Multimodal MRI prediction of cognitive functioning across the lifespan: separating between-person differences from within-person changes

This study utilizing longitudinal data from 450 adults demonstrates that while a multimodal MRI marker combining five imaging techniques effectively predicts stable between-person differences in cognitive functioning, it has limited capacity to capture within-person changes over time.

The Gist

Imagine your brain is a massive, bustling city. For decades, scientists have tried to understand how the "traffic" (your thoughts and memories) and the "roads" (your brain's structure) change as the city ages.

This paper is like a massive traffic report that tries to answer two very different questions:

1. The "Who" Question: Why is Person A's city generally more efficient than Person B's city, even when they are the same age? (Between-person differences).
2. The "When" Question: Why does Person A's city get a little more clogged up this year compared to last year? (Within-person changes).

The researchers used a powerful tool called MRI (a camera that takes pictures of the brain) to try and predict how well a person's "city" is running. They didn't just take one picture; they took five different types of pictures (modalities) to get a full 3D view of the brain.

Here is the breakdown of their findings in simple terms:

1. The Five Camera Angles (The Modalities)

Think of these as five different ways to photograph the brain city:

• DWI (Diffusion Imaging): This is like a satellite map of the highways. It shows the white matter tracts (the roads) connecting different parts of the brain.
• sMRI (Structural MRI): This is like a blueprint of the buildings. It measures the size and thickness of the brain's gray matter.
• FC (Functional Connectivity): This is like a phone call log. It shows which parts of the brain are "talking" to each other when you are just resting.
• Task-fMRI: This is like a security camera during a specific event. It watches the brain while you are doing a specific job (like looking at pictures or solving word puzzles).
• ASL (Arterial Spin Labeling): This is like a fuel gauge. It measures blood flow (the fuel) going to the brain.

2. The Prediction Game

The researchers built a "Super Computer" (Machine Learning) to look at these pictures and guess a person's Cognitive Score (how smart, quick, and memory-foolish they are).

• The Winner: When they combined all five camera angles into one giant model, it was the best predictor. It was like having a detective who uses a map, a blueprint, a phone log, a security camera, and a fuel gauge all at once.
• The Runner-Up: The "Highway Map" (DWI) was surprisingly the best single camera. It turns out, the integrity of the brain's "roads" is a huge clue to how well the city runs.
• The Loser: The "Fuel Gauge" (ASL) was the worst at predicting anything. The signal was too fuzzy, like trying to read a speedometer through a foggy windshield.

3. The Big Discovery: "Stable Traits" vs. "Daily Changes"

This is the most important part of the paper. The researchers realized their "Super Computer" was great at one thing but not so great at another.

• The "Who" (Between-Person): The model was excellent at predicting why Person A is generally smarter than Person B. It could explain about 60% of the differences between people.
  • Analogy: If you look at a photo of a city, you can easily tell which city is built better and which is falling apart. The MRI is great at spotting these permanent, structural differences.
• The "When" (Within-Person): The model was weak at predicting why Person A's brain got slightly worse (or better) this year compared to last year. It only explained about 17% of these changes.
  • Analogy: If you look at the same city today versus five years ago, the changes are subtle. The "roads" don't change much in a healthy person over five years. The MRI is like a high-resolution camera that is great at spotting the big cracks in the foundation, but it's not sensitive enough to see the tiny potholes that form in a single year.

4. The Age Factor

The researchers also asked: "How much of this is just because people are getting older?"

• Between People: Yes, older people generally have lower scores than younger people. The MRI markers captured almost all of this age-related difference (about 95%).
• Within People: Even though the MRI wasn't great at spotting small yearly changes, it did capture about 55% of the age-related decline that did happen within individuals.

The Bottom Line (The Takeaway)

Think of these MRI markers as a medical check-up tool:

• Great for Diagnosis (The "Who"): If you want to know if a patient is at high risk for cognitive decline compared to their peers, or if you want to group patients into "high functioning" vs. "low functioning" categories, these MRI scans are fantastic. They give you a very clear picture of the person's baseline.
• Not Great for Monitoring (The "When"): If you want to use these scans to track if a treatment is working right now or to see if a healthy person's brain is starting to slip this year, the current technology isn't sensitive enough yet. The changes are too small for the camera to catch clearly in healthy people.

In short: These brain scans are like a very good map of a city's permanent layout, but they aren't quite sensitive enough to track the daily traffic jams or the slow, subtle wear and tear of a single year. To track those small changes, we need even sharper cameras and data from people who are actually experiencing significant decline (like those with dementia), not just healthy aging adults.

Technical Summary

1. Problem Statement

While brain MRI has shown promise for predicting cognitive functioning, a critical gap exists in understanding whether these predictions rely on stable between-person differences (useful for diagnostic stratification) or longitudinal within-person changes (useful for prognosis and monitoring decline). Most existing predictive studies are cross-sectional, failing to distinguish these two sources of variance. Furthermore, it remains unclear how different MRI modalities (structural, functional, diffusion, perfusion) contribute to capturing these distinct variance components across the adult lifespan.

2. Methodology

Dataset and Participants:

• Source: Dallas Lifespan Brain Study (DLBS), a longitudinal dataset.
• Sample: 450 healthy adults (aged 21–90).
• Design: Up to three waves of data collection, spaced approximately 5 years apart.
• Exclusion: Participants with major psychiatric/neurological disorders, substance abuse history, or significant medical conditions were excluded to focus on healthy aging.

