Biological brain age prediction using machine learning on structural neuroimaging data: Multi-cohort validation against biomarkers of Alzheimer’s disease and neurodegeneration stratified by sex
Curation statements for this article:-
Curated by eLife
eLife assessment
The study has some significance for the field of dementia research and neurodegenerative diseases more broadly. Using the brain-age paradigm, the main findings are that having an older-appearing brain is associated with more advanced stages of amyloid and tau pathology, higher white matter hyperintensities, higher plasma NfL and carrying the APOE 34 allele. Findings were broadly similar in cognitively normal people and people with mild cognitive impairment and the evidence for these findings is convincing. Although sex differences are emphasized, the evidence for this is generally incomplete.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
- Neuroscience (eLife)
Abstract
Brain-age can be inferred from structural neuroimaging and compared to chronological age (brain-age delta) as a marker of biological brain aging. Accelerated aging has been found in neurodegenerative disorders like Alzheimer’s disease (AD), but its validation against markers of neurodegeneration and AD is lacking. Here, imaging-derived measures from the UK Biobank dataset (N=22,661) were used to predict brain-age in 2,314 cognitively unimpaired (CU) individuals at higher risk of AD and mild cognitive impaired (MCI) patients from four independent cohorts with available biomarker data: ALFA+, ADNI, EPAD, and OASIS. Brain-age delta was associated with abnormal amyloid-β, more advanced stages (AT) of AD pathology and APOE -ε4 status. Brain-age delta was positively associated with plasma neurofilament light, a marker of neurodegeneration, and sex differences in the brain effects of this marker were found. These results validate brain-age delta as a non-invasive marker of biological brain aging in non-demented individuals with abnormal levels of biomarkers of AD and axonal injury.
Article activity feed
-
-
eLife assessment
The study has some significance for the field of dementia research and neurodegenerative diseases more broadly. Using the brain-age paradigm, the main findings are that having an older-appearing brain is associated with more advanced stages of amyloid and tau pathology, higher white matter hyperintensities, higher plasma NfL and carrying the APOE 34 allele. Findings were broadly similar in cognitively normal people and people with mild cognitive impairment and the evidence for these findings is convincing. Although sex differences are emphasized, the evidence for this is generally incomplete.
-
Reviewer #1 (Public Review):
This work provides a comprehensive assessment of volumetric-MRI-based brain age estimates in relation to AD-related biomarkers and AD risk factors. Brain age modeling has been studied extensively in recent years. Brain age estimates are suggested surrogate markers for aging-associated changes in the brain. This paper provides findings on how brain age estimates are associated with AD-related amyloid and tau accumulation, cerebrovascular white matter disease, and unspecific neurodegeneration detected by plasma NfL and to some extent CSF NfL as well. The authors also provide important results on sex-specific differences in these associations.
Strengths:
Modeling and analyses were performed on different observational cohorts. Analysis was repeated for the cognitively unimpaired, and individuals with MCI …
Reviewer #1 (Public Review):
This work provides a comprehensive assessment of volumetric-MRI-based brain age estimates in relation to AD-related biomarkers and AD risk factors. Brain age modeling has been studied extensively in recent years. Brain age estimates are suggested surrogate markers for aging-associated changes in the brain. This paper provides findings on how brain age estimates are associated with AD-related amyloid and tau accumulation, cerebrovascular white matter disease, and unspecific neurodegeneration detected by plasma NfL and to some extent CSF NfL as well. The authors also provide important results on sex-specific differences in these associations.
Strengths:
Modeling and analyses were performed on different observational cohorts. Analysis was repeated for the cognitively unimpaired, and individuals with MCI separately.
Weaknesses:
Although the authors concluded that brain age prediction is a biomarker of AlD pathology, only associations were assessed in this study. Further analyses are required to truly assess the biomarker value of brain age prediction for AD pathology.
-
Reviewer #2 (Public Review):
In this work, the authors used machine learning techniques to predict chronological age in the large UK Biobank dataset using structural neuroimaging measures of regional brain volumes and cortical thickness in sex-stratified models. From these predictions, the authors calculated the brain-age delta, which is thought to reflect biological brain aging. The authors applied these models to four independent cohorts and calculated brain-age delta, which they then associated with several markers of Alzheimer's disease pathology, neurodegeneration, and cerebrovascular disease. The aim of these analyses was to validate brain-age delta as a clinically relevant marker of AD.
