Spatially Masked Regression Reveals Local and Distributed Predictability in Electrophysiological Recordings

This paper introduces a Spatially Masked Regression (SMR) framework that quantifies the balance between local and distributed information in electrophysiological recordings by reconstructing electrode signals while systematically excluding neighboring channels, revealing that individual channels reflect both immediate local redundancy and broader network-wide structure.

The Gist

The Big Question: Is the Signal Local or Global?

Imagine you are standing in a crowded room where everyone is talking. You have a microphone right in front of you.

• The Local View: You hear the person standing right next to you very clearly.
• The Global View: But you also hear the hum of the whole room, the music playing in the background, and the general chatter of the crowd.

In neuroscience, scientists record brain activity using electrodes (tiny microphones). A common debate is: Does the signal on one electrode come mostly from the tiny patch of brain right underneath it, or does it also carry information from the entire brain network?

Usually, it's hard to tell because the signals from nearby electrodes are so similar (like hearing the same conversation from two seats next to each other). This paper introduces a new tool to separate these two things.

The New Tool: "Spatially Masked Regression" (SMR)

The authors created a method called Spatially Masked Regression (SMR). Think of it as a "blindfolded prediction game."

1. The Setup: Imagine you want to guess what a specific person (the "Target") is saying.
2. The Normal Way: You listen to everyone in the room. Naturally, you rely heavily on the people sitting right next to the Target because their voices are loudest and clearest.
3. The SMR Way: The researchers put a "mask" over the people sitting next to the Target. You are not allowed to listen to the immediate neighbors. You have to guess what the Target is saying using only the people sitting far away across the room.

By gradually making the "mask" bigger (covering more neighbors), the researchers can see:

• How much of the Target's signal was just "local noise" from the neighbors?
• How much of the signal is actually part of a "global pattern" that can be predicted from far away?

What They Did (The Experiments)

They tested this on two different types of brain recordings, which are like two different types of rooms:

1. Scalp EEG (The "Big Room"): Electrodes are stuck on the outside of the head. Because the skull and skin mix the signals (like sound echoing in a large hall), the signals are very smooth and similar across the whole head.
2. Intracranial EEG (The "Small, Specific Room"): Electrodes are placed directly on the brain surface inside the skull. These signals are very sharp and specific to tiny areas, but the placement varies wildly from patient to patient (like a room where the furniture is rearranged differently every time).

The Results: What Did They Find?

1. The "Neighbors" Matter, But They Aren't Everything
When they blocked the immediate neighbors, the model could still guess the Target's signal reasonably well.

• Analogy: Even if you can't hear the people sitting next to you, you can still guess what the Target is saying by listening to the general vibe of the room.
• Finding: This proves that a single electrode doesn't just record its immediate neighborhood; it also carries a "broadcast" of information from the wider brain network.

2. The "Room Type" Matters (EEG vs. iEEG)

• Scalp EEG (The Big Room): The model was very good at predicting one person's signal from another person's data, even without seeing that specific person before.
  • Why? Because the signals are so mixed up and similar across the whole head, the "rules" of the room are the same for everyone.
• Intracranial EEG (The Specific Room): The model was less successful at transferring rules from one person to another.
  • Why? Because the electrodes are in different spots for different people, and the signals are very specific to tiny brain areas. It's like trying to guess the layout of a house just by looking at a blueprint from a different house; the walls might be in different places.

3. It's Not Just Random Noise
The researchers tried to trick the model by scrambling the data (shuffling the timing or randomizing the order of events). When they did this, the model failed.

• Analogy: If you take a song and play the notes in a random order, it's no longer a song.
• Finding: This confirms that the model isn't just guessing based on average volume or simple statistics. It is actually learning the structure and timing of how the brain communicates.

The Bottom Line

This paper shows that brain signals are a mix of local redundancy (neighbors saying the same thing) and distributed predictability (the whole network talking to each other).

The "Spatially Masked Regression" tool is a new way to measure exactly how much of a brain signal is "local" and how much is "global." It proves that even when you block out the immediate neighbors, the brain's wider network still leaves a clear fingerprint on every single electrode.

Technical Summary

Technical Summary: Spatially Masked Regression Reveals Local and Distributed Predictability in Electrophysiological Recordings

1. Problem Statement

Neural electrophysiological recordings (EEG and iEEG) are often interpreted as local measurements. However, due to volume conduction, spatial mixing, and distributed network dynamics, a single sensor's signal frequently reflects both immediate local activity and broader global structure. A fundamental challenge in systems neuroscience is distinguishing the extent to which an electrode's signal is determined by its immediate neighborhood versus distributed information across the entire array.

Standard functional connectivity measures (e.g., correlation, coherence, mutual information) often blur the distinction between local redundancy (where nearby sensors share similar signals due to biophysical smoothing) and distributed predictability (where information is encoded across the network). Furthermore, many connectivity approaches answer whether two signals are related, rather than quantifying how much of a specific signal is contained within the rest of the array and where that information originates. This distinction is particularly critical when sensor geometry and coverage vary significantly across individuals, as is common in clinical intracranial EEG (iEEG).

2. Methodology: Spatially Masked Regression (SMR)

The authors propose Spatially Masked Regression (SMR), a reconstruction-based framework designed to operationalize the trade-off between local and distributed information.

