Locally embedded presages of global network bursts

Satohiro Tajima, Takeshi Mita, Douglas J. Bakkum, Hirokazu Takahashi, Taro Toyoizumi

Abstract

Spontaneous, synchronous bursting of neural populations is a widely observed phenomenon in nervous networks, considered critical for both the functional and dysfunctional operations of the brain. However, the mechanism by which global synchrony emerges from an initially non-bursting state remains poorly understood. In this study, we develop a new state-space reconstruction method, Trace-Back Embedding (TBE), combined with high-resolution recordings of cultured neurons. This method extracts deterministic signatures of upcoming global bursts from the "local" dynamics of individual neurons during non-bursting periods. We find that local information within a single-cell time series can compare with or even outperform global mean-field activity for predicting future global bursts. Moreover, inter-cell variability in burst predictability reflects the underlying network structure. These findings demonstrate the deterministic mechanisms underlying locally concentrated early-warnings of global state transitions in self-organized networks.

Introduction

In both healthy brain function and pathological states like epilepsy, neural populations often exhibit sudden, massive synchrony known as "network bursts." Understanding how a network transitions from a quiet, non-bursting state to a global explosion of activity is a fundamental challenge in neuroscience. Traditionally, these transitions have been modeled as stochastic (random) processes. This research investigates whether there is a hidden, deterministic signature—a "presage"—buried in the activity of individual neurons before a burst occurs. Specifically, the study asks: can local information from a single cell predict a global state transition better than macroscopic information derived from the entire population?

Methodology: Trace-Back Embedding (TBE)

To address this, the researchers introduced Trace-Back Embedding (TBE), a model-free technique extending nonlinear state-space reconstruction (based on Takens’ Theorem and Convergent Cross-Mapping). The premise of TBE is that in a coupled dynamical system, the history of a single variable—in this case, one neuron's firing pattern—contains encoded information about the state of the entire system.

The TBE process involves three primary steps:

  1. Reconstructing the State-Space: Using "delay coordinates" to transform a one-dimensional time series of neural activity into a multi-dimensional trajectory.
  2. Retrograde Tracing: Analyzing the trajectory leading up to known bursts to identify "neighboring" states in the past that consistently precede global synchrony.
  3. Local-to-Global Mapping: Quantifying the accuracy with which the local dynamics of a single cell can "postdict" or predict the global mean-field activity of the whole population.

To validate this method, the authors utilized high-density CMOS multi-electrode arrays (MEAs) containing 11,011 electrodes to record from cultured rat cortical neurons. These preparations provided high spatiotemporal resolution, allowing for the tracking of 60–98 individual cells per preparation over several weeks (12–41 days in vitro).

The "Power of One": Key Insights

Local Superiority over Global Mean Fields

One of the most striking findings of the study is that approximately one-third of the recorded neurons could predict an upcoming global burst more accurately than the global mean-field activity itself. While it is often assumed that the average activity of a network is the most representative metric, the researchers found that the mean field can act as a "blurry" projection. In contrast, the temporal sequence of a single, well-positioned cell can capture the high-dimensional state of the system with greater precision.

Temporal Patterns vs. Firing Rates

Prediction accuracy was significantly higher when using TBE—which accounts for patterns over time—compared to simply monitoring momentary firing rates. This suggests that the "early-warning signal" is embedded in the specific sequence of firing rather than just the intensity of activity. In some cells, successful burst prediction rates exceeded 50% even when the network appeared "silent" to the naked eye, up to 500–1000ms before the burst occurred.

Predictability and Network Structure

Predictability was not uniform across the population. The "best" predictors were found to be "hub-like" neurons—cells with strong excitatory connections to the rest of the network. By using independent datasets where synaptic interactions were measured via electrical stimulation, the study established a direct causal link: a cell's ability to predict global events is a reflection of its functional role and hierarchy within the network.

Results and Robustness

The superiority of TBE-based detection using single-neuron dynamics over mean-field-based detection was statistically significant across all preparations (P < 5×10⁻¹⁴). Furthermore, burst predictability proved to be a stable property of individual neurons; cells that functioned as high-quality predictors in the first half of an experiment maintained that role in the second half (Spearman’s ρ = 0.79~0.93).

While the timing of the burst initiation appears to follow deterministic rules, the study noted that the size of the burst (the total number of spikes once the burst begins) remains largely stochastic and unpredictable. This suggests that while the trigger for a transition is deterministic, the eventual scale of the event is governed by different dynamics.

Discussion and Future Directions

This research demonstrates that global state transitions in neural networks are preceded by local, deterministic signatures. While the study was conducted in vitro, the findings have significant implications for understanding self-organizing systems.

Future work may focus on translating these findings to in vivo environments, where the complexity of a living brain may introduce additional noise. However, the discovery of these "presages" offers a potential pathway for developing more effective early-warning systems for clinical conditions like epilepsy. Beyond neuroscience, the "local-predicts-global" phenomenon may also be applicable to other complex systems, such as the propagation of rumors in social networks or the spread of disease in epidemiology.

Summary updated

Abstract

Spontaneous, synchronous bursting of neural population is a widely observed phenomenon in nervous networks, which is considered important for functions and dysfunctions of the brain. However, how the global synchrony across a large number of neurons emerges from an initially non-bursting network state is not fully understood. In this study, we develop a new state-space reconstruction method combined with high-resolution recordings of cultured neurons. This method extracts deterministic signatures of upcoming global bursts in "local" dynamics of individual neurons during non-bursting periods. We find that local information within a single-cell time series can compare with or even outperform the global mean field activity for predicting future global bursts. Moreover, the inter-cell variability in the burst predictability is found to reflect the network structure realized in the non-bursting periods. These findings demonstrate the deterministic mechanisms underlying the locally concentrated early-warnings of the global state transition in self-organized networks.
BibTeX
@inproceedings{tajima2017locally,
  author = {Satohiro Tajima and Takeshi Mita and Douglas J. Bakkum and Hirokazu Takahashi and Taro Toyoizumi},
  title = {{Locally embedded presages of global network bursts}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2017},
  doi = {10.1073/pnas.1705981114},
}
Figure