Copyright Notice. Do not remove this notice. COMMONWEALTH OF AUSTRALIA Copyright Regulations 1969 WARNING

Transcription

1 Copyright Notice University of South Australia (UniSA) Do not remove this notice. COMMONWEALTH OF AUSTRALIA Copyright Regulations 969 WARNING This material has been produced and communicated to you by or on behalf of the University of South Australia pursuant to Part VB of the Copyright Act 968 (the Act). The material in this communication may be subject to copyright under the Act. Any further reproduction or communication of this material by you may be the subject of copyright protection under the Act. Do not remove this notice.

2 Computational Modelling of Cognitive Functions using EEG Data Student: Nabaraj Dahal Supervisors: Prof. Nanda Nandagopal Prof. Javaan Chahl Prof. Paul Gartner Dr. Zorica Nedic Dr. Mark McDonnell 2

3 Outline Background Objective and Approach Parametric modelling using TVAR with Kalman smoother Distraction Eperiment with Driving Simulator Event Related Synchronization / Desynchronization Results Conclusion and Future work 3

4 Background Cognition Computing massive number of processes inside the brain Stimuli Cognition Action Boundry Physiology behind the encoding, storage, retrieval and decision making? Applying mathematical models and Engineering tools- Reverse Engineering Represent cognitive functions from EEG responses to stimuli 4

5 Objective/Approach To investigate methodologies and techniques to model the human cognitive function using EEG responses to a set of cognitive stimuli To stimulate Cognition in Driving Scenario 5

6 Cognitive Load Lab Eperiments: Distraction Eperiment with Driving Simulator Response Yes Response No 5U 6

7 Data Acquisition EEG Equipments: NuAmps Amplifier Curry Acquisition System STIM2 stimulation provider system Triggers of stimuli and responses Recording Criteria: 3 EEG channels, 2 reference channels, trigger channel Hz (sampling rate) Hz -7 Hz with 5 Hz notch filter 7

8 Pre-processing Blinks detected using template matching and corrected using Independent Component Analysis. EEG chunks where response time falls between 5 and 3 ms only selected. 2 sec chunks(trials) after the stimulus (5 ms to 25 ms) were selected for further processing 256 Samples Data down sampled to 28 Hz. stimulus response 8

9 Parametric Modelling of EEG data AAR/ TVAR (Adaptive/Time-Varying Auto Regressive Model) has been etensively used Captures EEG pattern effectively Improves classification accuracy TVAR equation realized in state space model enables use of Kalman filter Kalman smoother uses future measurements available to further improve (smooth out) estimation performance 9

10 TVAR model for EEG EEG sequence: () are TVAR coefficients, p is the model order, 2 zero mean and variance v. TVAR coeffs follow Gauss Markov model: (2) [ a a... a ]' t ( t) ( t) ( t) 2 p w ~ N(, Q) t A is the transition matri. In State Space Model: v t is the array of TVAR coeffs is the i.i.d Gaussian noise. is Gaussian noise with h [ y y... y ]' t t t 2 t p is the vector of past p measurements.

11 TVAR Coefficients The initial state is assumed Gaussian: Model parameters will be: ~ N( ), The model parameters are estimated using Epectation maimization (EM) algorithm proposed by Khan& Dutta, 27. The EM algorithm based Kalman Smoother( Khan & Dutta, 27) gives the AR coeffs that captures the dynamics of EEG much better.

12 TVAR Coefficients with Kalman Smoother Forward recursion equations: Backward recursion equations: 2

13 TVAR Modelling using Kalman Order of filter chosen:7 Smoother 7 AR coefficients are predicted for each sample of each trials. 3

14 All Pole Modelling using TVAR Coefficients The AR coefficients of each samples are averaged over all the trials. 7 TVAR coefficients at each time instant from (8 to 256 samples). The transfer function of the AR model: (3) The poles of the model are given by the roots of the denominator. 4

15 All Pole plot for ONE Channel Pole calculation

16 All Pole plot for ONE Channel

17 All Pole plot for all channels 7

18 Event Related Synchronization/Desynchronization Increase or decrease in power in specific frequency band The cause is believed to be the increase of decrease of the synchrony of the underlying neuronal population. ERP-time locked EEG activity ERS/ERD- phase(frequency) locked activity 8

19 Synchronization/Desynchronization Synchronization Desynchronization

20 Driving Only Results FT7 FC3 FC4 FT T C C T8 2

21 Audio Distraction Driving Results FT7 FC3 FC4 FT T C C T8 2

22 Results with Audio Distraction 22

23 Conclusions and Future Work Conclusion: Pole analysis of a TVAR model is a sensitive and practical technique for visualizing ERS/ERD patterns. Future Work: Use the poles consistency and variations as feature Vector to build a model. Classify feature vector for distracted driving from base driving. Identify feature vector during visual distracted driving. Proceed to other cognitive functions like arousal, fatigue, memory, emotion in driving 23

24 24