Can we predict the onset of seizures? Behnaam Aazhang J.S. - - PowerPoint PPT Presentation

Can we predict the onset of seizures?

Behnaam Aazhang J.S. Abercrombie Professor Electrical and Computer Engineering Rice University

Can we predict the onset of seizures?

• Let’s step back with a few more fundamental questions.

• understanding various disorders
• developing therapies
• patient-specific
• episode-specific
• scalability
• cost

How can engineers contribute to medicine?

engineers

• problem solving with constraints
• developing tools
• sense and measure
• nano-electronics
• control—modulation, stimulation, pacing
• machine learning and data analytics

example

• pacemakers

example

• pacemakers
• Can we modulate our neurological circuit?
• 86 billion neurons
• 10 micron diameter
• 100 Hz clock speed
• 100 trillion synapses

Rice neuroengineering initiative

Hardware Algorithms Get Data (Nanotechnology) Interpret and Use Data (Signal Processing)

Robinson

• St. Pierre

Veerarag- havan Kemere Xie Luan Szablowski Baraniuk Aazhang Pitkow Patel Allen O’ Malley Seymour

Raphael

Rice neuroengineering initiative

Hardware Algorithms Get Data (Nanotechnology) Interpret and Use Data (Signal Processing)

Robinson

• St. Pierre

Veerarag- havan Kemere Xie Luan Szablowski Baraniuk Aazhang Pitkow Patel Allen O’ Malley Seymour

Raphael

What am I excited about?

• Can we predict the onset of seizures?

What am I excited about?

• Can data analytics predict and prevent the onset of seizures in epileptic

patients?

epilepsy

• unprovoked and recurring seizures
• seizure
• no standard definition
• abnormally hyper-excited neuronal activities

epilepsy

• celebrities

epilepsy

• 1% of world’s population
• causes: stroke, tumors, infection, genetic, developmental,…
• 1/3 of patients do not respond to medication
• resection!!!!!
• deep brain stimulation?

the challenge ictal

the challenge inter-ictal

the challenge pre-ictal

approach

• patient and episode specific
• identify the seizure onset zone
• understand the dynamics of the underlying system
• predict seizures
• modulate (stimulate) to prevent the onset of seizure

epilepsy

• identify seizure onset zone

10 20 30 RAH1 RAMY2 RPBT1 Time (s)

Seizure Start Time

seizure zone

epilepsy

• identify seizure onset zone

10 20 30 RAH1 RAMY2 RPBT1 Time (s)

Seizure Start Time

seizure onset zone

epilepsy

• identify seizure onset zone

10 20 30 RAH1 RAMY2 RPBT1 Time (s)

Seizure Start Time

causality

causality

• ne time series forecasting another
• economics
• transportation
• …
• n. wiener (1956), c. granger (1969), h. marko (1973)
• j. massey (1990), g. kramer (1998),
• c. quinn, et. al. (2011)

a little background

• directed information and causality
• directional with temporal information

XN

1 ≡ (X1, X2, . . . , XN)

Y N

1

≡ (Y1, Y2, . . . , YN)

I(XN

1

→ Y N

1 ) = N

X

n=1

I(Xn

1 ; Yn|Y n−1 1

)

a little background

• mutual information of time series
• no temporal and no causal information

XN

1 ≡ (X1, X2, . . . , XN)

Y N

1

≡ (Y1, Y2, . . . , YN)

I(XN

1 ; Y N 1 ) = N

X

n=1

I(XN

1 ; Yn|Y n−1 1

)

I(XN

1

→ Y N

1 ) = N

X

n=1

I(Xn

1 ; Yn|Y n−1 1

)

a little background

• directed information of time series
• where

I(XN

1 → Y N 1 ) = H(Y N 1 ) − H(Y N 1 ||XN 1 )

I(XN

1 ; Y N 1 ) = H(Y N 1 ) − H(Y N 1 |XN 1 )

H(Y N

1 ||XN 1 ) = N

X

n=1

H(Yn|Y n−1

1

, Xn

1 )

a little background

• directed information of time series
• where

I(XN

1 → Y N 1 ) = H(Y N 1 ) − H(Y N 1 ||XN 1 )

causal conditional entropy

H(Y N

1 ||XN 1 ) = N

X

n=1

H(Yn|Y n−1

1

, Xn

1 )

I(XN

1 ; Y N 1 ) = H(Y N 1 ) − H(Y N 1 |XN 1 )

back to seizures

• causal relation among electrodes
• directed information
• model free—data driven
• k-nearest neighbor density estimation
• identify time series with largest directed information

