Report

French community for functional NIRS Signal processing techniques for fNIRS and application to Brain Computer Interfaces Gautier Durantin, ISAE/CERCO ISAE, Toulouse, 15.04.2014 1 Introduction EEG and fNIRS today encompass the most active areas of Brain Computer Interfaces research (Min 2010) fNIRS is currently mainly used as a complement of EEG (Takeuchi 2009, Fazli 2011) Noise reduction techniques and signal improvement techniques are the next step to improve BCI performance (Mitsukura 2013, Izzetoglu 2010) (Bashashati 2007) 2 Filtering for fNIRS – What do we need to filter ? Low frequency components : linear trends, measurement bias (Jang 2008) High frequency components : physiological noise (cardiac frequency), measurement noise (Huppert 2006) Raw signal (x) Filtered signal (y) FILTERING MODULE Use of filters to remove these components (e.g. linear filters) ! Delay, stability, performance 3 Moving Average Convergence Divergence Filter We propose a specific linear filter used by economists (Utsugi 2007, Cui 2010) The MACD (moving average convergence divergence), based on Exponential moving average (EMA) filters. Raw signal (x) Exponential moving average (EMA) = , −1 , … , −+1 Filtered signal (y) = ( ) MACD is obtained from two EMA : one short (k small), and one long (k big) Raw signal (x) Shortterm EMA Long-term EMA MACD + - Filtered signal (y) 4 Moving Average Convergence Divergence Filter MACD (short = 6s ; long = 13s) Economists use a signal line, obtained from a short EMA (5s) of MACD data, to predict stable increases on the curve. (Appel 1999) A MACD crossover of the signal line predicts a stable increase in the signal measured A MACD crossover of the signal line predicts a stable increase in the signal measured 5 Real-time hemodynamic response onset detection Controlled experiment of digit sequence memorization task 9 subjects 24 trials 8 9 4 51 x-x-x X STIMULATION REST (6-9sec) ANSWER (8 sec) Stimulus Increase in the signal MACD crossover 6 Towards a real-time BCI using fNIRS 4 control knobs : Speed, Heading, Altitude, Vertical speed Low load (speed 200, heading 200, Alt. 2000…) High load (speed 245, heading 315, altitude 8600…) 7 Towards a real-time BCI using fNIRS Pilot in ISAE flight simulator ATC msg mi Resp. windows SOA Signal filtering and synchronization (MACD) fNIRS output REST Real-time information on pilot’s mental state (Rest VS Task) Air Traffic Control (simulated) Overall accuracy : 98 % (std. Dev : 2,6%) or Classification Process TASK Load detection Data knowledge from phase L 8 Towards a real-time BCI using fNIRS Pilot in ISAE flight simulator ATC msg mi Resp. windows SOA Signal filtering and synchronization (MACD) fNIRS output REST Real-time information on perceived workload Air Traffic Control (simulated) or Classification Process TASK Load detection Data knowledge from phase L 9 Classification process fNIRS output of 20 training trials Signal processing (MACD) and feature extraction Use of different features (Tai & Chau 2009) [HbO2], [Hhb], peak response, kurtosis, skewness on different time windows Classifier design (LDA, SVM) TRAINING TESTING fNIRS output CLASSIFIER Real-time information on workload 10 Towards a real-time BCI using fNIRS : results 28 sessions 20 training trials 20 testing trials Overall accuracy obtained during testing phase : 79 % (std. dev : 12,8%) 19 subjects out of 28 have more than 75% accuracy 11 Further improvements in signal processing To improve signal processing and BCI accuracy, a solution would be to add a priori information in processing models Use of hemodynamic response models for temporal dynamic estimation of fNIRS. (Boynton 1996, Buxton 1997) Use of Kalman filtering to include estimation of temporal dynamics in signal processing (Abdelnour 2009, Gagnon 2011) 12 Kalman filtering Participant Stimulus NIRS Physiological processing model (HRF) Measurement model NIRS signal Dynamical model of hemodynamic response and fNIRS measurement fNIRS raw data KALMAN FILTER Confidence in the measures fNIRS filtered signal Confidence in the model Kalman filtering 13 Kalman filtering : results Tested offline on digit span memorization task data (9 subjects) with three levels of difficulty LINEAR FILTERING KALMAN FILTERING (effect size eta²=0,2) (effect size eta²=0,34) Kalman filtering is a promising tool to improve signal useability Challenges remain concerning Kalman tuning and real-time implementation 14 Conclusion Signal processing is a key step towards efficient Brain Computer Interface using fNIRS. Linear filtering brings good results, but improvements can be made to improve the accuracy of BCI designed with this type of filters. Kalman filtering or adaptive filtering are the best opportunities to improve signal useability. 15 Thank you for your attention 16 Digit span task 6 levels of difficulty 4 trials for each level of difficulty 17 Kalman modeling Physiological response model Measurement model Kalman filter 18 Possible improvement of Kalman modeling Physiological response model Measurement model Kalman filter MACD filter for onset prediction 19