Machine Learning-Based Classification of Patterns of EEG Synchronization for Seizure Prediction Piotr Mirowski, Deepak Madhavan MD, Yann LeCun PhD, Ruben Kuzniecky MD Courant Institute of Mathematical Sciences
Slide 2Observation window Seizure onset Extraction of elements from EEG, design acknowledgment + arrangement intracranial EEG interictal stage preictal stage ictal stage The seizure forecast issue Review of writing: most strategies actualize 1D choice limit machine learning utilized just for highlight determination Trade-off between: affectability (having the capacity to foresee seizures) specificity (keeping away from false positives) Benchmark information: 21-persistent Freiburg EEG dataset; current best results are: 42 % affectability 3 false positives for each day (0.25 fp/hour) [Litt and Echauz, 2002; Schulze-Bonhage et al, 2006]
Slide 3Hypotheses examples of brainwave synchronization: could separate preictal from interictal stages would be remarkable for every epileptic patient meaning of a " design " of brainwave synchronization: accumulation of bivariate " highlights " got from EEG, on all sets of EEG stations (central and extrafocal) taken at back to back time-focuses catch transient changes a bivariate " include ": catches a relationship : over a brief span window objective: quiet particular programmed figuring out how to separate preictal and interictal examples of brainwave synchronization highlights interictal preictal ictal [Le Van Quyen et al, 2003; Mirowski et al, 2009]
Slide 41min of preictal EEG 1min of interictal EEG 1min preictal design 1min interictal design Examples of examples of cross-connection Patterns of bivariate components Varying synchronization of EEG stations Non-frequential highlights: Max cross-relationship [Mormann et al, 2005] Nonlinear association [Arhnold et al, 1999] Dynamical entrainment [Iasemidis et al, 2005] Frequency-particular elements: [Le Van Quyen et al, 2005] Phase locking synchrony Entropy of stage contrast Wavelet intelligence [Le Van Quyen et al, 2003; Mirowski et al, 2009]
Slide 5c) 60-outline patterns (5min) d) Legend a) 1-outline patterns (5s) b) 12-outline patterns (1min) Separating examples of elements 2D projections (PCA) of wavelet synchrony SPLV highlights, tolerant 1 [Mirowski et al, 2009]
Slide 6Features registered on 5s windows ( N =1280 tests) 6x5/2= 15 bivariate elements on 6 EEG stations (Freiburg dataset) Wavelet investigation based synchrony values gathered in 7 electrophysiological recurrence groups : δ [0.5Hz-4Hz], θ [4Hz-7Hz], α [7Hz-13Hz], low β [13Hz-15Hz], high β [15Hz-30Hz], low γ [30Hz-45Hz], high γ [55Hz-120Hz] Features are collected into worldly examples y t : 12 outlines (1min) or 60 outlines (5min) # deed C, S, DSTL SPLV, H, Coh 1min 12 15=180 12 157=1260 5min 60 15=900 12 157=6300 Patterns of bivariate elements [Mirowski et al, 2009]
Slide 7Machine Learning Classifiers Input affectability Input example of elements: p x60 Layer 1 5@ p x48 Layer 2 5@ p x24 Layer 3 5@1x16 Layer 5 3 Layer 4 5@1x8 preictal interictal 1x8 convolution (crosswise over time) 1x2 sub-inspecting p x9 convolution (crosswise over time and space/freq) 1x13 convolution (crosswise over time) 1x2 subsampling L 1 - regularized convolutional systems (LeNet5, above) L 1 - regularized calculated relapse Support vector machines (Gaussian pieces) L1-regularization highlights sets of stations and recurrence groups discriminative for seizure expectation [LeCun et al, 1998; Mirowski et al, AAAI 2007, 2009]
Slide 821-understanding Freiburg EEG dataset therapeutically obstinate > 24h interictal 2 to 6 seizures Train + x-val on 66% information (57 prior seizures) PATIENT SPECIFIC Test on 33% information (31 later seizures) Previous best results: 42% affectability, 0.25 fpr/h [Aschenbrenner-Scheibe et al, 2003; Schelter et al, 2006a, 2006b; Maiwald, 2004; Winterhalder et al, 2003]
Slide 9Results on 21 patients (Freiburg) For every patient , no less than 1 technique predicts 100% of seizures , by and large a hour prior to the onset , with no false alert . Be that as it may, not generally a similar strategy… 16 mixes (highlight, classifier): how to pick a decent one? Classifiers : Features : Wavelet cognizance + conv-net : 15/21 patients ( 0 fp/hour ) Wavelet SPLV + conv-net : 13/21 patients (0 fp/hour) Wavelet intelligibility + SVM: 14/21 patients (<0.25 fp/hour) Nonlinear association + SVM: 13/21 patients (<0.25 fp/hour) [Mirowski et al, 2009]
Slide 10Example of seizure expectation True positives False negatives False negatives True negatives Wavelet lucidness + convolutional organize, tolerant 8 [Mirowski et al, 2009]
