TL;DR
This paper introduces a novel method for predicting brain age by combining structural and functional brain imaging data, achieving higher accuracy than using either modality alone. This multimodal approach not only predicts age more precisely but also reveals that individuals with cognitive impairments show accelerated brain aging, suggesting its potential as a biomarker for early detection of neurocognitive disorders.
What Problem Does This Paper Address?
The aging brain exhibits changes in both its structure and function. Researchers are interested in developing "brain age" prediction models, which estimate a person's biological brain age based on imaging data. The discrepancy between this predicted brain age and a person's chronological age (often termed "brain age gap" or "brain aging") can potentially serve as a sensitive biomarker for various neurological conditions, especially those that emerge later in life. Previous studies have primarily relied on either structural (e.g., cortical thickness) or functional (e.g., resting-state fMRI connectivity) data for age prediction. However, these modalities capture different aspects of brain health, and it's unclear whether combining them would lead to more accurate predictions and a more robust measure of age-related brain changes, particularly in the context of cognitive impairment.
Methods
The authors developed a two-level machine learning approach for age prediction.
- Feature Extraction: They extracted features from five different sources: two functional connectivity matrices (at different resolutions) and three structural measures (cortical thickness, cortical surface area, and subcortical volumes).
- Single-Source Models: For each of these five sources, they trained individual Support Vector Regression (SVR) models to predict age.
- Multi-Source Models: To leverage the complementary information from different modalities, they stacked the predictions from the single-source models. These stacked predictions were then fed into Random Forest (RF) regression models. They trained three types of multi-source models:
- Stacked-function: Combined predictions from the two functional connectivity matrices.
- Stacked-anatomy: Combined predictions from the three structural measures.
- Stacked-multimodal: Combined predictions from all five single-source models.
- Robustness and Generalization: They specifically tested the models' robustness against head motion (a common confound in MRI) and their ability to generalize to an independent dataset from a different site.
Datasets
- LIFE-Adult-Study: A large dataset from the Leipzig Research Centre for Civilization Diseases (N=2354, age 19-82) was used for training and testing the primary models.
- Enhanced Nathan Kline Institute - Rockland sample (NKI): A smaller, independent dataset (N=475) was used to evaluate the generalization performance of the models.
Key Insights and Ideas
- Multimodal Fusion is Key: Combining different types of brain imaging data significantly improves the accuracy of brain age prediction compared to using any single modality. This suggests that structural and functional information capture distinct, yet complementary, aspects of brain aging.
- Brain Aging as a Biomarker for Cognitive Impairment: The discrepancy between predicted and chronological brain age (brain aging) is directly related to the severity of cognitive impairment. This finding is crucial for identifying individuals at risk or those experiencing decline.
- Robustness to Confounds: The developed multimodal models are not unduly influenced by head motion, a common artifact in MRI data. This is a critical step towards reliable clinical application.
- Generalizability Matters: While models perform best on data from the same site, training on larger and more diverse datasets improves their ability to generalize to new, unseen data from different sites.
Key Results
- The stacked-multimodal model achieved the lowest mean absolute prediction error (MAPE) of 4.29 years on the test set, outperforming single-modality models and the stacked-function/anatomy models.
- Subjects with mild or major Objective Cognitive Impairment (OCI) showed significantly advanced brain aging compared to OCI-norm subjects, with the effect being most pronounced in the stacked-multimodal model.
- Head motion did not drive the age prediction models; in fact, attempts to aggressively remove motion artifacts sometimes reduced prediction accuracy, suggesting that motion regression might remove meaningful age-related variance.
- Models trained on the LIFE dataset showed a significant drop in performance when tested on the independent NKI dataset (MAPE increased from ~4.3 to ~8.0 years), but this gap narrowed when training data included a small subset from the NKI site, highlighting the benefit of diverse training data.
Limitations and Future Work
- Individual Differences: The authors acknowledge that some prediction error will always persist due to inherent individual differences in brain structure and function.
- White Matter Anatomy: Future studies could explore incorporating white matter anatomy to further refine prediction accuracy.
- Pathology-Specific Models: While the multimodal approach is best for general age prediction, specific pathologies might be better characterized by individual modalities. Future work could investigate the effectiveness of different uni- and multimodal approaches for specific diseases.
- Predicting Cognitive Performance Directly: The authors suggest that directly predicting future cognitive performance, rather than brain age, could be a more targeted approach for identifying individuals at risk and developing interventions.
- Standardization of MRI Protocols: Further standardization of MRI acquisition protocols across sites could improve model generalizability.
- Head Motion in Structural MRI: While the study addressed head motion in functional MRI, its impact on structural MRI and how to best account for it in existing datasets remains an area for further investigation.
Summary updated