Core Framework

Instead of using pairwise connectivity as a primitive, SMR treats neural recordings as partially observed samples of a structured dynamical system. For a target electrode $i$, the model attempts to reconstruct its time series $\hat{x}_i(t)$ using a linear mapping from all other electrodes $j$ in the array, subject to a spatial constraint:
$$\hat{x}_i(t) = \sum_{j=1}^{N_s} (1 - K_{ij}) w_{ij} x_j(t)$$
Where:

• $N_s$ is the number of channels.
• $w_{ij}$ are learnable regression weights.
• $K_{ij}$ is a binary spatial mask. If electrode $j$ falls within a predefined neighborhood $N(i)$ of the target $i$, $K_{ij}=1$ (excluding it); otherwise, $K_{ij}=0$.

The "Experimental Knob"

The size of the excluded neighborhood (controlled by the number of nearest neighbors $k$) serves as an experimental control:

• Weak/No Masking: The model relies heavily on nearby channels, reflecting the local redundancy regime.
• Strict Masking: Nearby channels are withheld, forcing the model to rely on distant, non-local predictors.
  By progressively increasing the mask size, the framework quantifies how much predictive information survives when local neighbors are removed.

Datasets and Modalities

The framework was applied to two distinct modalities to test robustness across spatial mixing spectra:

1. Intracranial EEG (iEEG): The AJILE12 dataset (12 subjects, heterogeneous electrode placement, clinical epilepsy monitoring).
2. Scalp EEG: A motor execution/imagery dataset (15 subjects, standardized 61-channel montage over sensorimotor cortex).

Evaluation Metrics and Protocols

• Metric: Distance Correlation (DistCorr) was used to quantify the dependence between original and reconstructed signals. Unlike Pearson correlation, DistCorr captures both linear and nonlinear relationships.
• Intra-subject Evaluation: Models were trained and tested on held-out data from the same subject.
• Cross-subject Evaluation:
  • EEG: Direct transfer of the learned relationship matrix from a reference subject to a target subject (due to standardized montage).
  • iEEG: A Correlation-Based Electrode Mapping (CBEM) was used to align heterogeneous electrode sets via a rectangular assignment problem (Hungarian algorithm) before transferring the relationship matrix. No target-specific weights were fitted; only the best donor matrix was selected.
• Surrogate Controls: To ensure the model relies on structured temporal and cross-channel organization rather than marginal statistics, the authors tested performance on:
  • Phase-shuffled data (preserves spectrum, destroys phase/temporal dynamics).
  • IAAFT (preserves amplitude distribution and autocorrelation, destroys phase structure).
  • Block-shuffled data (preserves local trends, destroys long-range temporal order).

3. Key Results

Intra-subject Performance

• Scalp EEG: High reconstruction performance (Mean DistCorr $\approx 0.908$). This reflects the high spatial smoothness and redundancy of scalp recordings due to volume conduction.
• iEEG: Moderate reconstruction performance (Mean DistCorr $\approx 0.553$). Despite lower absolute performance, significant predictability remained even when local neighbors were excluded.
• Lagged Models: Introducing short temporal lags (up to 60ms) provided additional predictive power, suggesting that temporal offsets contribute to the distributed structure.

Cross-subject Generalization

• EEG: Robust transfer was observed (Mean DistCorr $\approx 0.783$). The standardized montage and global nature of scalp signals allow learned inter-electrode structures to generalize well across individuals.
• iEEG: Transfer was significantly lower (Mean DistCorr $\approx 0.389$) compared to EEG. The heterogeneity of electrode placement and the focal nature of iEEG signals make establishing exact functional correspondence across subjects difficult.

Masking and Surrogate Findings

• Local vs. Distributed: Masking experiments revealed that while nearby electrodes contribute strongly to reconstruction, they do not account for all predictive information. A substantial residual predictability persists even when immediate neighbors are excluded, indicating that individual channels reflect both local redundancy and broader distributed structure.
• Surrogate Validation: Performance dropped substantially when applied to phase-shuffled, IAAFT, and block-shuffled surrogates. This confirms that SMR success depends on the specific structured temporal and cross-channel organization of the neural data, not merely on marginal statistics (e.g., power spectra) or amplitude distributions.

4. Significance and Claims

The paper positions SMR as an interpretable framework for quantifying the balance between local and distributed information in neural recordings.

• Operationalizing Locality: The authors claim to provide a method to turn "spatial locality" into an experimental control, allowing researchers to measure how much of a signal is embedded in the immediate neighborhood versus the broader network.
• Beyond Connectivity: The approach shifts the question from "Are these signals related?" to "How much of this signal is contained in the rest of the array, and where does that information come from?"
• Modality Differences: The results highlight a fundamental difference in the spatial organization of information between scalp EEG (highly distributed, generalizable) and iEEG (more focal, subject-specific), driven by both biophysical properties (volume conduction) and experimental constraints (electrode placement).
• Diagnostic Utility: The framework serves as a diagnostic tool to assess the reconstructability of activity at one site from activity elsewhere, offering insights into the redundancy and distributed nature of neural representations without requiring a specific generative model of the underlying dynamics.

The authors do not claim to replace existing connectivity frameworks but rather to complement them by offering a reconstruction-based perspective on spatial redundancy and distributed predictability.