Target Variable (Cognitive Functioning):

• A general cognitive factor (g-score) was derived using Exploratory Structural Equation Modeling (ESEM).
• Based on 11 cognitive tests covering working memory/reasoning, episodic memory, and processing speed.
• The g-score served as the ground truth for all machine learning predictions.

Neuroimaging Features (37 Phenotypes across 5 Modalities):

1. Diffusion-Weighted Imaging (DWI): Structural connectivity (streamline counts/lengths via SIFT2) and tractometry (FA, MD) using Schaefer and Gordon atlases.
2. Structural MRI (sMRI): Cortical thickness, surface area, grey matter volume, subcortical volumes (FreeSurfer), and global brain metrics.
3. Functional Connectivity (FC): Resting-state and task-based residual connectivity (correlations in BOLD time series after removing task effects).
4. Task-fMRI: Event-related BOLD activation contrasts from "Scenes" (episodic memory) and "Words" (semantic processing) tasks.
5. Arterial Spin Labeling (ASL): Cerebral blood flow (CBF) maps using BASIL processing.

Machine Learning Workflow:

• Strategy: Nested 5-fold cross-validation. Crucially, all waves from a single participant were kept together in either the training or test set to prevent data leakage.
• Level 1 (Phenotype-specific): Trained 5 algorithms (Random Forest, Elastic Net, XGBoost, PLS, Kernel Ridge Regression) on each of the 37 feature sets individually.
• Level 2 (Stacking):
  • Modality-specific stacking: Combined predictions from features within a single modality.
  • Multimodal stacking: Combined predictions from all 37 features across all 5 modalities.
• Statistical Analysis:
  • Linear Mixed-Effects (LME) Models: Used to decompose variance into between-person (mean across waves) and within-person (deviation from individual mean) components.
  • Variance Decomposition: Quantified how much MRI markers explained within- vs. between-person variance in the g-score.
  • Commonality Analysis: Partitioned variance to determine how much of the age-cognition relationship was captured uniquely by MRI vs. shared with age.

3. Key Contributions

• Longitudinal Benchmarking: First study to systematically benchmark multimodal MRI prediction of cognition while explicitly separating between-person stability from within-person change in a healthy aging cohort.
• Variance Decomposition Framework: Introduced a rigorous statistical approach to quantify the specific utility of MRI markers for stratification (between-person) vs. monitoring (within-person).
• Multimodal Stacking Validation: Demonstrated that combining diverse neuroimaging modalities significantly outperforms single-modality approaches.
• Age-Cognition Disentanglement: Quantified the extent to which MRI markers capture age-related cognitive variance at both the between- and within-person levels.

4. Key Results

Predictive Accuracy:

• Multimodal Stacking: Achieved the highest accuracy ($R^2 = 0.51$, Pearson $r = 0.73$).
• Modality Hierarchy:
  1. DWI: Strongest single modality ($R^2 \approx 0.47$), driven by structural connectivity (streamline count/length).
  2. FC: Improved significantly with stacking ($R^2 \approx 0.46$), nearly matching DWI. Task-based FC outperformed resting-state FC.
  3. sMRI: Strong performance ($R^2 \approx 0.43$), particularly driven by subcortical volumes and global metrics.
  4. Task-fMRI: Moderate performance ($R^2 \approx 0.43$).
  5. ASL: Weakest predictor ($R^2 \approx 0.13$), likely due to low signal-to-noise ratio.

Variance Decomposition (Between vs. Within):

• Asymmetry: MRI markers were highly effective at capturing between-person differences but modest at capturing within-person changes.
  • Between-person variance explained: Up to 60.3% (Multimodal model).
  • Within-person variance explained: Up to 17.2% (sMRI global variables); Multimodal model explained only 6.3%.
• Interpretation: In this healthy cohort, cognitive functioning is relatively stable over time (ICC = 0.88), making within-person change difficult to detect.

Commonality Analysis (Age-Related Variance):

• Between-person Age: The multimodal MRI marker captured 94.92% of the between-person age-related variance in cognition.
• Within-person Age: The multimodal MRI marker captured 54.72% of the within-person age-related variance.
• Interaction: MRI markers explained a small but significant portion of the interaction between age and cognitive decline.
• ASL Exception: ASL showed minimal overlap with age-related variance, consistent with its poor predictive performance.

5. Significance and Implications

• Clinical Utility Distinction: The study clarifies that current multimodal MRI models are excellent tools for phenotyping and risk stratification (identifying individuals with high/low cognitive baselines) but are currently limited for individual-level monitoring of cognitive decline in healthy populations due to the low magnitude of within-person change.
• Modality Selection:
  • For stratification: DWI and FC are the most efficient and robust choices.
  • For longitudinal tracking: sMRI features showed the highest sensitivity to within-person change, though the signal remains subtle.
• Future Directions: To improve within-person prediction, future studies require datasets with greater cognitive variability (e.g., preclinical dementia) and potentially more cognitively demanding fMRI tasks. The findings also highlight the need for improved ASL acquisition protocols to make perfusion data viable for cognitive prediction.
• Generalizability: The study acknowledges the "WEIRD" (Western, Educated, Industrialized, Rich, Democratic) nature of the sample, suggesting that findings regarding the magnitude of within-person change may differ in more diverse or clinical populations.

In summary, this paper provides a critical framework for interpreting MRI-based cognitive predictions, emphasizing that high predictive accuracy does not automatically equate to sensitivity for tracking longitudinal change, and that the utility of a marker depends heavily on the specific clinical or research question (diagnosis vs. prognosis).