TL;DR
This systematic review of multimodal brain age studies reveals that while combining different neuroimaging techniques (multimodal models) significantly improves the accuracy of predicting biological age, this enhanced accuracy doesn't always translate to greater clinical utility. The findings suggest that the choice of imaging modality and the specific phenotype being studied are crucial for determining a brain age model's usefulness.
What Problem Does This Paper Address?
Predicting a person's "brain age" from neuroimaging data is a promising approach for understanding brain health and identifying risks for adverse outcomes. Brain age models aim to estimate an individual's biological age based on brain structure and function, and the difference between this estimated age and chronological age (brain age gap) can be a sensitive marker for various health conditions. Initially, these models relied on a single type of neuroimaging data (unimodal), but more recent research has explored combining data from multiple sources (multimodal) to improve predictions. However, there's a lack of clear guidance on which neuroimaging features and combinations are best for both accurately predicting age and being clinically useful. This paper addresses this gap by systematically reviewing existing multimodal brain age studies to understand how different approaches impact model accuracy and utility.
Methods
The researchers conducted a systematic review of multimodal brain age studies, adhering to PRISMA guidelines. They searched four major databases for relevant literature and screened a large number of articles to identify studies that compared multiple brain age models using different imaging modalities.
Key methodological aspects include:
- Quantitative Synthesis: They aggregated data from included studies to perform quantitative analyses on model accuracy (measured by mean absolute error) and model utility (measured by the correlation between brain age gaps and phenotypes).
- Min/Max Scaling: To account for variations in sample size and feature dimensionality across studies, they min/max scaled accuracy and dimensionality metrics within each study. Accuracy was reverse-scored to align with utility (higher score = better).
- Mixed-Effects Models: These models were used to assess the impact of imaging modality and model dimensionality on accuracy and utility, while accounting for study-level variations.
- Phenotype-Specific Analysis: For utility, they specifically examined how different phenotypes (e.g., cognitive function, chronic brain disorders) were associated with brain age gaps across various model types.
Datasets
The paper does not introduce new datasets. Instead, it synthesizes findings from existing studies that utilized a variety of neuroimaging datasets. The most popular datasets mentioned are:
- Cambridge Centre for Ageing and Neuroscience (CamCAN) dataset
- UK Biobank
The studies included in the review used data from various MRI sequences (T1-weighted, T2-weighted, diffusion-weighted), positron emission tomography (PET), magnetoencephalography (MEG), and electroencephalography (EEG).
Key Insights and Ideas
- Accuracy vs. Utility Divergence: The most striking insight is that increased model accuracy (better age prediction) does not automatically lead to increased clinical utility (better detection of health-related differences). This challenges the common assumption that the "best" model is simply the most accurate one.
- Multimodal Advantage for Accuracy: Combining data from multiple neuroimaging modalities consistently leads to more accurate brain age predictions compared to using a single modality.
- Unimodal Strengths for Utility: While multimodal models excel at accuracy, unimodal models, particularly those using functional magnetic resonance imaging (fMRI), showed greater sensitivity to a broader range of phenotypes.
- Chronic Brain Disorders as a Sweet Spot: Brain age models, especially multimodal ones, are particularly effective at detecting differences in chronic brain disorders like Alzheimer's disease and schizophrenia. This is likely due to the progressive nature of these conditions and their widespread impact on brain structure.
- Feature Selection Matters: The choice of neuroimaging features significantly impacts both accuracy and utility. Anatomical features (e.g., cortical thickness, volume) and white matter microstructure are generally more accurate predictors of age than functional features (e.g., functional connectivity). However, functional features might be more sensitive to certain transient or state-like phenotypes.
- Dimensionality's Role: Increasing the number of features generally improves model accuracy, but this effect is less pronounced for model utility.
Key Results
- Accuracy: Multimodal brain age models were the most accurate in 18 out of 20 studies reviewed. Accuracy generally increased with the number of features used, but this effect was less pronounced than the benefit of using multimodal data.
- Utility: Model utility varied considerably across studies and phenotypes. No single set of features emerged as universally superior for utility.
- Multimodal models showed the strongest utility for detecting differences in chronic brain disorders (e.g., Alzheimer's disease, schizophrenia).
- fMRI-based models were more sensitive to a broader array of phenotypes.
- The number of features did not significantly impact model utility.
- Accuracy and Utility Unrelated: The study found no significant correlation between model accuracy and model utility, reinforcing the idea that these are distinct performance metrics.
Limitations and Future Work
- Heterogeneity: The review acknowledges significant heterogeneity in study designs, sample characteristics, and methodological choices, which can make direct comparisons challenging.
- Statistical Power: For some specific phenotypes, the analysis was based on data from only a single study, limiting statistical power and the ability to draw definitive conclusions.
- Phenotype Specificity: While multimodal models are good for chronic brain disorders, improving their utility for other phenotypes remains an open question.
- Future Directions:
- Researchers need to move beyond solely optimizing for accuracy and also rigorously evaluate model utility across diverse phenotypes.
- Exploring novel approaches, such as applying fitness functions to brain age gaps to maximize separability, could enhance utility.
- Investigating higher-order brain network features (e.g., graph theory indices) might offer further improvements.
- More research is needed to understand how to systematically improve the utility of brain age models for a wider range of clinical applications.
Summary updated