Feature consistency in transdiagnostic connectome-based models of sustained attention and autism symptoms

This study demonstrates that functional connectivity patterns predicting sustained attention and autism symptoms are phenotype-specific and vary significantly based on participant age and clinical diagnosis, highlighting the importance of considering these factors in transdiagnostic connectome-based models.

The Gist

Imagine your brain is a massive, bustling city with millions of roads (connections) linking different neighborhoods (regions). Some roads are for heavy traffic like the "Default Mode" (daydreaming), others are for sensory input like the "Visual" district, and some are for quick reactions like the "Motor" zone.

This paper is like a group of detectives trying to find the perfect map that explains how people keep their focus on a single task—like watching a movie or driving a car—without getting distracted. They wanted to know: Is there one single "Focus Map" that works for everyone, regardless of whether they are neurotypical, have autism, or have ADHD?

Here is the story of their investigation, broken down simply:

The Five Different Maps

The researchers didn't just look at one group of people. They gathered five different "maps" created by other scientists in the past. Each map was built to predict a specific behavior:

1. Map A (The Autism Map): Predicted how severe autism symptoms were in a group of kids.
2. Map B (The Youth Autism Map): Predicted autism symptoms in a different group of kids.
3. Map C (The Youth Attention Map): Predicted how well those same kids could focus on a task.
4. Map D (The Audio-Visual Map): Predicted focus in adults doing a listening/seeing task.
5. Map E (The Adult Attention Map): Predicted focus in neurotypical adults.

The Big Question: If you overlay these five maps, do they all show the exact same roads? Is there a "Core Focus Highway" that everyone uses?

The Investigation: What They Found

The researchers superimposed these maps on top of each other, like tracing paper, to see where the lines matched up.

1. The "No Perfect Overlap" Surprise
They hoped to find a single, perfect set of roads that every brain uses to stay focused. They didn't find it.

• Analogy: Imagine five different people giving you directions to the same coffee shop. They all agree on the general area (downtown), but they all take slightly different streets. One person turns left at the bakery; another turns right at the park. There is no single "perfect route" that everyone takes.
• The Twist: Two of the maps were made from almost the exact same group of kids (90% overlap!). But because one map was trying to predict "Autism Symptoms" and the other was trying to predict "Focus Scores," the roads they highlighted were completely different. It's like two chefs using the same kitchen and ingredients but making a soup vs. a cake; the tools and steps look nothing alike.

2. The "Neighborhoods" That Do Agree
While the specific roads (connections) weren't identical, the neighborhoods (brain networks) were often the same.

• The Pattern: Most maps agreed that the Visual district (seeing), the Motor district (moving hands), and the Heteromodal district (complex thinking) were crucial.
• Analogy: Even if the specific streets differ, all the maps agree that to get to the coffee shop, you have to pass through the "Downtown" and "Park" neighborhoods. You can't get there without them.
• Why? To stay focused, you need to see what's happening, move your hand to respond, and use your brain to process it all.

3. The "Age and Diagnosis" Filters
The researchers found that the maps looked more similar when the people were similar.

• Age: Maps made from kids looked more like other kid-maps. Maps from adults looked more like other adult-maps.
  • Analogy: A map of a city designed for toddlers (who walk slowly and need hand-holding) looks very different from a map designed for teenagers (who run and take shortcuts). The "Focus Highway" changes as you grow up.
• Diagnosis: Maps predicting autism symptoms looked different from maps predicting attention scores.
  • Analogy: If you are trying to navigate a city with a broken leg (autism symptoms), the path you take is different than if you are just trying to run a race (attention task). The brain adapts its "traffic flow" based on the specific challenge.

The Big Takeaway

The main lesson of this paper is that the brain is flexible, not rigid.

There isn't one single "Focus Switch" or a universal "Attention Blueprint" that fits every human. Instead, the brain builds a unique "Focus Route" for every person, depending on:

• Who they are (their age).
• What they are dealing with (autism, ADHD, or neurotypical).
• What they are trying to do (watch a movie, solve a puzzle, or listen to a sound).

The "So What?"
This is actually good news for science. It means we can't just use one "one-size-fits-all" test to diagnose attention problems or autism. We need to understand that the brain's wiring is a personalized, dynamic system. To help someone focus, we might need to look at their specific "road map" rather than assuming they are driving the same route as everyone else.

In short: Focus is a journey, and everyone takes a slightly different path to get there.

Technical Summary

1. Problem Statement

Sustained attention is a critical neurobiological process, yet its neural correlates remain poorly understood, particularly in transdiagnostic contexts involving neurodevelopmental disorders like Autism Spectrum Disorder (ASD) and Attention-Deficit/Hyperactivity Disorder (ADHD). While these conditions share behavioral and neurobiological features (e.g., executive dysfunction, sensory processing issues), it is unclear whether a core, consistent functional connectivity architecture underpins sustained attention across different populations, ages, and diagnostic statuses. Previous studies have produced connectome-based predictive models (CPM) for attention and autism, but the degree of overlap and generalizability of these models across datasets has not been systematically interrogated.

2. Methodology

The authors conducted a meta-analytic re-examination of five previously published, transdiagnostic connectome-based models.

