Exploring Links between Brain Image-Derived Phenotypes and Accelerometer-Measured Physical Activity in the UK Biobank

Using UK Biobank data, this study reveals that objectively measured physical activity is strongly associated with motor- and attention-related brain features and offers superior predictive utility for cardiometabolic diseases compared to neuroimaging measures alone.

The Gist

Imagine your brain as a massive, bustling city. This city has two main things we care about: the roads (how different neighborhoods talk to each other) and the buildings (how big and healthy the neighborhoods themselves are).

For a long time, scientists have known that if you don't move your body, the city starts to crumble. But they didn't know exactly which roads were getting paved over or which buildings were shrinking because of a lack of movement.

This study is like a massive detective investigation using data from 100,000 people in the UK. The researchers put wristwatches (accelerometers) on everyone to track exactly how much they moved, sat, and slept. Then, they took MRI scans of everyone's brains to see the "roads" (functional connectivity) and the "buildings" (gray matter volume).

Here is what they found, broken down into simple stories:

1. The "Movement Highway" Connection

The researchers used a special math tool (called Canonical Correlation Analysis) to see if there was a pattern linking how people moved to how their brains looked.

• The Finding: They found a strong, direct link between physical activity and the roads (brain connectivity), but a much weaker link to the buildings (brain size).
• The Analogy: Think of the brain's "roads" as a highway system. The study found that people who moved more had highways that were better organized, specifically in the Motor Control and Attention districts. It's like finding that people who jog regularly have perfectly smooth, well-lit roads connecting the "Gym District" to the "Focus District."
• The Surprise: The link to the "buildings" (brain size) was there, but it was faint. It's like saying, "Yes, moving helps the buildings stay standing, but it helps the traffic flow (the roads) much more dramatically."

2. The "Three Magic Metrics"

The study looked at 11 different ways to measure movement (like how many minutes you sat, how fast you walked, or how your activity changed from day to night). They wanted to know: Which specific movement habits matter most for brain health?

• The Finding: It turns out, you don't need to track everything. Just three specific habits told the whole story:
  1. Total Intensity: How much total energy you burned (the "volume" of movement).
  2. Peak Hours: How active you were during your most energetic 10 hours of the day.
  3. Day/Night Rhythm: How clearly you could tell the difference between your "active day" and your "resting night."
• The Analogy: Imagine trying to describe a person's fitness. You could list 100 different stats, but if you just tell me "how hard they worked," "how long they worked," and "whether they rested at night," I can guess almost everything else about their fitness. These three metrics were the "keys" that unlocked the secrets of the brain's health.

3. The "Crystal Ball" for Disease

Finally, the researchers asked: If we want to predict if someone will get sick (like diabetes, heart disease, stroke, or cancer), is it better to look at their brain scan or their activity watch?

• The Finding: The activity watch won, hands down.
• The Analogy: Imagine you are trying to predict if a car is going to break down.
  • Option A: You look at the engine's internal wiring (the brain scan).
  • Option B: You look at how much the car was driven and how well it was maintained (the activity data).
  • The Result: Option B was a much better predictor of whether the car would fail. Specifically, for Heart Disease and Diabetes, knowing how much a person moved was far more useful than looking at their brain structure. The brain scan added very little extra information.

The Big Takeaway

This study tells us that movement is a powerful language the brain understands.

When you move your body, you aren't just burning calories; you are essentially "paving the roads" in the parts of your brain that control movement and focus. While moving doesn't drastically change the size of your brain neighborhoods, it dramatically improves how those neighborhoods talk to each other.

Most importantly, if you want to know if someone is at risk for heart disease or diabetes, watching how they move is a better crystal ball than scanning their brain. It suggests that keeping an active, rhythmic lifestyle is one of the most direct ways to keep your brain's "highways" open and your body healthy.

Technical Summary

1. Problem Statement

Despite extensive evidence linking physical activity (PA) to neurodegenerative disease prevention and brain health, the specific multivariate relationships between high-resolution, objectively measured PA and complex brain imaging phenotypes remain poorly understood. Previous studies often focused on isolated brain regions or single PA metrics. There is a critical gap in understanding:

• How PA relates simultaneously to functional connectivity (FC) networks and gray matter volume (GMV) across distributed brain systems.
• The predictive capacity of PA for individual brain features.
• The joint utility of PA and brain imaging biomarkers in predicting the risk of prevalent diseases (diabetes, stroke, coronary heart disease, and cancer), specifically quantifying their relative contributions to explained variance.

2. Methodology

The study leveraged data from the UK Biobank, utilizing a subset of participants with both accelerometer data and multimodal brain imaging.

Data Sources:

• Physical Activity (PA): Measured via Axivity AX3 wrist-worn triaxial accelerometers over 8 days for ~106,000 participants. Data was processed into 11 summary variables reflecting activity volume (e.g., Total Log Acceleration, TLAC), intensity (Light vs. Moderate/Vigorous), fragmentation (bout transitions), and circadian rhythm (M10, L5, Relative Amplitude).
• Brain Imaging:
  • Functional Connectivity (FC): Derived from resting-state fMRI (rs-fMRI) using 55 independent components, resulting in 1,485 nodewise connections.
  • Gray Matter Volume (GMV): Derived from T1-weighted structural MRI across 139 anatomical Regions of Interest (ROIs).
• Covariates: Extensive adjustment for confounders including age, sex, head size, scanner position, motion parameters, and lifestyle factors (BMI, smoking, alcohol).