10 20 30 RAH1 RAMY2 RPBT1 Time (s) Seizure Start Time

→ ˆ fX,Y → ˆ H(X), ˆ H(X, Y ) → ˆ I(X → Y )

seizure onset zone

• causal influence—directed connectivity
• a graph with electrodes as nodes and directed information as edge
• pre-ictal (period prior to seizure)

electrode

RAH1 RAH2 RPH4 RAMY2 RAINS3 RAMY4 RAMY3 RPH3 RAMY12 RAMY9 RAMY5 RAH3 LAH8 RPH2 RAMY7 LAH5 RAH5 RPBT11 RMOF5 RAMY10 RMOF10 RAMY11 RAINS4 RPBT1 RAMY6 RPH10 RAH13 RAH6 LAH12 RAH10

seizure onset zone

• causal influence—directed connectivity
• a graph with electrodes as nodes and directed information as edge
• pre-ictal (period prior to seizure)

flow between population of neurons

RAH1 RAH2 RPH4 RAMY2 RAINS3 RAMY4 RAMY3 RPH3 RAMY12 RAMY9 RAMY5 RAH3 LAH8 RPH2 RAMY7 LAH5 RAH5 RPBT11 RMOF5 RAMY10 RMOF10 RAMY11 RAINS4 RPBT1 RAMY6 RPH10 RAH13 RAH6 LAH12 RAH10

seizure onset zone

• causal influence—directed connectivity
• a graph with electrodes as nodes and directed information as edge
• pre-ictal, ictal, post-ictal

RAH1 RAH2 RPH4 RAMY2 RAINS3 RAMY4 RAMY3 RPH3 RAMY12 RAMY9 RAMY5 RAH3 LAH8 RPH2 RAMY7 LAH5 RAH5 RPBT11 RMOF5 RAMY10 RMOF10 RAMY11 RAINS4 RPBT1 RAMY6 RPH10 RAH13 RAH6 LAH12 RAH10 RAH1 RAH2 RPH4 RAMY2 RAINS3 RAMY4 RAMY3 RPH3 RAMY12 RAMY9 RAMY5 RAH3 LAH8 RPH2 RAMY7 LAH5 RAH5 RPBT11 RMOF5 RAMY10 RMOF10 RAMY11 RAINS4 RPBT1 RAMY6 RPH10 RAH13 RAH6 LAH12 RAH10

seizure onset zone

• causal influence—directed connectivity
• a graph with electrodes as nodes and directed information as edge
• pre-ictal (period prior to seizure)
• net degree of a node = out degree - in degree

RAH1 RAH2 RPH4 RAMY2 RAINS3 RAMY4 RAMY3 RPH3 RAMY12 RAMY9 RAMY5 RAH3 LAH8 RPH2 RAMY7 LAH5 RAH5 RPBT11 RMOF5 RAMY10 RMOF10 RAMY11 RAINS4 RPBT1 RAMY6 RPH10 RAH13 RAH6 LAH12 RAH10

seizure onset zone

L A H 5 R A H 1 R A H 2 R A M Y 2 R P H 4 R A I N S 3 R A M Y 4 R A M Y 3 R P H 3 R A M Y 1 2 2 4 6 8 10

Net Outlfow

• causal influence—directed connectivity
• a graph with electrodes as nodes and directed information as edge
• pre-ictal (period prior to seizure)
• net degree of a node = out degree - in degree

RAH1 RAH2 RPH4 RAMY2 RAINS3 RAMY4 RAMY3 RPH3 RAMY12 RAMY9 RAMY5 RAH3 LAH8 RPH2 RAMY7 LAH5 RAH5 RPBT11 RMOF5 RAMY10 RMOF10 RAMY11 RAINS4 RPBT1 RAMY6 RPH10 RAH13 RAH6 LAH12 RAH10

seizure onset zone

L A H 5 R A H 1 R A H 2 R A M Y 2 R P H 4 R A I N S 3 R A M Y 4 R A M Y 3 R P H 3 R A M Y 1 2 2 4 6 8 10

Net Outlfow

• causal influence—directed connectivity
• a graph with electrodes as nodes and directed information as edge
• pre-ictal (period prior to seizure)
• net degree of a node = out degree - in degree

RAH1 RAH2 RPH4 RAMY2 RAINS3 RAMY4 RAMY3 RPH3 RAMY12 RAMY9 RAMY5 RAH3 LAH8 RPH2 RAMY7 LAH5 RAH5 RPBT11 RMOF5 RAMY10 RMOF10 RAMY11 RAINS4 RPBT1 RAMY6 RPH10 RAH13 RAH6 LAH12 RAH10

seizure onset zone

L A H 5 R A H 1 R A H 2 R A M Y 2 R P H 4 R A I N S 3 R A M Y 4 R A M Y 3 R P H 3 R A M Y 1 2 2 4 6 8 10

Net Outlfow

electrodes in seizure

• nset zone
• causal influence—directed connectivity
• a graph with electrodes as nodes and directed information as edge
• pre-ictal (period prior to seizure)
• net degree of a node = out degree - in degree

