Critical ages in the life course of the adult brain: Nonlinear subcortical aging

TL;DR: The main conclusions are that most brain structures do not follow a simple path throughout adult life and that accelerated decline in high age is not the norm of healthy brain aging.

About: This article is published in Neurobiology of Aging.The article was published on 2013-10-01 and is currently open access. It has received 326 citations till now.

Summary

1. Introduction

• This is likely to produce more complex trajectories than what can be described by linear or the usually employed higher order polynomial (quadratic or even cubic) models (Fjell et al., 2010a).
• The present study was undertaken with the purpose of estimating trajectories across age of 17 brain structures in a large cross-sectional sample (n = 1100).
• This makes it possible to identify critical ages where life-phases characterized by relative stability are followed by periods where estimated atrophy accelerates, or critical ages where periods of estimated reduction eventually level off.
• The rational for the present study was to go beyond these general trends, by more accurately delineating the trajectories for the different structures across adult life, and to identify critical ages characterized by changes in estimated rate of atrophy.

2.1 Samples

• Distribution of participants across decades are shown in Table 1.
• All the healthy samples were screened for diseases and history of neurological conditions and dementia, and none of the participants showed signs of cognitive dysfunction.
• [Insert Table 1 about here] 2.1.2 Longitudinal sample:.
• The longitudinal sample consisted of 142 (60-90 years, mean age = 75.6 years, 48% females) participants from the Alzheimer Disease Neuroimaging Initiative (ADNI) database (www.loni.ucla.edu/ADNI), followed for one year.
• The raw data were obtained from the ADNI database, Principal Investigator Michael W. Weiner, VA Medical Center and University of California – San Francisco.

2.2 MRI processing

• All scans were obtained from 1.5T magnets from two different manufacturers (Siemens, Erlangen, Germany; General Electric CO, Milwaukee, WI), and from five different models (Siemens: Avanto, Symphony, Sonata, Vision/ GE: Signa).
• All segmentations were manually inspected for accuracy by an experienced operator, and corrected in case of errors.
• 8 Intracranial volume (ICV) was estimated by use of an atlas-based normalization procedure, where the atlas scaling factor is used as a proxy for ICV, shown to correlate highly with manually derived ICV (r = .93) (Buckner et al., 2004).
• In a previous publications with an overlapping sample pool, the results for the pooled samples were replicable in each of the subsamples (Fjell et al., 2009b, Walhovd et al., 2011), indicating that the sensitivity of detecting effects are upheld and the statistical power are increased manifold.
• Two MPRAGEs at each time-point were averaged to increase the ANR and CNR.

2.4 Statistical analyses

• To reduce the number of comparisons, mean values for left and right hemisphere were used in all ROI analyses.
• For the cross-sectional analyses, a nonparametric local smoothing model, the smoothing spline, implemented in Matlab, was fitted to the data.
• The quadratic function is always a parabola, which sometimes causes the model to indicate a non-monotonous agerelationship when non-linear but monotonous trajectories are a more likely.
• BIC rewards goodness of fit, but includes a penalty that is an increasing function of the number of estimated parameters.
• First, the authors tested whether the structures or regions that showed increases or decreases in the 11 full cross-sectional sample also showed longitudinal increases or decreases, respectively, in the independent ADNI sample.

3.1 Cross-sectional data

• To compare the linear and the smoothing spline models, the authors calculated BIC for the relationship between each brain volume and age (Supplementary Table 2, also including the quadratic model for comparison purposes).
• Scatterplots illustrating the estimated trajectories are presented in Figure 1 and Figure 2 (ventricular system).
• Of the 17 tested regions, a non-linear model represented the data best for 13 (total brain volume, cerebral cortex and WM, hippocampus, caudate, cerebellar WM, brain stem, pallidum, putamen, and lateral, inferior lateral, 3rd and 4th ventricle).
• Inspections of the plots revealed substantial differences in estimated trajectories for the non-linear models.
• Early critical age varied greatly across structures, from 31 to 59 years, and volume-age correlations differed between the defined periods.

3.2 Longitudinal validation

• All ROIs showed significant longitudinal change at p < .05.
• This confirmed the finding of substantial atrophy/ ventricular expansion observed cross-sectionally for all ROIs, except the caudate nucleus.
• For caudate, a weak positive correlation with age was observed after 59 years in the cross-sectional data, which was not found in the longitudinal analyses.
• Next, the authors studied how well the pattern of cross-sectionally estimated change matched the longitudinal findings.
• In the age-range 60-90 years, Spearman ρ between the cross-sectional estimate of shrinkage and the longitudinally measured volume loss was .91 (p < 10-5).

4. Discussion

• There were three main findings: First, a heterogeneous pattern of discontinuous agecorrelations in different age-spans characterised the majority of brain regions, and critical ages for changes in estimated rates of atrophy could be identified.
• Second, accelerated estimated reduction with advanced age is not the norm of brain aging.
• Rather, different structures showed a mix of trajectories.
• When more negative (positive for CSF) age-volume correlations were seen in the last part of the age-span, this would typically start in mid-life.
• Finally, the longitudinal analyses in general supported the cross-sectional results, with a reasonably coherent pattern of atrophy across structures.

4.1 Trajectories of estimated change across the adult life-span

• Cross-sectional studies have shown non-linear age-relationships for the volume of several brain structures (Raz et al., 2004, Allen et al., 2005, Lupien et al., 2007), including studies with samples overlapping the present (Walhovd et al., 2009).
• The authors found that after a period of relative stability during mid-life, accelerated estimated reductions started at about 50 years of age, followed by a strongly negative linear age-relationship from 60 years.
• Cerebral WM was the only structure positively correlated with age in the earliest part of the age-range, followed by a strong negative relationship.
• This pattern is in line with a previous publication reporting multi-modal imaging data from 8-85 years, partly overlapping sample five (Westlye et al., 2010b).
• 16 Caudate was the most deviant structure, best described by a U-shaped trajectory.