Strengths
This is a well-written manuscript that explains a well-powered study of multiple deeply-phenotyped cohorts. An impressive amount of …Reviewer #2 (Public Review):
In this work, the authors used machine learning techniques to predict chronological age in the large UK Biobank dataset using structural neuroimaging measures of regional brain volumes and cortical thickness in sex-stratified models. From these predictions, the authors calculated the brain-age delta, which is thought to reflect biological brain aging. The authors applied these models to four independent cohorts and calculated brain-age delta, which they then associated with several markers of Alzheimer's disease pathology, neurodegeneration, and cerebrovascular disease. The aim of these analyses was to validate brain-age delta as a clinically relevant marker of AD.
Strengths
This is a well-written manuscript that explains a well-powered study of multiple deeply-phenotyped cohorts. An impressive amount of work went into this manuscript and that is evident from reading it. The manuscript was enjoyable to read and easy to follow, and the authors provided an informative summary figure visualizing the analysis plan of this work. More specifically there are five key strengths in this present work.
First, instead of aiming for a brain-predicted age model with optimal predictive accuracy, as is typically the case in studies using brain-age delta measures, the authors used a model with a restricted feature set and a limited age range to allow for better neurobiological interpretability and to increase the relevance of this model to ageing cohorts.
Second, the authors corrected for the proportional bias that is seen in brain age models and controlled subsequent analyses (i.e. associations between brain-age delta and markers of AD pathology, etc.) for chronological age. This is an important and necessary step when working with brain-age delta but is not always implemented across studies.
Third, the authors computed Shapley Additive explanation values (SHAP) which quantified the contribution of different brain regions to the brain age prediction. This ensured that the model had neurobiological interpretability which is not always the case with brain-age prediction models. This was further improved by using a relatively restricted feature set that is often used in brain-age prediction studies as the most important regions could be easily visualized and therefore more readily interpreted. This is in contrast to other models that use a large number of smaller brain features, which are less easily vizualised and less interpretable.
Fourth, importantly, the authors used sex-stratified models as they generated the brain-age delta measures separately in men and women. This allowed for sex-specific analyses of the associations between brain-age delta and markers of AD pathology, cerebrovascular disease, and neurodegeneration, which is important given evidence of sex differences in AD. These sex-stratified models also enabled the authors to compare the most relevant brain regions in the brain age prediction models. While previous work has reported sex differences in brain-age delta, the sex-specific contribution of specific brain features is important information that is not usually reported.
Finally, in addition to investigating the association of brain-age delta with specific markers of AD pathology, cerebrovascular disease, and neurodegeneration, the authors also analyzed the association between brain-age delta and amyloid and tau status stages which provides important clinically relevant information. This information is important if future work aims to further investigate the use of brain-age delta in the field of AD.Limitations
There are three important weaknesses in this present work. First, the conclusion that "These results validate brain-age delta as a non-invasive marker of biological brain aging related to markers of AD and neurodegeneration" (from the Abstract) may be overstated. While we assume that brain-age delta reflects an accelerated ageing process, this is still a cross-sectional measure and the results show cross-sectional associations with markers of AD and neurodegeneration. For true validation of this measure as a non-invasive marker of biological brain aging with respect to markers of AD and neurodegeneration, we would need longitudinal data to show that changes in brain age are longitudinally associated with changes in markers of AD and neurodegeneration.
