Multimodal MRI-based neuromarkers trace longitudinal changes in cognitive functioning in ADHD

This study demonstrates that multimodal MRI-based neuromarkers derived from the Oregon ADHD-1000 dataset can accurately predict and longitudinally track both interindividual differences and intraindividual changes in cognitive functioning among children with and without ADHD, highlighting their potential for prognosis and treatment monitoring.

The Gist

The Big Picture: Reading the Brain's "Weather Report"

Imagine the human brain is like a complex, bustling city. Sometimes, the city runs smoothly; other times, there's traffic jams, construction, or power outages. For children with ADHD (Attention Deficit/Hyperactivity Disorder), the city often has unique patterns of traffic and noise that make it harder to focus or sit still.

For a long time, scientists have tried to take a single snapshot of this city to diagnose ADHD. But this study asked a different question: Can we watch the city over time to see how it changes?

The researchers wanted to build a "weather forecast" for the brain. Instead of just saying, "It's raining today" (a diagnosis), they wanted to predict, "The rain is getting heavier, or the sun is coming out" (how a child's thinking skills are improving or changing over time).

The Ingredients: A "Multimodal" Smoothie

To make this prediction, the scientists didn't just look at one thing. They blended different types of brain scans into a "super-smoothie" (which they call multimodal MRI).

1. The Structure (sMRI): This is like looking at the blueprints and buildings of the city. How thick are the walls? How big is the library (subcortical volume)?
2. The Activity (fMRI): This is like listening to the radio waves and traffic flow while the city is resting. Who is talking to whom? How synchronized are the neighborhoods?
   • They looked at standard "traffic flow" (Functional Connectivity).
   • They also looked at "local chatter" (ReHo) and "signal strength" (ALFF), which are often ignored but turned out to be helpful.

They fed all this data into a smart computer (Machine Learning) to see if it could guess a child's general thinking ability (called "g" or general intelligence).

The Results: A Crystal Ball That Works for Everyone

The computer learned to predict how well a child would do on thinking tests just by looking at their brain scans. Here is what they found:

• It Works for Everyone: The "weather forecast" was accurate for both children with ADHD and children without it. It didn't matter if you had the diagnosis or not; the brain patterns predicted thinking skills equally well.
• It Tracks Change (The "Intra" Magic): This is the most exciting part. Most studies only compare Person A to Person B. This study watched Person A over time.
  • Analogy: Imagine a gardener. Most studies take a photo of a rose bush and say, "This one is taller than that one." This study watched the same rose bush grow from spring to summer. They found that the brain scans could predict 33% of how much a child's thinking skills would improve or change as they got older.
• It Follows the Seasons (Age): As children get older, their brains naturally mature, and their thinking gets sharper. The study showed that the brain scans could track about 60% of this natural "growing up" process.

The Connection to Symptoms: Why Do They Act That Way?

ADHD has two main symptoms: Inattention (daydreaming, losing focus) and Hyperactivity (fidgeting, running around).

The researchers wanted to know: Does the brain scan explain why these symptoms happen?

• Hyperactivity: The brain scans explained almost everything about the link between hyperactivity and thinking skills. If a child's brain activity changed, their hyperactivity and thinking skills changed right along with it.
• Inattention: The scans explained about 26% of the link. Interestingly, inattention seemed more "stable" (like a personality trait) and didn't change as much day-to-day as hyperactivity did, so the brain scans had a harder time tracking those tiny daily shifts.

Why Does This Matter?

Think of this research as moving from a static map to a live GPS.

1. Better Prognosis: Instead of just saying, "You have ADHD," doctors might one day say, "Based on your brain's current 'traffic patterns,' your thinking skills are likely to improve by X amount over the next year."
2. Treatment Monitoring: If a child starts a new medication or therapy, doctors could scan their brain to see if the "traffic" is actually clearing up, rather than just waiting to see if the child behaves better in class.
3. Personalized Care: It proves that we can track the unique journey of one person's brain, not just compare them to a crowd.

The Bottom Line

This study is like building a high-tech dashboard for the brain. It shows that by combining different types of brain scans, we can create a tool that doesn't just diagnose a problem, but tracks the story of a child's development, helping us understand how their brain grows and how to best support them along the way.

Technical Summary

1. Problem Statement

Current machine learning research in Attention Deficit/Hyperactivity Disorder (ADHD) predominantly focuses on interindividual prediction (distinguishing between individuals, e.g., diagnosis or stratification). There is a significant gap in understanding whether neuroimaging markers can predict intraindividual variation (longitudinal changes within the same person over time).

• The Gap: Most studies rely on cross-sectional designs, limiting their utility for prognosis and treatment monitoring.
• The Objective: To determine if multimodal structural (sMRI) and functional (fMRI) markers can predict longitudinal cognitive trajectories, age-related cognitive development, and the dynamic link between cognitive functioning and ADHD symptoms (inattention and hyperactivity) in children with and without ADHD.

2. Methodology

Dataset and Participants

• Source: Oregon ADHD-1000 study, a large-scale longitudinal dataset.
• Sample: $N = 594$ participants (ages 8–17), comprising 1,053 annual timepoints (mean 1.77 timepoints per participant).
• Groups: Includes controls, clinically diagnosed ADHD, sub-clinical ADHD, and mixed groups.
• Exclusions: History of epilepsy, IQ < 75, non-stimulant psychotropic medication use, and comorbidities like autism or psychosis (though depression was allowed as a trajectory outcome). Stimulant medication was paused 48 hours prior to assessment.