4.2 Critical ages in estimated regional brain change

• The trajectory of a neuroanatomical volume across age represents the additive combination of several neurobiological processes.
• The authors suggest that changes in the relative impact of these can be observed as turning points in the estimated change in brain volumes, what they refer to as critical ages (see Figure 4).
• This will be affected by medical conditions such as hypertension, cholesterol, diabetes or metabolic syndrome, genetic variations such as apolipoprotein E (APOE), and variables such as cognitive activity and education.
• To speculate, one scenario may be as follows:.
• At this point, the second derivative of the age-volume trajectory will change, representing a critical age.

4.3 Cross-sectional vs. longitudinal results

• It is impossible to infer changes in brain structures based on cross-sectional data alone (Raz and Lindenberger, 2010), as this depends on assumptions of no cohort-effects and selection 18 bias.
• Thus, at least in their rank order of magnitude, the cross-sectional results for the age-range above 60 years seem to be largely in coherence with independent longitudinal data.
• Nonetheless, caution must still be exercised in interpreting the results, as longitudinal data were available for the oldest part of the sample only.
• Of more general concern is that the inherent problem of mapping life-span trajectories from cross-sectional examinations cannot easily be resolved with longitudinal data, since longitudinal examinations of brain structures over decades are not feasible, and longitudinal studies have methodological problems of their own (e.g. selective recruitment and attrition).
• An ideal approach to reproduce the dynamic process of change would be longitudinal studies with high density of measures and assessment of multiple time windows across the life span (Raz et al., 2010; Raz & Lindenberger, 2011).

Figures

Table 2 Testing of linearity

Frequently asked questions

Q1. What are the contributions mentioned in the paper "Critical ages in the life-course of the adult brain: nonlinear subcortical aging" ?

The main conclusions are that most brain structures do not follow a simple path throughout adult life, and that accelerated decline in high age is not the norm of healthy brain aging. 

Q2. How many MPRAGEs were averaged before pre-processing?

For samples 1, 2, 3 and 5, 2-4 MPRAGEs were averaged before pre-processing to increase signalto-noise (SNR) and contrast-to-noise ratio (CNR). 

Q3. What is the next step to increasing understanding of the mechanisms of brain aging?

Important next step to increased understanding of the mechanisms of brain aging will be to conduct large-scale, multi-modal imaging studies, combining e.g. volumetry, DTI and intensity/contrast measures (Fjell et al., 2008, Westlye et al., 2010b, a), as well as longitudinal studies with high density of measurements to examine the trajectories across age with regards to the critical phases proposed on the basis of the cross-sectional analyses.21 

Q4. Why did the longitudinal and cross-sectional results differ?

Because the methods used to calculate longitudinal change and to fit the cross-sectional trajectories differ in important aspects, and the samples do not overlap, direct comparisons of estimations of absolute rates of atrophy between the longitudinal and cross-sectional results were not performed. 

Q5. What is the point of maximum acceleration of slope change?

for quadratic models, the second derivative is assumed to be constant across the life span, and hence the point of maximum acceleration of slope change cannot be determined. 

Q6. What is the approach to reproduce the dynamic process of change?

An ideal approach to reproduce the dynamic process of change would be longitudinal studies with high density of measures and assessment of multiple time windows across the life span (Raz et al., 2010; Raz & Lindenberger, 2011). 

Q7. What is the algorithm for smoothing spline?

The authors used an algorithm that optimizes smoothing level based on a version of Bayesian Information Criterion (BIC), i.e. the smoothing level that minimizes BIC for each analysis was chosen. 

Q8. What was the average volume of the brain in the cross-sectional data?

Brain volumes in the cross-sectional data were regressed on sample and ICV, and age-reductions estimated from the cross-sectional data were measured in standard deviation decline in volume in the age-range 60 to 90. 

Q9. What is the structure that distinguishes between healthy and AD patients?

Even though hippocampal volume is the structure that distinguishes best between AD-patients and healthy elderly, amygdala is also affected in early stages of the disease (Fjell et al., 2010b). 

Q10. What is the significance of the hippocampus in AD?

Hippocampus is especially important due to its role in memory and early AD (de Leon et al., 2006, Du et al., 2007, Jack et al., 2008, Fennema-Notestine et al., 2009, McEvoy et al., 2009). 

Q11. What is the rank order of magnitude of the cross-sectional results?

at least in their rank order of magnitude, the cross-sectional results for the age-range above 60 years seem to be largely in coherence with independent longitudinal data. 

Q12. How many major regions and structures were included in the study?

The authors included volume for 17 major regions and structures estimated from the whole-brain segmentation approach in FreeSurfer (Fischl et al., 2002). 

Q13. What is the way to infer changes in brain volume?

It is impossible to infer changes in brain structures based on cross-sectional data alone (Raz and Lindenberger, 2010), as this depends on assumptions of no cohort-effects and selection18bias. 

Q14. What were the critical ages for the structures that showed deviations from linearity?

For the structures that showed deviations from linearity (except putamen), critical ages, i.e. the ages where estimated atrophy started to accelerate or decelerate, were identified. 

Q15. What was the way to test the stability of the results?

To test the stability of the results, a split half analysis was performed for WM volume (Supplementary Figure 2), yielding identical spline curves. 

Q16. How did the authors calculate the slope of the local smoothing curve?

The authors calculated the ages where the slope of the local smoothing curve changed (the secondderivative), using the expression − d2 f age( )d age2 .