A Peak Synchronization Measure for Multiple Signals

Rahul Biswas, Koulik Khamaru, Kaushik Kumar MajumdarIEEE TSP 2014 · DOI

TL;DR

This paper introduces a novel "peak synchronization" measure to quantify how well events (peaks) in multiple time-series signals align. This method is particularly useful for analyzing signals like intracranial EEG where spikes represent critical events, and it shows improved performance over traditional amplitude correlation in detecting seizure activity.

What Problem Does This Paper Address?

Understanding the interdependence between multiple time-series signals is crucial in many scientific fields, from physics to neuroscience. Traditional methods for measuring synchronization, like amplitude correlation or phase synchronization, often treat entire signal segments uniformly. However, in many real-world signals, specific events or features (like sharp peaks or spikes) carry more significant information. The challenge is to develop a synchronization measure that specifically focuses on the simultaneous occurrence of these important events, even if they are slightly time-lagged, and can be generalized to more than two signals. This is especially relevant in neuroscience, where simultaneous spikes across brain regions are known indicators of neural events.

Methods

The core of this paper is the proposed "peak synchronization" measure. It operates by:

  • Peak Detection: First, significant peaks are identified in each individual signal. The paper uses a standard thresholding method (median + 2 standard deviations within a sliding window) but notes that the synchronization measure itself is independent of the specific peak detection algorithm.
  • Weighting Peaks: A novel aspect is the use of a probability density function (like a Gaussian) to assign weights to peaks. Peaks occurring at the exact same time point receive the highest weight, while peaks occurring with a slight time lag receive progressively lower weights. This weighting scheme captures the idea that near-simultaneous events are more indicative of synchronization than distant ones.
  • Calculating Synchronization: For two signals, the measure combines weighted contributions of peaks from both signals. For multiple signals (r > 2), the measure is generalized by averaging the pairwise synchronization values.
  • Parameter Independence: The authors demonstrate that the measure is largely insensitive to the specific parameter values chosen for the weighting function, making it robust.

Datasets

The authors applied their method to intracranial EEG (iEEG) signals from 21 epileptic patients, specifically focusing on 87 focal onset seizures. This data was obtained from the Freiburg Seizure Prediction Project. The dataset consists of one-hour recordings (preictal, ictal, and postictal) for each seizure, acquired using a 128-channel system at a 256 Hz sampling rate.

Key Insights and Ideas

  • Event-Centric Synchronization: The paper successfully shifts the focus of synchronization from continuous signal segments to discrete, information-rich events (peaks). This is a crucial conceptual advancement for analyzing signals where spikes are paramount.
  • Robustness to Time Lags: The weighting scheme allows the measure to capture synchronization even when peaks are not perfectly simultaneous, which is a significant advantage over methods that require exact alignment.
  • Generalizability: The proposed measure can be straightforwardly extended from pairwise synchronization to synchronization among multiple signals, making it applicable to complex systems with many interacting components.
  • Improved Seizure Detection: The peak synchronization measure demonstrably highlights seizure activity more effectively than traditional amplitude correlation, suggesting its potential for clinical applications in seizure detection and prediction.

Key Results

  • High Synchronization During Seizures: For 78 out of 87 analyzed seizures (approximately 89.66%), the peak synchronization measure showed a very high value during the seizure event, compared to pre- and post-seizure periods. This contrasts with amplitude correlation, where this trend was observed in only 24 out of 87 seizures (27.59%).
  • Statistical Significance: The authors validated their findings by comparing the peak synchronization values against shifted surrogate signals, confirming that the observed high synchronization during seizures is statistically significant and not due to chance.
  • Parameter Independence: Figures 2 and 3 illustrate that the average peak synchronization measure remains relatively stable across a wide range of central coefficients (weights), indicating the robustness of the method.
  • Computational Efficiency: The algorithm's time complexity is O(r²N) for r signals and N time points, which is efficient and suitable for online implementation.

Limitations and Future Work

  • Peak Definition: While the measure is independent of the peak detection algorithm, the quality of the synchronization measure relies on accurate peak identification. Artifacts or noise could lead to spurious peaks.
  • Focus on Upward Peaks: The current formulation primarily deals with upward peaks. The authors mention that the concept can be extended to downward peaks (troughs) or a combination of both.
  • Future Directions: The authors suggest applying this method to study interictal discharges in scalp EEG during sleep, where artifacts are reduced. They also propose exploring the replacement of peaks with Fourier coefficients for a measure closer to mutual information, and investigating its use in cognitive science research for event-related potentials (ERPs) and brain-computer interfaces.

Summary updated

BibTeX
@article{biswas2014peak,
  author = {Rahul Biswas and Koulik Khamaru and Kaushik Kumar Majumdar},
  title = {{A Peak Synchronization Measure for Multiple Signals}},
  journal = {IEEE Trans. Signal Process.},
  year = {2014},
  doi = {10.1109/tsp.2014.2333568},
}
Figure