Analytical Framework:
The authors employed three distinct analytical approaches:

1. Bipartial Canonical Correlation Analysis (CCA): Used to identify low-dimensional axes of shared variation between PA variables and brain IDPs (FC and GMV separately) after regressing out confounders. The study focused primarily on the first canonical mode for stability.
2. Variable Importance Assessment:
   • Linear Regression: Tested the significance of individual PA variables in predicting each brain feature (FC connection or GMV ROI).
   • Machine Learning: Used Random Forest and Stepwise BIC (Bayesian Information Criterion) to rank PA variables based on their predictive importance and non-linear contributions.
   • Predictive Performance: Evaluated via 10-fold cross-validation (repeated 100 times) using Pearson correlation between observed and predicted values.
3. Disease Risk Prediction:
   • Nested Logistic Regression: Three models were compared for four diseases:
     • $M_1$: Full model (Confounders + PA + Brain IDP).
     • $M_2$: PA only (Confounders + PA).
     • $M_3$: Brain IDP only (Confounders + Brain IDP).
   • Metrics: Performance was assessed using ROC-AUC and McFadden's pseudo $R^2$ (to quantify the proportion of variation explained, addressing issues with imbalanced datasets).

3. Key Results

A. Multivariate Associations (CCA)

• PA vs. Functional Connectivity (FC): A statistically robust but low-dimensional association was found.
  • Canonical Correlation: $r = 0.50$ (Primary mode).
  • Brain Regions: The strongest contributions came from connections within and between control/attention and sensorimotor networks.
  • PA Drivers: Vigorous activity metrics (TLAC, M10, LIPA) were negatively correlated with the FC canonical variate, while sedentary measures (SBout) were positively correlated.
  • Variance: The primary mode explained ~34% of PA variance but only ~0.19% of FC variance, indicating PA aligns with a narrow subspace of brain functional variation.
• PA vs. Gray Matter Volume (GMV): The association was significantly weaker.
  • Canonical Correlation: $r = 0.19$.
  • Brain Regions: Contributions were small and diffuse, with the cerebellar vermis (associated with coordination) showing the strongest coefficients.
  • Variance: The primary mode explained ~40% of PA variance but only ~1.4% of GMV variance.

B. Variable Importance & Predictive Modeling

• Predicting Brain Features: PA variables successfully predicted a significant portion of brain features (66.6% of FC connections and 97.8% of GMV ROIs were statistically significant).
• Key Predictors: Across both FC and GMV, three variables consistently emerged as top predictors: TLAC (Total Log Acceleration), M10 (average acceleration during the 10 most active hours), and LIPA (Light Intensity Physical Activity).
• Network Specificity: PA was a stronger predictor for sensorimotor network connectivity and specific GMV regions (prefrontal cortex, hippocampus, amygdala) than for other networks.

C. Disease Risk Prediction

• Performance Metrics:
  • AUC: Models performed best for Coronary Heart Disease (CHD) and Diabetes (AUC $\approx$ 0.80), followed by Stroke ($\approx$ 0.70) and Cancer ($\approx$ 0.64).
  • McFadden's $R^2$: This metric revealed that PA variables alone ($M_2$) contributed more to explaining disease variance than brain IDPs alone ($M_3$).
  • Incremental Value: Adding brain IDPs to PA models ($M_1$ vs $M_2$) provided only marginal improvements in predictive power.
  • Ranking: For all diseases, $R^2_{M1} > R^2_{M2} > R^2_{M3}$. The gap between PA-only and Brain-IDP-only models was largest for CHD and Diabetes.

4. Key Contributions

1. Multivariate Integration: First comprehensive analysis linking high-dimensional, objective PA data with both functional (FC) and structural (GMV) brain phenotypes simultaneously.
2. Identification of Specific Networks: Demonstrated that PA is most strongly associated with motor-related and attention-related brain networks, rather than global brain reorganization.
3. Predictive Hierarchy: Established that objective PA measurements are superior predictors of cardiometabolic disease risk (CHD, Diabetes) compared to cross-sectional neuroimaging biomarkers.
4. Methodological Rigor: Applied bipartial CCA and nested logistic regression with McFadden's $R^2$ to rigorously quantify the relative explanatory power of lifestyle vs. biological markers in large-scale cohorts.

5. Significance and Limitations

Significance:

• The findings suggest that behavioral indicators (PA) may be more immediate and powerful proxies for cardiometabolic health risks than static brain imaging measures in the general population.
• It validates the use of wearable accelerometry as a scalable tool for monitoring brain health and disease risk.
• The study highlights that the link between PA and the brain is low-dimensional, affecting specific networks (sensorimotor/attention) rather than the entire connectome.

Limitations:

• Cross-Sectional Design: PA and imaging were collected at different time points, preventing causal inference regarding whether PA changes brain structure or vice versa.
• Low Dimensionality: While the association is statistically robust, it explains a small fraction of total brain variance, meaning PA is not the sole driver of brain phenotypic variation.
• Cancer Prediction: The weak prediction for cancer is attributed to the heterogeneity of cancer types and the multifactorial nature of cancer etiology (genetics, environment) which PA alone cannot fully capture.

Conclusion:
The study concludes that objectively measured physical activity is strongly associated with a specific set of motor-related brain features and provides substantial predictive information for cardiometabolic diseases, often outperforming neuroimaging biomarkers in isolation. This supports the integration of lifestyle data into precision medicine models for disease prevention.