adults (Scheller et al., 2018), which could also result in null findings. Fourthly, we did not account for other comorbidities and possible confounders alongside age or presence of the APOE4 allele in the prediction models (Gozal, 2018). Finally, previous MRI studies mostly recruited patients with polysomnography‐diagnosed OSA from sleep clinics, which might be a different population from those recruited in memory clinics with a self‐reported assessment of SDB based on their clinical interview.

Interestingly, we were not able to demonstrate any interaction between SDB and MCI or AD with brain atrophy. This is indicative that despite the frequent clinical co‐occurrence of SDB and AD, there may be no synergy between them in accelerating gray matter atrophy. Recent investigations using cerebrospinal fluid and PET imaging suggest an interplay between amyloid production/clearance and SDB (Bubu et al., 2019; Liguori et al., 2017; Sharma et al., 2018; Spira et al., 2013; Yun et al., 2017). These include an impairment in the cerebrospinal fluid–interstitial fluid exchange (Ju et al., 2016), cerebral edema secondary to an intermittent hypoxia (Spira et al., 2018) (similar to the increase in brain volume and pseudoatrophy observed in multiple sclerosis), and compensatory excessive neuronal synaptic activity (Polsek et al., 2018) in SDB, all of which could potentially lead to an increase in beta‐amyloid deposition and its clearance reduction. It is, therefore, possible that the presence of SDB is associated with AD risk only through beta‐amyloid deposition (Bubu et al., 2019; Sharma et al., 2018) or altered brain function (Chen et al., 2018; Park et al., 2016; Thomas, Rosen, Stern, Weiss, & Kwong, 2005), as mentioned before. While amyloid burden has been linked to SDB in several observational studies (Bubu et al., 2019; GBD 2016 Disease and Injury Incidence and Prevalence Collaborators et al., 2017; Sharma et al., 2018), a recent study on non‐demented elderly subjects, has found no association between self‐reported sleep disturbances and brain amyloid PET burden (Gabelle et al., 2019). In addition, in our results, we expected an interaction with MCI or AD where it is generally accepted that neuronal loss follows amyloid deposition. More studies are needed to better understand the compensatory increase in gray matter volume in SDB suggested by several studies, as well as the precise progression of brain atrophy in AD, as both may have contributed to obtaining such negative findings.

4.2 BrainAGE prediction in AD and SDB

Brain age prediction methods have been previously used in cognitively normal subjects (Aycheh et al., 2018; Franke et al., 2012). In addition, several studies have used the ADNI dataset and other datasets of middle‐aged adult and elderly population with MAE ranging from 3.8 to 6 years (Cole et al., 2019; Varikuti et al., 2018). Conversely, in our study, an advanced sensitive BrainAGE estimation method has been implemented to detect pathologic brain aging. A repeated support vector machine (SVM) models were fitted on parcel‐wise gray matter volume data of on stratified subsamples from external cohorts, making the model notably less sensitive to heterogeneity in images (Varikuti et al., 2018). In addition, compared to previous studies on the middle‐aged adults and elderly subjects (Cole et al., 2019), while using multiple datasets for training prediction model, our age prediction results were accurate with an MAE of 3.6 years in older adults and elderly subjects. Replication of previous findings in patients with AD, taken together with acceptable MAE, is indicative of the reliability of our proposed method in gray matter volume assessment and age estimation.

While there is no exact definition for accelerated brain aging, the BrainAGE score has been shown to be a sensitive predictor of disease progression in dementia (Cole et al., 2019; Gaser et al., 2013; Löwe et al., 2016). Previous findings on increased BrainAGE score in MCI and AD course (Beheshti, Maikusa, & Matsuda, 2018; Caballero, Klöppel, Dichgans, & Ewers, 2016; Liem et al., 2017), are in agreement with the reported accelerated aging of the demented brain shown in‐vivo and ex‐vivo studies (Mecocci et al., 2018). The BrainAGE score in studies using ADNI ranged from almost zero for patients with stable MCI, to 5.7–6.2 years for patients with progressive MCI, and reached up to 10 years for patients with AD (Cole et al., 2019). We found the average 4.1 and 9 BrainAGE scores in patients with AD and MCI, in agreement with previous findings using ADNI data. Since we did not distinguish patients with progressive from stable MCI, our results in the MCI group were modest compared to other studies including patients with late or progressive MCI.

DATA AVAILABILITY STATEMENT

Data used in the preparation of this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf.

ACKNOWLEDGMENTS

The authors gratefully acknowledge the efforts, time, and dedication of the participants and staff of the Alzheimer's Disease Neuroimaging Initiative (ADNI). Collection and sharing of data for this project was funded by ADNI (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH‐12‐20012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer's Association; Alzheimer's Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol‐Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann‐La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health (www.fnih.org). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer's Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California. Of note, Simon B. Eickhoff is supported by the Deutsche Forschungsgemeinschaft, the National Institute of Mental Health (R01‐MH074457), the Helmholtz Portfolio Theme “Supercomputing and Modeling for the Human Brain” and the European Union's Horizon 2020 Research and Innovation Programme under Grant Agreement No. 785907 (HBP SGA2). Ricardo Osorio's salary is supported by NIH/NIA'a R01AG056031, R01AG056531, and R21AG055002. Nooshin Abbasi was funded by Preston Robb Fellowship.