Katholieke Universiteit Leuven K.U.Leuven PARAFAC: a powerful tool in EEG monitoring Sabine Van Huffel Dept. Electrical Engineering ESAT-SCD SCD Katholieke Universiteit Leuven, Belgium 1
Contents Overview Algorithms (see workshop presentations) Examples in EEG monitoring Conclusions and new directions 2
Algorithms: Canonical Decomposition (PARAFAC) Decomposition of an arbitrary tensor as a minimal linear combination of possibly non-orthogonal rank 1 tensors PARAFAC: compute via alternating least squares (Smilde, Bro, and Geladi, 2004) most popular Simultaneous matrix diagonalizations (De Lathauwer, 2006) Simultaneous generalized Schur decomposition (De Lathauwer, 2004) Other schemes (Paatero, 1999; Vorobyov, Sidiropoulos and Gerschman, 2005, ) With orthogonality constraints (Kolda, 2001; Zhang and Golub, 2001) Online PARAFAC (Nion and Sidiropoulos, 2009) 3
Contents Overview Algorithms Examples in EEG monitoring Epileptic seizure onset localization Neonatal seizure localization Event-Related Potential analysis Conclusions and new directions 4
Introduction: epileptic seizure onset localization Interictal EEG Ictal EEG 21 electrode positions @UZ Gasthuisberg Spikes, slow waves (epileptiform activity?) artifacts Ictal source localization eye blink muscle 5
PARAFAC for seizure onset localization C 1 B 1 C R B R Y = +. + + A 1 A R E time Split EEG in different frequencies using wavelets. frequency time space space => Analysis in 3 dimensions instead of just 2 6
Interpretation of a trilinear component PARAFAC: Example extracting 1 component B 1 : time course A 1 : distribution over channels C 1 : frequency content (distribution across scales). frequency time place = 7
PARAFAC for seizure onset localization * * * * B 1 B 2 C 1 C 2 8 A 1 A 2
Reconstructed epileptic atom (De Vos et al., NeuroImage 2007) (E. Acar et al, Bioinformatics 2007) 9
More interesting seizure A B Fp2 100µV F8 T4 T6 02 F4 C4 P4 Fz Cz Pz Fp1 F7 T3 T5 C O1 F3 C3 P3 T2 T1 0 1 2 3 4 5 6 7 8 9 10 Time (sec) 10
More interesting seizure A B Fp2 150µV F8 T4 T6 02 F4 C4 P4 Fz Cz Pz Fp1 C F7 T3 T5 O1 F3 C3 P3 T2 T1 0 1 2 3 4 5 6 7 8 9 10 Time (sec) 11
Why does this work? Muscle artifacts are distributed over frequencies by wavelet transformation and can not be modeled by a trilinear structure In 2 seconds, seizures are stable in time, frequency and space PARAFAC is ``unique 12
Added value in clinical practice? Validation study using the ictal EEG of 37 patients: Visual analysis : 21 well localized New method: 34 well localized PARAFAC is more reliable than visual analysis and matrix techniques (ICA, SVD) as preprocessing step for source localization Present study: using simultaneous EEG-fMRI Prospective validation of EEG-fMRI in the presurgical work-up for epilepsy surgery. PARAFAC used to delineate the ictal onset zone 13
Contents Overview Algorithms Examples in EEG monitoring Epileptic seizure onset localization Neonatal seizure localization Event-Related Potential Analysis Conclusions and new directions 14
Neonatal brain monitoring seizure detection Lack of oxygen supply leads to brain damage The occurrence of seizures best indicator for neurological damage and can increase damage if not properly treated Most seizures subclinical (90%): only detectable via EEG monitoring. Need for automated EEG monitoring. Collaboration with Sophia Child Hospital, Rotterdam, NL In the USA: 1 on 8 births premature 11.1% of prematures have seizures 15
Neonatal Seizure detection Algorithm, mimicking the human observer : (Deburchgraeve et al., Clinical Neurophysiology 2008 & 2009) 2 seizure types: Spike train Oscillation Combination 16 For each type a separate detection algorithm was developed.
Extract & localize oscillations using PARAFAC Seizure on C3-Cz-F3 + artefact on T5 O-CP: Extract spatial distribution of the seizure 17 without distortion of the artifact.
