Multimodal brain age estimates relate to Alzheimer disease biomarkers and cognition in early stages: a cross-sectional observational study
Curation statements for this article:-
Curated by eLife
eLife assessment
This is a useful study exploring multi-modality brain age (structural plus resting state MRI) in people in the early stages or at risk of Alzheimer's disease. They found solid evidence that people with cognitive impairment had older-appearing brains and that older-appearing brains were related to Alzheimer's risk factors such as amyloid and tau deposition. They claim to show that the multi-modality brain age model is more accurate than a unimodal structural MRI model, though the evidence for that is incomplete.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
- Neuroscience (eLife)
Abstract
Estimates of ‘brain-predicted age’ quantify apparent brain age compared to normative trajectories of neuroimaging features. The brain age gap (BAG) between predicted and chronological age is elevated in symptomatic Alzheimer disease (AD) but has not been well explored in presymptomatic AD. Prior studies have typically modeled BAG with structural MRI, but more recently other modalities, including functional connectivity (FC) and multimodal MRI, have been explored.
Methods:
We trained three models to predict age from FC, structural (S), or multimodal MRI (S+FC) in 390 amyloid-negative cognitively normal (CN/A−) participants (18–89 years old). In independent samples of 144 CN/A−, 154 CN/A+, and 154 cognitively impaired (CI; CDR > 0) participants, we tested relationships between BAG and AD biomarkers of amyloid and tau, as well as a global cognitive composite.
Results:
All models predicted age in the control training set, with the multimodal model outperforming the unimodal models. All three BAG estimates were significantly elevated in CI compared to controls. FC-BAG was significantly reduced in CN/A+ participants compared to CN/A−. In CI participants only, elevated S-BAG and S+FC BAG were associated with more advanced AD pathology and lower cognitive performance.
Conclusions:
Both FC-BAG and S-BAG are elevated in CI participants. However, FC and structural MRI also capture complementary signals. Specifically, FC-BAG may capture a unique biphasic response to presymptomatic AD pathology, while S-BAG may capture pathological progression and cognitive decline in the symptomatic stage. A multimodal age-prediction model improves sensitivity to healthy age differences.
Funding:
This work was supported by the National Institutes of Health (P01-AG026276, P01- AG03991, P30-AG066444, 5-R01-AG052550, 5-R01-AG057680, 1-R01-AG067505, 1S10RR022984-01A1, and U19-AG032438), the BrightFocus Foundation (A2022014F), and the Alzheimer’s Association (SG-20-690363-DIAN).
Article activity feed
-
-
eLife assessment
This is a useful study exploring multi-modality brain age (structural plus resting state MRI) in people in the early stages or at risk of Alzheimer's disease. They found solid evidence that people with cognitive impairment had older-appearing brains and that older-appearing brains were related to Alzheimer's risk factors such as amyloid and tau deposition. They claim to show that the multi-modality brain age model is more accurate than a unimodal structural MRI model, though the evidence for that is incomplete.
-
Reviewer #1 (Public Review):
The authors sought to explore the brain age paradigm in the early stages of Alzheimer's disease, focusing on the combination of different MRI modalities (brain structure derived from T1-weighted MRI and functional connectivity derived from resting state fMRI). Their goal was to understand how different unimodal brain ages and a combined multi-modal brain age related to risk factors related to Alzheimer's disease, namely hippocampal volume, cognitive performance, amyloid or tau positivity from PET scans or CSF data and neurofilaments. As part of this, they aimed to ascertain which brain age models performed more accurately.
The major strength of the methods is the novel combination of different MRI modalities using Gaussian Processes and stacking to predict brain age. Another strength is the use of multiple …
Reviewer #1 (Public Review):
The authors sought to explore the brain age paradigm in the early stages of Alzheimer's disease, focusing on the combination of different MRI modalities (brain structure derived from T1-weighted MRI and functional connectivity derived from resting state fMRI). Their goal was to understand how different unimodal brain ages and a combined multi-modal brain age related to risk factors related to Alzheimer's disease, namely hippocampal volume, cognitive performance, amyloid or tau positivity from PET scans or CSF data and neurofilaments. As part of this, they aimed to ascertain which brain age models performed more accurately.
