Co-Activation Patterns Characterize Early Resting-State Networks in Newborn Infants: A High-Density Diffuse Optical Tomography Study

This study applies co-activation pattern (CAP) analysis to high-density diffuse optical tomography (HD-DOT) data from sleeping newborns to reveal distinct, transient cortical network states and early frontoparietal connectivity that are not detectable through conventional static functional connectivity methods.

The Gist

The Big Picture: Listening to a Baby's Brain "Chat"

Imagine you are trying to understand how a group of friends interacts at a party.

• The Old Way (Static Analysis): You take a photo of the whole room and count how many people are standing near each other. You get a general idea of who is friends with whom, but you miss the moments when they actually start talking, laughing, or arguing. You only see the average.
• The New Way (This Study): Instead of a photo, you record a video. You look for specific, high-energy moments—like when two friends suddenly start a deep conversation. You group these specific moments together to see what kind of "conversations" (brain networks) happen most often.

This study is the first time scientists have used this "video" approach (called Co-Activation Patterns or CAPs) on newborn babies using a special, baby-friendly camera.

---

The Problem: Why Can't We Just Use MRI?

Usually, to see inside a baby's brain, scientists use an MRI machine. But MRIs are like loud, scary tunnels.

• Babies move a lot.
• The machine is loud and isolating.
• To get a clear picture, the baby has to stay perfectly still, often requiring sedation (sleeping pills), which changes how the brain works.

The Solution: The researchers used a wearable hat called HD-DOT (High-Density Diffuse Optical Tomography).

• The Analogy: Think of this hat as a flashlight with many tiny beams. It shines safe, near-infrared light through the baby's thin skull to see how blood flows in the brain.
• It's quiet, comfortable, and can be used right next to the baby's crib while they sleep naturally.

---

The Method: Finding the "Spark"

The researchers recorded 44 sleeping newborns. They wanted to see how different parts of the brain (Frontal, Central, and Parietal) talk to each other.

1. The Seed: They picked a specific spot in the brain (like a "seed") and waited for it to get very active (like a lightbulb turning on bright).
2. The Snapshot: Whenever that seed lit up, they took a "snapshot" of the whole brain to see what else was happening at that exact second.
3. The Clustering: They took thousands of these snapshots and used a computer algorithm (K-means) to sort them into groups.
   • Analogy: Imagine you have a pile of 1,000 photos of a party. You sort them into piles: "The Dancing Group," "The Deep Conversation Group," and "The Nap Group."
   • In the brain, these piles are called Co-Activation Patterns (CAPs).

---

The Discoveries: What Did They Find?

1. The Brain is a "Flash Mob," Not a Static Map

The study found that the baby's brain isn't just one steady network. It's more like a flash mob.

• For a few seconds, the brain might look like a "Default Mode Network" (a state of daydreaming or self-reflection).
• Then, it switches to a completely different pattern.
• Why this matters: Traditional methods (the "photo") just showed the average of all these states, making the brain look blurry. The new method (the "video") showed that these distinct patterns exist clearly, just for short bursts.

2. Early Signs of "Adult" Thinking

The researchers found some very exciting patterns where the front of the brain (planning/thinking) and the back of the brain (memory/senses) lit up together.

• The Analogy: In adults, these two areas talk to each other to form the "Default Mode Network" (the network you use when you are daydreaming or thinking about yourself).
• The Finding: Even though newborn brains are supposed to be "immature" and messy, this study showed that these complex, adult-like conversations do happen in newborns. They just happen in short, fleeting bursts rather than being constant.

3. The "Global Glow"

They also found patterns where the entire brain lit up at once.

• The Analogy: This is like the whole room suddenly turning on all the lights.
• The Meaning: This might be the brain "waking up" or a general surge of activity, which is different from specific conversations between friends.

---

Why Is This a Big Deal?

1. It Works on Babies: They proved that this complex math works even with messy, moving baby data. It's like proving you can take a perfect group photo even if the kids are squirming, as long as you take enough snapshots.
2. It's More Honest: It shows us that brain development isn't a straight line. It's a series of dynamic, changing states.
3. Future Hope: If we can understand how these "flash mobs" work in healthy babies, we might be able to spot problems earlier in babies who are sick or born prematurely. We could see if their "flash mobs" are too quiet, too chaotic, or missing entirely.

The Bottom Line

This paper is a breakthrough because it moved from taking a blurry average photo of a baby's brain to watching a high-definition video of its activity. It revealed that even in the first days of life, the brain is already practicing complex, adult-like conversations, but it does so in short, dynamic bursts that we previously couldn't see.

Technical Summary

1. Problem Statement

• Limitations of Static Analysis: Traditional functional connectivity (FC) studies in neonates rely on "static" methods (averaging correlations over time), which fail to capture the non-stationary, dynamic nature of brain activity.
• Imaging Constraints: While fMRI is the gold standard, it is ill-suited for neonates due to motion sensitivity, the need for sedation (which alters brain activity), and the unnatural scanning environment.
• Methodological Gap: While Co-Activation Pattern (CAP) analysis has been successfully applied to adult fMRI and EEG to identify transient, high-activity states, it has not yet been validated or applied to neonatal High-Density Diffuse Optical Tomography (HD-DOT) data.
• Objective: To validate CAP analysis for neonatal HD-DOT and characterize the transient, dynamic organization of early functional brain networks in sleeping term newborns.