Slide 11Patient 12, nonlinear reliance 15 extrafocal TLB3 TLC2 TLB2 TLC2 [HR_7] TLC2 [TBB6] TLC2 [TBA4] TLC2 TLB2 TLB3 [HR_7] TLB3 [TBB6] TLB3 [TBA4] TLB3 [HR_7] TLB2 [TBB6] TLB2 [TBA4] TLB2 [TBB6] [HR_7] [TBA4] [HR_7] [TBA4] [TBB6] central extrafocal 10 extrafocal central extrafocal 5 intrafocal 0 30 40 50 60 0 10 20 Time (outlines) Patient 8, wavelet rationality 4 High γ (55-100Hz) Low γ (31-45Hz) 3 High β (14Hz – 30Hz) Low β (13Hz – 15Hz) 2 α (7Hz – 13Hz) 1 θ (4Hz – 7Hz) δ (< 4Hz) 0 20 30 40 50 60 0 10 Time (outlines) Feature affectability (and choice) L 1 regularization → meager weights Analysis of information affectability : Logistic relapse: take a gander at weights Conv nets: inclination on sources of info High γ frequencies could be discriminative for seizure forecast order ? [Mirowski et al, 2009]
Slide 12Thank You Litt B., Echauz J., Prediction of epileptic seizures , The Lancet Neurology 2002 EEG Database at the Epilepsy Center of the University Hospital of Freiburg, Germany , accessible: https://epilepsy.uni-freiburg.de/freiburg-seizure-expectation extend/eeg-database/Le Van Quyen M., Soss J., Navarro V., et al, Preictal state distinguishing proof by synchronization changes in long haul intracranial recordings , Clinical Neurophysiology 2005 Mormann F., Kreuz T., Rieke C., et al, On the consistency of epileptic seizures, Clinical Neurophysiology 2005 Mormann F., Elger C.E., Lehnertz K., Seizure foresight: from calculations to clinical practice , Current Opinion in Neurology 2006 Iasemidis L.D., Shiau D.S., Pardalos P.M., et al, Long-term planned online constant seizure forecast , Clinical Neurophysiology 2005 B. Schelter, M. Winterhalder, T. Maiwald, et al, Do False Predictions of Seizures Depend on the State of Vigilance? A Report from Two Seizure-Prediction Methods and Proposed Remedies , Epilepsia , 2006 B. Schelter, M. Winterhalder, T. Maiwald, et al, Testing factual hugeness of multivariate time arrangement examination methods for epileptic seizure forecast", Chaos , 2006 T. Maiwald, M. Winterhalder, R. Aschenbrenner-Scheibe, et al, Comparison of three nonlinear seizure expectation strategies by method for the seizure forecast trademark , Physica D , 2004 R. Aschenbrenner-Scheibe, T. Maiwald, M. Winterhalder, et al, How well can epileptic seizures be anticipated? An assessment of a nonlinear strategy , Brain , 2003 M. Winterhalder, T. Maiwald, H. U. Voss, et al, The seizure expectation trademark: a general system to survey and think about seizure forecast strategies , Epilepsy Behavior , 2003 J. Arnhold, P. Grassberger, K. Lehnertz, C. E. Elger, A strong strategy for identifying reliance: applications to intracranially recorded EEG , Physica D , 1999 LeCun Y., Bottou L., et al, Gradient-Based Learning Applied to Document Recognition , Proc IEEE , 86(11), 1998 Mirowski P., Madhavan D., et al, TDNN and ICA for EEG-Based Prediction of Epileptic Seizures Propagation , 22nd AAAI Conference 2007 Mirowski P., et al, Classification of Patterns of EEG Synchronization for Seizure Prediction , Clinical Neurophysiology, under update Mirowski P., et al, System and Method for Ictal Classification , US Patent Application, 2009 12
Slide 14Appendix
Slide 15Detailed results
Slide 16Maximum cross-connection Cross-relationship between's EEG channels x an and x b : Maximum cross-relationship for deferrals | τ |< 0.5s: Cross-relationship between's channels For every channel, decision of postpone giving best cross-relationship [Mormann et al, 2005] 16
Slide 17Time-delay inserting x a ( t ) and x b ( t ) are time-postpone embeddings of d EEG tests from channels x an and x b around time t . Elec b Elec a 1 second [Iasemidis et al, 2005], [Mormann et al, 2005]
Slide 18Nonlinear reliance Measure Euclidian separations , in state-space , between directions of x a ( t ) and x b ( t ) . Similitude of direction of x a ( t ) to the direction of x b ( t ) : K closest neighbors of x a ( t ) : Distance of neighbors of x a ( t ) to x a ( t ) : Symmetric measure of comparability of directions: K closest neighbors of x b ( t ) : Distance of neighbors of x b ( t ) to x a ( t ) : [Arnhold et al, 1999] [Mormann et al, 2005]
Slide 19Difference of Lyapunov examples Estimate of the biggest Lyapunov type of x a ( t ) , i.e. exponential rate of development of an irritation in x a ( t ): STL b STL a Short-term Lyapunov example (registered more than 10sec) diminishes (i.e. solidness of EEG direction increments ) before seizure 1 hour Measure of joining of confused conduct of EEG channels x an and x b : disentrainment entrainment [Iasemidis et al, 2005] 19
Slide 20Phase locking, synchrony Phase locking = stage synchrony (Wavelet or Hilbert changes) stage [Le Van Quyen et al, 2005], [Mormann et al, 2005] 20
Slide 21Phase locking insights φ a,f ( t ) and φ b,f ( t ) are periods of Morlett wavelet coefficients from EEG channels x an and x b , at recurrence f , time t Phase-locking esteem at recurrence f : Related measure: wavelet rationality Coh a,b ( f ) Shannon entropy of stage distinction at recurrence f utilizing M b
SPONSORS
SPONSORS
SPONSORS