The major strength of the methods is the novel combination of different MRI modalities using Gaussian Processes and stacking to predict brain age. Another strength is the use of multiple data sources for both model training and testing, reducing the reliance on a single site and decreasing the likelihood of overfitting, which should improve generalisability. A weakness is the poor fit of the functional connectivity model to the data, whereby the vast majority of test participants were shown to have younger appearing brains, even those with cognitive impairment. This indicates that an alternative fMRI processing pipeline could have been beneficial, however, no experimentation on this important facet of the analysis was included. Another weakness is the relatively limited sample size compared too much of the brain age literature and the failure to report the R^2 metric, which is important for the comparison of this study with previously published reports. Potentially, more accurate models would have led to clearer results, as there are a number of borderline findings which hinder clear interpretation.
In general, the study did meet its stated goals and was able to generate a multi-modality brain-age model and this model did show older appearing brains in people with cognitive impairment. This model also showed that people with older appearing brains had poorer cognition, lower hippocampal volume, and greater amyloid deposition. In people who met the criteria for being cognitively impaired, greater tau deposition on PET scans was associated with an older appearing brain. One claim of the study is that the multi-modality brain-age model was more accurate than the brain-volume model, however, it is unclear from the report whether appropriate statistics were used for this. The authors need to clarify exactly what procedure they undertook to compare the models, as they potentially employed an erroneous method (determining statistical significance based on the number of bootstraps instead of the number of observations) which may have led them to mistakenly claim better performance.
Given the relatively poor or equivocal performance of the brain age models and the relatively small sample sizes available, it is not clear that the modelling or dataset will have a big impact on the field. More accurate modelling methods are openly available, as are larger datasets. Nevertheless, the study is well-motivated and scientifically rigorous, so the results themselves are informative regarding the interrelationships of key Alzheimer's biomarkers and risk factors.
-
Reviewer #2 (Public Review):
Millar et al. have used multimodal brain magnetic resonance imaging (MRI) data from subjects with preclinical AD (cognitively normal with amyloid pathology), cognitive impairment, and matched cognitively normal (CN) subjects to predict the subject's age (i.e. brain age). To do so they have trained a Gaussian Process Regression (GPR) model using data from 3 datasets. The predicted age was then compared to the chronological age to calculate the brain age gap (BAG). Using resting-state functional MRI (rsfMRI) they calculated functional connectivity across 300 brain regions. Similarly, T1w MRI images have been used to calculate volumetric measures across 68 cortical and 33 subcortical regions. The results were then used as features in two separate models resulting in FC-BAG and Vol-BAG measures, respectively. A …
Reviewer #2 (Public Review):
Millar et al. have used multimodal brain magnetic resonance imaging (MRI) data from subjects with preclinical AD (cognitively normal with amyloid pathology), cognitive impairment, and matched cognitively normal (CN) subjects to predict the subject's age (i.e. brain age). To do so they have trained a Gaussian Process Regression (GPR) model using data from 3 datasets. The predicted age was then compared to the chronological age to calculate the brain age gap (BAG). Using resting-state functional MRI (rsfMRI) they calculated functional connectivity across 300 brain regions. Similarly, T1w MRI images have been used to calculate volumetric measures across 68 cortical and 33 subcortical regions. The results were then used as features in two separate models resulting in FC-BAG and Vol-BAG measures, respectively. A third model is then devised by "stacking" the previous model's predictions as features in a new model resulting in Vol+FC-BAG. The models were then applied to the test dataset and BAG measures were calculated for subjects in CN, preclinical AD, and cognitively impaired (CI) groups. All models show significantly higher BAGs for subjects with cognitive impairments. Finally, the authors have examined the relationship between BAGs measures and Amyloid Markers, Tau Markers, Neurodegeneration Markers, and Cognition.
Strengths:
The manuscript is very clearly written. The study is well designed and for the most parts the method section contains all the necessary information to replicate the steps. The sample size is comparable to similar studies investigating brain age in clinical populations and the inclusion and exclusion criteria are clearly stated. In order to avoid bias and overfitting in the predictive models the authors have (1) used a separate training (+validation) set and test set and (2) removed any subjects with potential pathology or impairment from the training set. Using data points on the plots as well as a combination of boxplot+violin plots makes the results clear and data distributions are provided when necessary. Furthermore, chronological age has been used as a covariate to correct the relationship between BAG and age, making the results more interpretable and reliable. Focusing on the stronger section of the results, the study shows a higher Vol-BAG (or Vol+FC-BAG) in subjects with CI, which is significantly related to higher Amyloid PET, higher PET and CSF-related Tau measures, and lower global cognition. In conclusion, the Vol-BAG results are clear and clinically relevant, based on a model with a reasonable prediction performance.