2. Methodology

Participants and Data Acquisition

• Cohort: 44 healthy, term-born newborns (median postmenstrual age: 40+3 weeks) recruited from Rosie Hospital, Cambridge.
• Modality: Wearable HD-DOT (LUMO system by Gowerlabs).
  • Setup: Custom caps with 12 tiles (36 sources, 48 detectors) covering frontal, central, and parietal regions.
  • Parameters: Dual wavelengths (735 nm, 850 nm), 10 Hz sampling rate, recorded cot-side while infants slept.
• Preprocessing Pipeline:
  • Motion Correction: Extraction of motion-free segments (≥100s) using Homer2.
  • Physiological Noise Removal: Short-separation regression to remove scalp hemodynamics; linear regression to remove high-frequency noise (respiration/cardiac) and low-frequency drifts (Legendre polynomials).
  • Reconstruction: Conversion of optical density to [HbO] and [HbR] concentrations; image reconstruction onto a neonatal head mesh using DOTHUB and Toast++.
  • Downsampling: Data reduced to 1 Hz to match hemodynamic timescales.

Analysis Pipeline (CAP Analysis)

1. Seed Selection:
   • Regions of Interest (ROIs): Frontal, Central, and Parietal (based on M-CRIB atlas).
   • Participant-Specific Seeds: Instead of group-averaged seeds, optimal seed locations were identified for each infant by maximizing the Dice similarity coefficient between the seed's correlation map and the ROI mask.
2. Frame Selection:
   • Data frames were sorted by [HbO] activity at the seed locations.
   • The top 15% of frames (highest seed activity) were retained for clustering.
3. Dimensionality Reduction:
   • Combined [HbO] and [HbR] data from ~11,000 sensitive nodes (21,968 dimensions).
   • PCA reduced dimensions to 275 components (retaining 99% variance).
4. Clustering:
   • K-Means clustering applied to the reduced data.
   • Optimal cluster count ($k=7$) determined via the elbow method and silhouette scores.
5. Metric Calculation:
   • Consistency: Intra-cluster spatial correlation (validates clustering quality).
   • Fractional Occupancy: Proportion of time spent in a specific CAP.
   • Dwell Time: Mean duration of consecutive frames within a CAP.
   • Transition Probability: Likelihood of moving from one CAP to another.

3. Key Results

Validation of CAP Analysis

• Feasibility: The top 15% of seed-selected frames showed very high correlation ($r \geq 0.95$) with conventional static seed-based correlation maps, validating the frame selection strategy.
• Consistency: CAPs achieved high consistency scores (Frontal: 0.51, Central: 0.48, Parietal: 0.48), exceeding benchmarks established in adult fMRI studies (Liu & Duyn, 2013). This confirms that neonatal HD-DOT data contains recurring, coherent spatial configurations.

Characterization of Patterns

• Distinct Configurations: Three general trends were observed across all regions:
  1. Seed-like patterns: Spatial maps resembling the static seed correlation.
  2. Anti-correlated patterns: Positive activation in the seed region paired with negative activation elsewhere.
  3. Global activation: Widespread positive [HbO] increases across the cortex.
• Frontoparietal Co-activation: Several CAPs (e.g., Frontal CAP 4, Parietal CAP 3, 4, 5, 7) exhibited transient frontoparietal co-activation.
  • Some patterns resembled the Default Mode Network (DMN) (medial frontal + lateral parietal).
  • Others suggested early Frontoparietal Network (FPN) configurations (lateral frontal + lateral parietal), sometimes with DMN suppression.

Temporal Dynamics

• Occupancy and Dwell Time: Significant differences were found in fractional occupancy and dwell times across different CAPs within each region.
• Transition Matrices: Transitions were predominantly "reflexive" (staying in the same CAP), consistent with the slow hemodynamic response. However, non-symmetric transitions indicated directed relationships between specific network states.

4. Key Contributions

1. First Application: This is the first study to apply CAP analysis to neonatal HD-DOT data, bridging a gap between adult dynamic connectivity methods and infant neuroimaging.
2. Methodological Validation: Demonstrated that HD-DOT data, despite motion artifacts and shorter recording times, is robust enough for unsupervised clustering and dynamic analysis.
3. Dynamic Insight: Revealed that neonatal resting-state networks are not continuous but exist as transient, intermittent states. Static maps are shown to be an average of these distinct, short-lived configurations.
4. Network Maturation: Provided evidence for the early, intermittent emergence of higher-order networks (DMN/FPN) at term age, which are often invisible to static correlation analyses.

5. Significance

• Clinical and Developmental Impact: By capturing the dynamics of network formation, this method offers a more sensitive tool for studying typical vs. atypical brain development. It allows researchers to observe how networks "come online" rather than just their average connectivity.
• Technological Advancement: Validates the use of wearable HD-DOT for complex dynamic analyses, making it a viable alternative to fMRI for studying infants in naturalistic settings (e.g., cot-side, home).
• Future Directions: Establishes a pipeline for longitudinal studies to track how these transient states evolve into mature adult networks over the first year of life, potentially serving as early biomarkers for neurodevelopmental disorders.

Limitations Noted

• Coverage: The current HD-DOT caps did not cover the occipital (visual) or temporal (auditory) regions.
• Data Loss: Motion artifacts still resulted in the exclusion of ~46% of the initial cohort (38/82 infants), though this is consistent with fNIRS infant studies.
• Standardization: The field lacks consensus on preprocessing steps (e.g., Global Signal Regression) and hyperparameters for neonatal CAP analysis.