RAH1 RAH2 RPH4 RAMY2 RAINS3 RAMY4 RAMY3 RPH3 RAMY12 RAMY9 RAMY5 RAH3 LAH8 RPH2 RAMY7 LAH5 RAH5 RPBT11 RMOF5 RAMY10 RMOF10 RAMY11 RAINS4 RPBT1 RAMY6 RPH10 RAH13 RAH6 LAH12 RAH10

• causal influence—directed connectivity
• a graph with electrodes as nodes and directed information as edge
• pre-ictal (period prior to seizure)
• net degree of a node = out degree - in degree

seizure onset zone

L A H 5 R A H 1 R A H 2 R A M Y 2 R P H 4 R A I N S 3 R A M Y 4 R A M Y 3 R P H 3 R A M Y 1 2 2 4 6 8 10

Net Outlfow

electrodes in seizure

• nset zone

nearly perfect match with the neurologist for all 12 patients

RAH1 RAH2 RPH4 RAMY2 RAINS3 RAMY4 RAMY3 RPH3 RAMY12 RAMY9 RAMY5 RAH3 LAH8 RPH2 RAMY7 LAH5 RAH5 RPBT11 RMOF5 RAMY10 RMOF10 RAMY11 RAINS4 RPBT1 RAMY6 RPH10 RAH13 RAH6 LAH12 RAH10

epilepsy

• focus on electrodes in the seizure onset zone—250 electrodes down to 6-10
• dynamics of time series to predict seizures

state space

• trajectory is nonlinear

state space

• trajectory is nonlinear
• inter-ictal and pre-ictal

state space

• trajectory is nonlinear
• inter-ictal and pre-ictal periods are not distinguishable

dynamics

• capturing dynamics of recordings
• K recordings in time m are
• a linear approximation is often insufficient to capture the dynamics

10 20 30 RAH1 RAMY2 RPBT1 Time (s)

Seizure Start Time

Xm+1 = f(Xm) Xm =       x(1)

m

x(2)

m

. . . x(K)

m

      Xm+1 = AXm where A is K × K

dynamics

• time embedding
• dynamics result in
• a linear approximation has shown to be sufficient in many applications

X1 =      X1 X2 . . . XM−h+1 X2 X3 . . . XM−h+2 . . . . . . ... . . . Xh Xh+1 . . . XM     

X2 =      X2 X3 . . . XM−h+2 X3 X4 . . . XM−h+3 . . . . . . ... . . . Xh+1 Xh+2 . . . XM+1      = f(X1)

X2 = AX1 where A is Kh × Kh

example

• Lorenz attractor

dynamic mode decomposition

• the main objective is to estimate
• dynamics of the system is captured by eigenvector and eigenvalues of
• the Kh x Kh matrix can be approximated by a smaller matrix

A A = X2X 1

1

= X2US1W> A A = ΦΛΦ−1

˜ A = W>

r AWr = W> r X2UrS1 r

extracting key feature

• spatiotemporal feature extraction

DMD Power DMD Phase

features

• DMD power versus frequencies and phase correlations among electrodes

44

Feature 2: Feature 1:

. vs.

back to seizure prediction

• dynamics

Xm+1 = AmXm

46

47

Seizure 7 Patient 020

48

seconds seconds

Seizure 4 Patient 038

49

seconds seconds

50

L2 between consecutive Hilbert Phase correlation windows L2 between consecutive PSD windows L2 between consecutive averaged DMD power windows L2 between consecutive averaged DMD phase windows

seconds seconds seconds seconds

extracting key feature

51

Prediction score Accuracy

(TN+TP)/all

Precision

TP/(TP+FP)

Sensitivity

TP/(TP+FN)

Specificity

FP/(FP+TN)

DMD 0.875 0.92 0.91 0.84 0.96 Fourier + Hilbert 0.82 0.83 0.71 0.83 0.83

SVM with kernel

seizure prediction

• promising results
• ECoG, DI, directed graphs, EmDMD, SVM
• patient specific
• real-time processing

control

• spatiotemporally focused modulation
• ultrasound
• electromagnetic signal

Target

ultrasound

• conjugate beam forming versus optimized beams

Off-target activation Otherwise Off-target activation Otherwise

electromagnetic waves

• conjugate beam forming versus optimized beams

take-home message

• Rice neuroengineering initiative
• sensing and imaging
• learning and data analytics
• extremely low-power small form-factor implementation
• control and modulation
• non-invasive or minimally invasive

the team

Funding:

projects

• non-invasive deep brain stimulation (fatima ahsan and boqiang fan)
• wireless multisite modulation of the diseased heart (romain

(romain cosentino and anton banta)

• real-time closed-loop modulation for depression (negar erfanian)
• learning and socialization in primates (sudha yellapantula)
• understanding olfactory circuit (joe young)
• modulation of epileptic circuit (dorsa moghaddam)