Abstracts

Rationale: Near infrared spectroscopy (NIRS) coupled with Electroencephalography (EEG) is a promising technique to study hemodynamic responses to epileptic activity [Gallagher et al Seizure 2008]. By measuring light absorption, NIRS monitors local fluctuations of deoxy (HbR) and oxy hemoglobin (HbO) concentrations with excellent temporal resolution (10Hz). EEG acquired during EEG/NIRS is visually inspected by an expert to detect epileptic discharges. NIRS response is then analyzed using a general linear model (GLM) combining a hemodynamic response function (HRF) after each EEG event. Because of NIRS higher sampling frequency than functional Magnetic Resonance Imaging (fMRI) and the explored volume being less localized, physiological noise is strongly present in the NIRS signal. The objective of this study is to evaluate the performances of two detection methods using realistic simulations Methods: Two linear regressions methods designed to model the physiological noise are evaluated: (1) a GLM using a set of sinusoidal regressors up to the frequency (fmax) [Cohen-Adad et al Med. Im. Analysis 2007], (2) a Weighted Least Square (WLS) regression in the time frequency domain [Matteau-Pelletier et al. IEEE BME 2008] modeling drifts using wavelets coefficients. Realistic physiological backgrounds consisted in resting state NIRS data acquired on the motor areas of two healthy subjects using a Techen CW5 system. 10 HbO and HbR baselines of 9 minutes were selected. Epileptic activity was generated by randomly simulating the occurrence of 50 interictal spikes, convolving each event with a standard HRF and adding it to one baseline scaled at a specified Signal-to-Noise Ratio (SNR). For each study, 10 epileptic signals were simulated for each baseline. Three parameters were evaluated: (a) the ability to detect activity in noisy signals varying the SNR from 1/1000 to 10, (b) the amount of slow drifts to be considered in the model varying fmax from 0 to 0.4 Hz, (c) the effect of discarding drift regressors correlating strongly with the epileptic response, varying this correlation threshold r from 0.02 to 0.37. Results: Analysis (a) suggests that the GLM generates more false positives than WLS exhibiting higher t values even in noisy conditions, suggesting WLS to be more reliable. Analysis (b) suggests that modeling slow drifts between 0 to 0.15 Hz is optimal for HbO for both methods. For HbR the optimal fmax is 0.025Hz, potentially due to the fact that HbR signals contain less physiological signal than HbO. Analysis (c) shows that the tuning of the correlation threshold r has an important impact on the performance (Figure 1). For r> 0.22, t values critically decrease, especially for the GLM, suggesting degeneracy in presence of correlated regressors. Conclusions: Modeling slow physiological drifts is absolutely needed to analyze accurately NIRS response to EEG epileptic activity, whereas this issue is less crucial in fMRI. Results show that WLS has an overall better behaviour than the GLM (better specificity and robustness to degeneracy), probably because of its better ability to model transient events.