• Datasets and Models:

  1. ABIDE Model: Resting-state fMRI predicting Social Responsiveness Scale (SRS) scores in youth with/without autism ($N=229$).
  2. Yale Youth Autism Model: GradCPT (task) fMRI predicting ADOS-2 scores in youth with/without autism ($N=63$).
  3. Youth Attention Model: GradCPT fMRI predicting task accuracy ($d'$) in the same youth cohort ($N=70$; shares ~90% subjects with Model 2 but predicts a different phenotype).
  4. avCPT Model: Audio-visual CPT fMRI predicting task accuracy in young adults ($N=43$).
  5. Adult Attention Model: GradCPT fMRI predicting task accuracy in neurotypical adults ($N=25$).

• Data Processing & Harmonization:

  • All datasets were processed using a consistent pipeline aligned to a shared 268-node functional atlas.
  • Connectome Generation: Pearson correlations between all node pairs created $268 \times 268$ matrices.
  • Feature Standardization: To control for varying network sizes across original studies, the authors re-analyzed the raw data to extract the top 1,000 positively and negatively correlated edges for each model (totaling 2,000 edges per model). Robustness checks were performed with 500 and 2,500 edge thresholds.
  • Directionality Adjustment: Due to the inverse relationship between autism symptoms and attention performance, the "positive" networks of autism models were compared against the "negative" networks of attention models (and vice versa) to align predictive directions.

• Statistical Analyses:

  • Edge Overlap: Hypergeometric cumulative density functions (CDF) and permutation testing (1,000 iterations) were used to assess the statistical significance of shared edges between models.
  • Network Consistency: Contribution of canonical functional networks (e.g., Default Mode, Frontoparietal) was quantified and corrected for network size.
  • Representational Similarity Analysis (RSA): Used to correlate feature importance vectors across datasets. A multiple linear regression model tested how age, scan type (rest vs. task), behavioral measure (attention vs. autism), and autism diagnosis prevalence predicted the similarity between brain-behavior vectors.

3. Key Contributions

• Systematic Cross-Study Comparison: This is the first study to directly compare five distinct CPM models derived from different sites, scanners, tasks, and phenotypes to test for a "core" attention network.
• Phenotype-Specificity: The study demonstrates that while broad network trends exist, specific predictive features are highly sensitive to the target phenotype (attention vs. autism) rather than being universal markers of sustained attention.
• Methodological Rigor: By re-processing raw data to standardize feature selection (top $N$ edges) and using RSA on unthresholded data, the authors isolated the effects of demographic and phenotypic variables from methodological artifacts.

4. Key Results

• Lack of a Universal Core Network: No single set of edges or network pairs was shared across all five models.
  • Only 408 edges (positive) and 423 edges (negative) were shared by at least two models.
  • No edges were shared across all five models.
  • Notably, the two models derived from the same ~90% overlapping subject pool (Yale Youth Autism vs. Youth Attention) did not share statistically significant edges, despite the high subject overlap. This suggests the predictive features are driven by the phenotype being predicted, not the sample itself.
• Phenotype-Driven Similarity:
  • Models predicting attention phenotypes (Models 3, 4, 5) shared significantly more edges and network pairs with each other than with models predicting autism phenotypes (Models 1, 2).
  • Models predicting autism symptoms shared specific network pairs involving heteromodal association networks.
• Network Architecture:
  • Positive Networks: Dominated by Visual, Motor, and Default Mode networks.
  • Negative Networks: Dominated by Visual, Default Mode, and Cerebellar networks.
  • The results support the Sensorimotor-Association (SA) axis, indicating that sustained attention relies on coordination between sensory/motor and higher-order association networks.
• Factors Influencing Similarity (RSA):
  • Age: Positively associated with model similarity (older samples were more similar).
  • Autism Diagnosis: Negatively associated with similarity (samples with higher proportions of autism diagnoses were less similar to each other).
  • Scan Type & Behavioral Measure: These were not statistically significant predictors of similarity in the regression model, though pairwise correlations showed that congruent phenotypes (e.g., attention vs. attention) yielded higher similarity than incongruent ones.

5. Significance and Implications

• Transdiagnostic Complexity: The findings challenge the notion of a single, universal "sustained attention network." Instead, the neural architecture supporting attention appears to be phenotype-specific and modulated by developmental stage (age) and clinical diagnosis.
• Clinical Generalizability: Predictive models trained on one specific phenotype (e.g., autism symptoms) may not generalize well to others (e.g., attention deficits), even within the same individuals. This highlights the need for precision medicine approaches that account for specific symptom profiles rather than broad diagnostic categories.
• Neurodevelopmental Insights: The negative association between autism diagnosis and model similarity suggests that the heterogeneity of autism subtypes creates diverse neural signatures, complicating the search for a unified biomarker.
• Future Directions: The study underscores the importance of large-scale, prospective studies that control for sociodemographic factors (SES, environment) and utilize longer scan times to disentangle the complex interplay between development, diagnosis, and attention.

In conclusion, while distributed patterns involving visual and heteromodal networks are consistently involved in sustained attention, the specific functional connections are not invariant. They vary significantly based on what is being predicted (phenotype), who is being studied (age/diagnosis), and the specific behavioral context.