Weaknesses:
The manuscript follows authors' recently published work on FC-BAG in symptomatic and preclinical Alzheimer disease (Millar et al, Neuroimage 2022) by adding T1w volumetric measures from Freesurfer. Based on the results the additional value of rsfMRI connectivity is at best marginal. The FC-BAG model has a weak performance and is outperformed by Vol-BAG. The marginal benefit of adding FC-BAG to the Vol-BAG model is around 10% which comes with the additional cost of a new and more computationally demanding modality as well as making the biological relevance of the model almost untraceable. The preclinical findings reported as "Specifically, FC-BAG may capture a unique biphasic response to preclinical AD pathology" while potentially interesting are based on an unreliable model (FC-BAG) and can be a spurious finding. These results need further validation both robustness analysis within the current sample and in independent datasets. The other findings related to preclinical AD are based on the hippocampal volume which as the authors have mentioned in the discussion limitation is part of the features included in the Vol-BAG model and absent from FC-BAG.
In conclusion, the manuscript has clear findings based on the Vol-BAG model differentiating the cognitively impaired subjects from other groups and these results relate to the clinical severity of the disease as measured by Amyloid Markers, Tau Markers, and Cognition.
-
Reviewer #3 (Public Review):
In this manuscript, the authors aim to study the relationship between the Brain Age Gap (BAG) measure - based on functional connectivity and structural features - and different AD biomarkers such as amyloid, tau, cognition, and neurodegeneration in cognitively healthy and demented individuals. The main results showed increased BAG in cognitively impaired individuals. In this subgroup of individuals BAG models based on structural data were associated with more advanced AD pathology and lower cognitive performance. The BAG models based on fMRI data seem to show a U-curve in the health-disease continuum. The authors discuss the results in terms of a biphasic response of fMRI - while structural-based BAG would capture progression as well as highlight the advantages of multimodal data to understand health and …
Reviewer #3 (Public Review):
In this manuscript, the authors aim to study the relationship between the Brain Age Gap (BAG) measure - based on functional connectivity and structural features - and different AD biomarkers such as amyloid, tau, cognition, and neurodegeneration in cognitively healthy and demented individuals. The main results showed increased BAG in cognitively impaired individuals. In this subgroup of individuals BAG models based on structural data were associated with more advanced AD pathology and lower cognitive performance. The BAG models based on fMRI data seem to show a U-curve in the health-disease continuum. The authors discuss the results in terms of a biphasic response of fMRI - while structural-based BAG would capture progression as well as highlight the advantages of multimodal data to understand health and disease in healthy aging.
While the study has its merits such as the use of novel metrics, a decent sample with biomarkers and fMRI, etc., I believe some of the main conclusions of this paper are not fully substantiated by the results. The results based on the structural BAG model are solid (i.e., CI participants have older BA compared to healthy controls). However, I find the conclusions regarding the fMRI-BAG and the multimodal-BAG models are not fully supported by the results. The biphasic response of fMRI-BAG results - and the subsequent advantage of multimodal BAG - is based on p-values between .05 and .10 which have very low evidential value (e.g., Benjamin et al, 2018). I strongly discourage reporting these results as "marginal" and drawing assertive interpretations on this basis. Further, the poor performance of fMRI seems to add little information (in the stacked model) to the structural-only BA model.
The aim of the authors is to be commended, that is to take advantage of powerful machine learning methods and multimodal imaging to better understand the health-disease continuum in aging. This path is promising and can lead to both, better predictive tools and a better understanding of the aging brain. Further, the sample is good, from a single cohort, with multiple MRI modalities and biomarker information and the manuscript is easy to read and includes a very informative introduction. Also, it has some interesting findings such as in Fig. 2C and 2E where the graphs seem to show how BAG seems to be most useful at younger ages if used to predict dementia. Having said that, the "marginal" effects are central to the conclusions of this paper and are a critical caveat. Other methodological limitations of this paper are the parcellation used for structural BAG, which is relatively gross, a possible effect of motion on preprocessed functional connectivity, and the lack of multiple comparisons correction. Finally, the lack of a detailed description of the higher-level statistical analysis is detrimental to the clarity of the manuscript and leads to some confusion regarding the carried analyses.
-
