Single-Trial Extraction of Pure Somatosensory Evoked Potential Based on Expectation Maximization Approach

• Published in: IEEE transactions on neural systems and rehabilitation engineering Vol. 24; no. 1; pp. 10 - 19
• Main Authors: Chen, Wei, Chang, Chunqi, Hu, Yong
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.01.2016; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Artificial neural networks; Bayes methods; Bayes Theorem; Bayesian model; Brain modeling; Computer Simulation; cortical source; Data Interpretation, Statistical; Electroencephalography; Electroencephalography - methods; Estimation; Evoked potentials; Evoked Potentials, Somatosensory - physiology; expectation Maximization (EM); Extraction; Feature extraction; Intraoperative Neurophysiological Monitoring - methods; Likelihood Functions; log-likelihood; Mathematical models; Maximization; maximum likelihood (ML); Models, Statistical; Monitoring; Pattern Recognition, Automated - methods; Reaction Time - physiology; regions of interest (ROI); Rehabilitation; Reproducibility of Results; Sample Size; Scalp; Sensitivity and Specificity; Signal Processing, Computer-Assisted; Signal-To-Noise Ratio; single-trial; somatosensory evoked potential (SEP) latency; Surgeries; Tracking; universal template
• ISSN: 1534-4320; 1558-0210; 1558-0210
• DOI: 10.1109/TNSRE.2015.2432835

It is of great importance for intraoperative monitoring to accurately extract somatosensory evoked potentials (SEPs) and track its changes fast. Currently, multi-trial averaging is widely adopted for SEP signal extraction. However, because of the loss of variations related to SEP features across different trials, the estimated SEPs in such a way are not suitable for the purpose of real-time monitoring of every single trial of SEP. In order to handle this issue, a number of single-trial SEP extraction approaches have been developed in the literature, such as ARX and SOBI, but most of them have their performance limited due to not sufficient utilization of multi-trial and multi-condition structures of the signals. In this paper, a novel Bayesian model of SEP signals is proposed to make systemic use of multi-trial and multi-condition priors and other structural information in the signal by integrating both a cortical source propagation model and a SEP basis components model, and an Expectation Maximization (EM) algorithm is developed for single-trial SEP estimation under this model. Numerical simulations demonstrate that the developed method can provide reasonably good single-trial estimations of SEP as long as signal-to-noise ratio (SNR) of the measurements is no worse than -25 dB. The effectiveness of the proposed method is further verified by its application to real SEP measurements of a number of different subjects during spinal surgeries. It is observed that using the proposed approach the main SEP features (i.e., latencies) can be reliably estimated at single-trial basis, and thus the variation of latencies in different trials can be traced, which provides a solid support for surgical intraoperative monitoring.