Comparison of Blind Source Separation Methods in Fast Somatosensory-Evoked Potential Detection

Hongtao Liu, Chunqi Chang, Keith D. K. Luk, Yong HuJournal of Clinical Neurophysiology 2011 · DOI

TL;DR

This paper evaluates various Blind Source Separation (BSS) algorithms for extracting Somatosensory Evoked Potentials (SEPs) from noisy physiological data. The study finds that Second-Order Statistics (SOS)-based methods, particularly SOBI6, significantly outperform Higher-Order Statistics (HOS)-based methods in this specific application, offering a more robust and reliable approach for real-time SEP monitoring.

What Problem Does This Paper Address?

Somatosensory Evoked Potentials (SEPs) are crucial for monitoring neurological function during surgery, but the signals are often contaminated by significant electrical and biological noise, leading to a low signal-to-noise ratio (SNR). Traditional methods like ensemble averaging (EA) require a large number of trials to achieve a measurable signal, which is time-consuming and may miss critical events during surgery. Existing signal processing techniques like filtering can distort SEP signals due to overlapping frequencies with noise. Therefore, there's a need for faster, more reliable methods to extract SEP signals from noisy data, especially for real-time applications. This paper addresses the gap in understanding which BSS algorithms are most effective for this specific challenge.

Methods

The authors compare several BSS algorithms, categorizing them into Higher-Order Statistics (HOS)-based (fastICA, Extended Infomax, JADE) and Second-Order Statistics (SOS)-based (SOBI, AMUSE, ESPRIT).

  • Simulation Study: They simulated SEP signals and added different types of noise (EEG, power-line, and mixed) at various SNRs to evaluate algorithm performance.
  • Clinical Experiment: Real SEP data from 15 scoliosis patients undergoing surgery was used.
  • Performance Metrics:
    • Correlation Coefficient (r): Measured the similarity between the known SEP template and the extracted SEP signal. An absolute value of |r| > 0.60 was considered a good extraction.
  • Success Ratio: The proportion of cases where |r| > 0.60.
  • Reproducibility: Assessed by comparing correlation coefficients of averaged single trials.

The paper details the mathematical underpinnings of each BSS method, including their assumptions about source signals (e.g., independence, non-Gaussianity, uncorrelatedness) and the statistical properties they exploit. Preprocessing steps like centering and pre-whitening were applied before BSS.

Datasets

  • Simulated Data: SEP signals were simulated and combined with synthetic EEG and power-line noise at an SNR of -20 dB. Three noise conditions were tested: EEG only, power-line only, and a mix of both.
  • Real Clinical Data: SEP signals were recorded from 15 patients undergoing spinal surgery. The template SEP signals were obtained by averaging 300 trials, as per the American Electroencephalographic Society guidelines.

Key Insights and Ideas

  • SOS vs. HOS for SEP: SOS-based BSS algorithms generally outperformed HOS-based algorithms for SEP extraction. This is attributed to the limited data points available in typical SEP recordings. HOS methods require significantly more data for reliable estimation compared to SOS methods.
  • SOBI's Superiority: The Second-Order Blind Identification (SOBI) algorithm, particularly the SOBI6 variant (using six time-lagged covariance matrices), consistently demonstrated the best performance across various noise conditions in both simulation and experimental studies.
  • Data Shortage is Key: The limited number of data points in SEP signals makes HOS-based methods less suitable, as they struggle to reliably estimate higher-order statistics. SOS methods, relying on second-order statistics, are more robust in such scenarios.
  • Noise Type Matters: While SOBI performed well across all noise types, ESPRIT showed strong performance specifically when only power-line noise was present. JADE and Extended Infomax showed poor performance across most conditions.

Key Results

  • Simulation Study:
    • In EEG-only noise, SOBI2 and SOBI6 showed significantly higher correlation coefficients than other methods.
  • With power-line noise, SOBI6 (0.96±0.01) and SOBI2 (0.95±0.02) achieved very high correlation coefficients.
  • In mixed EEG and power-line noise, SOBI6 achieved a correlation coefficient of 0.77±0.03, significantly higher than other methods.
  • SOBI6 had a success ratio greater than 90% for mixed noise, while other methods were below 10%.
  • FastICA seldom converged, and JADE/Extended Infomax had correlation coefficients below 0.5.
  • Experimental Study:
    • SOBI6 (0.81±0.14) and SOBI2 (0.72±0.21) showed the highest correlation coefficients with the SEP template.
  • SOBI6 exhibited the least variation in correlation coefficients.
  • SOBI6 demonstrated the highest reproducibility (0.94±0.03) between averaged single trials.

Limitations and Future Work

  • Data Shortage for HOS: The primary limitation for HOS-based methods is the insufficient number of data points for reliable estimation of higher-order statistics.
  • Subjectivity in Component Selection: While efforts were made to reduce bias (e.g., double-blind identification), the initial selection of SEP-related components still involved some degree of subjective visual inspection by an experienced neurophysiologist.
  • Future Work: The authors suggest further studies investigating automatic SEP-related component recognition to completely eliminate subjective bias. They also highlight the need for research into real-time, online, and automatic SEP extraction, specifically focusing on SOBI. The choice of the number of correlation matrices for SOBI is presented as a trade-off between computational complexity and performance, suggesting further exploration in this area.

Summary updated

Abstract

Blind source separation (BSS) is a promising method for extracting somatosensory-evoked potential (SEP). Although various BSS algorithms are available for SEP extraction, few studies have addressed the performance differences between them. In this study, we compared the performance of a number of typical BSS algorithms on SEP extraction from both computer simulations and clinical experiment. The algorithms we compared included second-order blind identification, estimation of signal parameters via rotation invariance technique, algorithm for multiple unknown signals extraction, joint approximate diagonalization of eigenmatrices, extended infomax, and fast independent component analysis. The performances of these BSS algorithms were determined by the correlation coefficients between the true and the extracted SEP signals. There were significant differences in the performances of the various BSS algorithms in a simulation study. In summary, second-order blind identification using six covariance matrix denoting SOBI6 was recommended as the most appropriate BSS method for fast SEP extraction from noisy backgrounds.
BibTeX
@article{liu2011comparison,
  author = {Hongtao Liu and Chunqi Chang and Keith D. K. Luk and Yong Hu},
  title = {{Comparison of Blind Source Separation Methods in Fast Somatosensory-Evoked Potential Detection}},
  journal = {Journal of Clinical Neurophysiology},
  year = {2011},
  doi = {10.1097/wnp.0b013e31821213bd},
}
Figure