Second, the authors reported that brain-age delta was not related to longitudinal brain change ('aging signature change'), which supports a recent finding that cross-sectional brain-age delta was not associated with longitudinal brain change but was associated with birthweight and polygenic risk scores for brain-age delta (Vidal-Pineiro et al., 2021 eLife). This previous finding led to the conclusion that brain-age delta may reflect early-life factors more so than longitudinal brain change or 'accelerated brain ageing'. This is a critical issue to contend with if we really wish to pursue further validation of the brain-age delta as a potential marker of aging
Third, the analyses for the associations between brain-age delta and other variables are not corrected for multiple comparisons, even though a large number of comparisons are conducted. This means that some of the apparently significant results could be false positives. Appropriately correcting these analyses for multiple comparisons would strengthen the results, allowing for greater confidence in the significant results, and would avoid mistaken interpretations of false positive findings.Appraisal
The authors developed accurate and generalizable sex-specific measures of the brain-age delta. The authors demonstrated that brain-age delta was associated with measures of AD pathology and neurodegeneration. These have the potential to be useful findings that may promote the use of the brain-age delta in AD research. However, as these results are not corrected for multiple comparisons it is possible that some of these results may be false positives. Moreover, the finding that brain-age delta was not associated with longitudinal brain change may undermine the conclusions, as it could suggest that brain-age delta is not reflective of accelerated brain ageing.Impact
I believe that this work has two important impacts. First, the methods demonstrated in the present study highlight that sex-stratified models may be necessary for future brain-age delta studies, and given that the models were externally validated in four separate cohorts, a key impact is that future researchers will be able to apply the well-described brain-age models here in their own work. Second, the finding that brain-age delta was not related to longitudinal brain change or atrophy, supports previous similar findings and could suggest that brain-age delta does not, as previously assumed, reflect accelerated brain ageing. This may indicate that the brain-age delta is not a satisfactory marker of brain ageing and therefore could discourage future work with this metric that attempts to validate it is a clinical marker of brain ageing. If this issue could be alternatively explained or if brain-age delta is, in fact, shown to reflect brain ageing, then an additional potential impact is that it may support the future investigation into the use of brain-age delta in longitudinal studies of brain ageing and neurodegeneration. -
Reviewer #3 (Public Review):
Cumplido-Mayoral and colleagues' study focused on the brain-age paradigm in the context of Alzheimer's disease risk. The goal was to valid brain-age 'deltas' by assessing how they relate to Alzheimer's biomarkers and related neurodegenerative measures. They did this by training a new brain-age model on FreeSurfer phenotypes (cortical and subcortical) using the UK Biobank dataset. They then tested multiple datasets including ALFA, ADNI, OASIS, and EPAD, focusing on cognitively unimpaired people and people with mild cognitive impairment. Using brain-age deltas calculated in the test sets, the authors then tested associations with a range of dementia-related measures, including the presence of MCI, APOE e4, amyloid and tau positivity, white matter hyperintensity volume and NfL levels from plasma or CSF.
Strength…
Reviewer #3 (Public Review):
Cumplido-Mayoral and colleagues' study focused on the brain-age paradigm in the context of Alzheimer's disease risk. The goal was to valid brain-age 'deltas' by assessing how they relate to Alzheimer's biomarkers and related neurodegenerative measures. They did this by training a new brain-age model on FreeSurfer phenotypes (cortical and subcortical) using the UK Biobank dataset. They then tested multiple datasets including ALFA, ADNI, OASIS, and EPAD, focusing on cognitively unimpaired people and people with mild cognitive impairment. Using brain-age deltas calculated in the test sets, the authors then tested associations with a range of dementia-related measures, including the presence of MCI, APOE e4, amyloid and tau positivity, white matter hyperintensity volume and NfL levels from plasma or CSF.
Strengths include using multiple independent datasets from different sources. This provides large sample sizes and access to different data types. Another strength is the efforts to understand drivers of brain age prediction, by using the SHAP technique. The authors include a newly trained brain-age prediction model, which appears to work as well as existing alternative methods.
A weakness is the number of tests conducted and the absence of multiple comparison corrections. A problem with the SHAP analysis is that it does not account for the correlated nature of the input features.
Overall, the study met the stated aims, and I anticipate the results to make a positive contribution to the research field. The results tended to support the conclusions, particularly regarding the relationship between brain-age delta and the markers of neurodegeneration, AD risk, and cerebrovascular health. The only concern around this is whether the number of tests conducted has inflated the type I error rate and resulted in some false positives. This could have been explored further. The conclusions are sex differences are less well supported by the evidence. While some delta-by-sex interactions were significant, others were not (e.g., Figure 3), however, the interpretation focuses only on the significant ones to support blanket statements about the differences between males and females with regard to neurodegeneration. Given the issues about multiple comparisons, this seems premature and somewhat uneven.
-