Data Processing & Feature Extraction

• Target Variable: A general cognitive factor ($g$) derived via Confirmatory Factor Analysis (CFA) from six cognitive tasks: N-Back accuracy, CPT d', Stop Signal Reaction Time (SSRT), Digit Span Forward/Backward, and WIAT word reading.
• Neuroimaging Features (7 Sets):
  • sMRI (4 sets): Cortical area, cortical thickness, subcortical volume, and total brain volume (derived via FreeSurfer).
  • fMRI (3 sets): Functional connectivity (FC), Regional Homogeneity (ReHo), and Amplitude of Low-Frequency Fluctuations (ALFF) (derived via XCP-D).
  • Parcellation: Glasser atlas (360 cortical regions) and FreeSurfer ASEG (19 subcortical regions).

Machine Learning Pipeline

The study employed a two-stage stacking approach to prevent data leakage and maximize predictive power:

1. First-Stage Models: Three distinct algorithms were trained on each of the 7 feature sets independently using nested 5-fold cross-validation:
   • Kernel Ridge Regression (KRR)
   • Partial Least Squares (PLS)
   • XGBoost
   • Result: 21 first-stage predictions per participant-timepoint.
2. Second-Stage (Stacked) Model: A Ridge Regression model combined the 21 first-stage predictions to generate the final cognitive estimate ($\hat{g}$).
3. Validation: Performance was assessed via out-of-sample Pearson correlation ($r$), $R^2$, MAE, and MSE.

Statistical Analysis for Longitudinal Dynamics

To disentangle interindividual (between-person) and intraindividual (within-person) effects, the authors used Linear Mixed-Effects (LME) models:

• Decomposition: Predicted $g$-scores were split into a participant-specific mean ($\hat{g}_{mean}$) representing interindividual variance, and a deviation from that mean ($\hat{g}_{deviation}$) representing intraindividual variance.
• Commonality Analysis: Used to quantify the shared variance between neuroimaging predictions, age, and ADHD symptoms (hyperactivity/inattention).
• Generalizability: Tested via ADHD-additive vs. ADHD-interactive LME models to see if the prediction accuracy differed by diagnosis.

3. Key Results

Predictive Performance

• Stacked Model Accuracy: The multimodal stacked model achieved an out-of-sample Pearson correlation of $r = 0.459$, significantly outperforming all single-feature or single-algorithm first-stage models.
• Feature Importance:
  • Functional Connectivity was the strongest predictor ($r \approx 0.39$), particularly the Frontoparietal, Cingulo-Opercular, and Default Mode networks.
  • ReHo and ALFF provided a statistically significant, albeit small, incremental improvement over models using only FC and sMRI.
  • sMRI: Cortical thickness was the most informative structural feature.
• Generalizability: The model performed comparably for children with and without ADHD (no significant interaction effect), suggesting the markers are transdiagnostic.

Variance Decomposition (Inter- vs. Intraindividual)

• Interindividual Variance: The model explained 29.09% of the variance in cognitive functioning between different individuals.
• Intraindividual Variance: The model explained 33.48% of the variance in cognitive changes within individuals over time.

Age-Related Development

• The neuroimaging markers captured 72.52% of the interindividual age-related variance and 60.87% of the intraindividual age-related cognitive development. This indicates the markers effectively track the natural maturation of cognitive function during adolescence.

Cognition–Symptom Relationships

• Hyperactivity: Showed significant associations with cognitive functioning at both inter- and intraindividual levels. The neuroimaging marker captured 58.79% of the total variance in this relationship (43.36% interindividual, 100% of the intraindividual component).
• Inattention: Associated with cognition only at the interindividual level. The marker captured 25.99% of this interindividual association.

4. Key Contributions

1. Longitudinal Validation: Demonstrated that s/fMRI markers can successfully predict within-person cognitive changes, a critical step toward using neuroimaging for treatment monitoring and prognosis, not just diagnosis.
2. Multimodal Stacking: Proved that integrating diverse features (FC, ReHo, ALFF, sMRI) via stacking significantly enhances predictive accuracy compared to unimodal approaches.
3. Transdiagnostic Utility: Established that these markers generalize across ADHD and neurotypical populations, supporting a dimensional view of psychopathology (RDoC framework).
4. Developmental Tracking: Showed that brain-based markers can capture the trajectory of age-related cognitive improvements in adolescents, including those with ADHD.

5. Significance and Implications

• Clinical Potential: These findings lay the groundwork for using MRI-based biomarkers to monitor treatment efficacy. If a treatment improves cognition, the model should detect the corresponding shift in the neuroimaging-derived $\hat{g}$ score.
• Research Paradigm Shift: The study challenges the field to move beyond cross-sectional "diagnosis" models toward longitudinal "trajectory" models.
• RDoC Alignment: By capturing the biological link between cognition and symptoms across the spectrum of typical to atypical development, the study supports the Research Domain Criteria (RDoC) framework.
• Limitations & Future Work: While promising, the predictive accuracy ($r=0.46$) leaves room for improvement. The authors suggest future integration of task-based fMRI (which typically yields higher correlations) and larger datasets to refine these neuromarkers for clinical application.

Conclusion: This study provides foundational evidence that multimodal MRI markers can longitudinally track cognitive functioning in ADHD, offering a viable path toward personalized prognosis and treatment monitoring.