18 Extract & localize oscillations using PARAFAC
Why trilinear structure to extract oscillation? PARAFAC models as much variance as possible in the tensor that fits in a trilinear structure. Sensitive for activity that is active during the whole time segment, stable in localization and frequency Oscillations in the EEG meet those requirements, thus PARAFAC is most sensitive for oscillations in the EEG. Less suitable for spike train type seizures as they are discontinuous and to local in time. 19
Extract & localize spikes using PARAFAC We use the output of the seizure detector: 20
Extract & localize spikes using PARAFAC Construction of the tensor: 21
Extract & localize spikes using PARAFAC SP-CP: 22
Robustness of PARAFAC based spike localization Comparison with spike averaging - Construct tensor with 20 identical EEG spike segments - Add random noise to the tensor with different SNR - Calculate correlation with the noise free spatial distribution 23
Validation study Comparison with visual analysis of the EEG by a neurophysiologist. - In all cases there is a good qualitative correspondence between the neurologist and the algorithm. - Localization plots are helpful tool for neurophysiologist in analyzing seizures. - Together with seizure detector: useful for brain monitoring at the bedside. 24
Further remarks -PARAFAC expects a fixed localization in time. Divide migrating seizures into smaller windows to capture migration - Long seizures >1min: divide into smaller windows - Current research: use the extracted spatial distribution as input to dipole source localization with a realistic head model. 25
Contents Overview Algorithms Examples in EEG monitoring Epileptic seizure onset localization Neonatal seizure localization Event-Related Potential Analysis Conclusions and new directions 26
Event-Related Potential Analysis ERPs have very low SNR and suffer from artifacts caused by non-brain and brain sources Variety of PARAFAC Applications, e.g.: -Brain topography (Field and Graupe, Brain Topogr. 1991) -Brain-computer interfacing (A. Cichocki, IEEE computer society magazine, 2008) - Detection of rhythmic activity, e.g. (α, θ), during cognitive task (Miwakeichi et al., NeuroImage 2004) (Martinez-Montes, NeuroImage 2004) (Vanderperren et al., MBEC 2008) -Inter-trial phase coherence analysis in event-related EEG (MØrup et al., NeuroImage 2005)(M. Weiss et al., ICASSP 2009) 27 - Event-related EEG during simultaneous fmri acquisition
Reduction of Alpha activity in Go/NoGo task Overlooked type of brain related distortion = alpha activity Frequency between 8 and 13 Hz Mostly in a relaxed state, but also in alert people Apply PARAFAC on (Morlet) wavelet transformed EEG per epoch (after normalization), remove alpha component and reconstruct epoch 28
ERP results Go/NoGo task Original ERP ERP after PARAFAC 3 29
ERP during simultaneous fmri acquisition Application of PARAFAC to Event-Related EEG allows including three or more data dimensions into the tensor for ERP analysis (e.g. subjects, trials, tasks, etc.) unique data decomposition without additional assumptions finding ERP properties not revealed by traditional averaging Preprocessing important, e.g. removal of scanner-related artifacts Imposing orthogonality in 1 mode (trial, subject) to avoid degeneracy Optimize parameters (number of components) 30
Visual detection task EEG with 62 electrodes+eog+ecg in 3T 5 types of equiprobable stimuli 4 checkerboard segments, one per quadrant 1 full central checkerboard Variable randomized SOA UR DL DL UL DR 31
PARAFAC on Channels x Time x Subjects ERP for upper right stimulus Recorded in scanner plus fmri Recorded in scanner, no fmri Averaged ERPs per subject Averaged ERPs per subject 5 left occipital channels all channels different P1 latencies found 2 components: grand in grand average ERP average (left) and separated in 2 components BCGartefact (right) 32
Contents Overview Algorithms Examples in EEG monitoring Epileptic seizure onset localization Neonatal seizure localization Event-Related Potential Analysis Conclusions and new directions 33
Conclusions and future directions Tensors increasingly popular in biomedical SP: Successful: e.g. epileptic seizure onset localization using multichannel EEG restricted to use of PARAFAC solved with alternating least squares PARAFAC is more sensitive than visual EEG reading for seizure localization Electrode artifacts disturb localization Use results of PARAFAC as starting point for 3D EEG source localization 34 Remark: Tucker3 model has also been used for seizure localization (Acar et al., IASTED 2007)
Conclusions and future directions Improvements: More appropriate decompositions, e.g. block term (De Lathauwer, 2009), Adding constraints, e.g. spatial, source independence (De Vos et al., 2009), nonnegativity (MØrup et al, 2008) (Cichocki et al., 2008) Monitor brain activity dynamics online PARAFAC (Nion and Sidiropoulos, 2009) Allow ERPs to shift over time across the channels and dynamic sources PARAFAC2 (Weis et al, IEEE EMBS 2010) (Bro et al, Chemom.1997) (Harshman, UCLA report 1970) Multimodal data acquisition, e.g. combine with fmri (MØrup et al., 2005-2006; Martinez-Montes, NeuroImage, 2004) 35
Generalized tensor decomposition applied to simultaneously acquired EEG-fMRI data Martinez-Montes, NeuroImage 2004 Applications: Brain activation (fmri) in response to epileptic spikes (EEG) Brain activation (changes) in response to cognitive task 36
Acknowledgments I would like to thank my collaborators: Deburchgraeve Wouter, Maarten De Vos, Katrien Vanderperren, Lieven De Lathauwer, Mariya Ishteva Wim Van Paesschen Perumpillichira J. Cherian, Renate M. Swarte, Joleen H. Blok, Gerhard H. Visser, Paul Govaert 37