Real time network modulation for intractable epilepsy

Transcription

1 Real time network modulation for intractable epilepsy Behnaam Aazhang! Electrical and Computer Engineering Rice University! and! Center for Wireless Communications University of Oulu Finland

2 Real time network modulation for intractable epilepsy Behnaam Aazhang! Electrical and Computer Engineering Rice University! and! Center for Wireless Communications University of Oulu Finland

3 Real time network modulation for intractable epilepsy Behnaam Aazhang! Electrical and Computer Engineering Rice University! and! Center for Wireless Communications University of Oulu Finland

4 acknowledgement

5 acknowledgement Rakesh Malladi

6 acknowledgement Rakesh Malladi Nitin Tandon, MD at UTHSC Giridhar Kalamangalam, MD at UTHSC

7 a scientific curiosity

8 a scientific curiosity How does human brain work? Ancient Egypt and Greece Roman empire the seat of intelligence

9 a scientific curiosity How does human brain work? Ancient Egypt and Greece Roman empire the seat of intelligence 19th century 90s the decade of the brain 2013 the brain initiative

10 understanding quantum leap neuron doctrine

11 understanding quantum leap neuron doctrine electrical circuit electrical excitability

12 understanding quantum leap neuron doctrine electrical circuit electrical excitability enabler tools microscope electrodes

13 grand challenges relation neuronal circuit connectivity and behavior

14 grand challenges relation neuronal circuit connectivity and behavior transition of neuronal circuits disease state to healthy state learning

15 our research focus network modulation as a reparative therapy epilepsy, parkinson, alzheimers circuits connectivity behavior common theme tools influence on network

16 this talk network modulation as a reparative therapy epilepsy, parkinson, alzheimers circuits connectivity behavior common theme tools influence on network

17 neurological disorders human nervous system a gigantic network with nano scale structures 200 billion neurons and trillions of connections

18 neurological disorders human nervous system a gigantic network with nano scale structures 200 billion neurons and trillions of connections over 200 identified neurological disorders cost of neurological diseases $600 billion annually

19 epilepsy unprovoked and recurring seizures seizure no standard definition

20 epilepsy unprovoked and recurring seizures seizure no standard definition abnormally synchronized hyper-excited neuronal activities variations sub-clinical seizure burst full blown seizure single focal seizure multifocal seizure

21 epilepsy effects 3 million patients in the USA medication resection stimulation (modulation) neurons respond to electric signals!

22 epilepsy effects 3 million patients in the USA medication resection stimulation (modulation) neurons respond to electric signals!

23 recording-stimulation deep brain

24 recording-stimulation deep brain subdural

25 recording-stimulation deep brain subdural trans-cranial

26 recording-stimulation deep brain subdural trans-cranial tradeoff: invasive versus effective

27 our methodology deep brain subdural trans-cranial

28 today s application subdural recording identify epileptic zone

29 today s application e Parameters As outlined earlier in this promethodology is to use low-frequency stimute the hyper-synchronous subdural recording links in an epilepby recent trends reported in studies on huand rodents [110, 111] showing identify epileptic zonethat low freon of individual sites can reduce IEDs. A his proposal is to apply paired pulse at low resection! th the electrodes connecting the hyperactive ocusing on just the highly excited regions. We the relative timing between the pulses applied electrodes with the actual propagation delay or instance, let us say we chose electrodes d1 es for applying LFS and the direction of causal epileptic zone d1 to d2. Also let t denote the time delay for ce from d1 to reach d2. Then LFS is applied at ediately after detecting inter-ictal spike as ded LFS is applied at d2 only after a delay of t. hat LFS will break the link just as it becomes Figure 7: An illustration depicting seizure network outlined in gray among subdural grid electrodes. See [41] for a methodology on how the electrodes are depicted onto the cortical surface (image courtesy of ).

30 potential application e Parameters As outlined earlier in this promethodology is to use low-frequency stimute the hyper-synchronous subdural stimulationlinks in an epilepby recent trends reported in studies on huand rodents [110, 111] showing that low freon of individual sites can reduce IEDs. A his proposal is to apply paired pulse at lowth the electrodes connecting the hyperactive ocusing on just the highly excited regions. We the relative timing between the pulses applied electrodes with the actual propagation delay or instance, let us say we chose electrodes d1 es for applying LFS and the direction of causal epileptic zone d1 to d2. Also let t denote the time delay for ce from d1 to reach d2. Then LFS is applied at ediately after detecting inter-ictal spike as ded LFS is applied at d2 only after a delay of t. hat LFS will break the link just as it becomes Figure 7: An illustration depicting seizure network outlined in gray among subdural grid electrodes. See [41] for a methodology on how the electrodes are depicted onto the cortical surface (image courtesy of ).

31 challenges in stimulation prevent the network from going to hyperexcitable state identify seizure zone identify temporal markers low frequency stimulation seizure markers!

32 subdural recording and stimulation wish list real time closed loop

33 let s do this!

34 subdural recording and modulation electro-cortico-graphy (ECoG) subdural 154 channels (electrodes)

35 recording electro-cortico-graphy (ECoG) subdural temporpolar 154 channels (electrodes) TP4 recording TP3 TP2 TP Time (secs)

36 recording electro-cortico-graphy (ECoG) subdural 154 channels (electrodes) TP4 recording TP3 TP2 these are not spikes TP Time (secs)

37 recording electro-cortico-graphy (ECoG) subdural 154 channels (electrodes) TP4 recording TP3 TP2 these are not spikes local field potentials TP Time (secs)

38 recording electro-cortico-graphy (ECoG) subdural 154 channels (electrodes) TP4 recording stimulation? TP3 TP2 TP Time (secs)

39 stimulation (modulation) protocol depress the influence of one population of neurons on another temporally precise low frequency stimulation of selected electrodes CCEPs ECoG Spectral Decomposition Temporal Decomposition Sparse Effective Connectivity Stimulation Locations Low-frequency Stimulation Stimulation Timing Stimulation Protocol

40 research agenda temporally precise low frequency stimulation of selected electrodes CCEPs ECoG Spectral Decomposition Temporal Decomposition Sparse Effective Connectivity Stimulation Locations Low-frequency Stimulation Stimulation Timing Stimulation Protocol

41 research agenda temporally precise low frequency stimulation of selected electrodes develop a model and protocols real-time, closed loop, build the system clinical trial CCEPs ECoG Spectral Decomposition Temporal Decomposition Sparse Effective Connectivity Stimulation Locations Low-frequency Stimulation Stimulation Timing Stimulation Protocol

42 today s talk develop sparse effective connectivity CCEPs ECoG Spectral Decomposition Temporal Decomposition Sparse Effective Connectivity Stimulation Locations Low-frequency Stimulation Stimulation Timing Stimulation Protocol

43 modeling time series of length N d channels X N 1 =(x 1, x 2,, x N ) x n =[x n (1),x n (2),,x n (d)] T 2< d 8n

44 modeling time series of length N d channels these are not spikes local field potentials TP4 TP3 TP2 TP1 X N 1 =(x 1, x 2,, x N ) Time (secs) x n =[x n (1),x n (2),,x n (d)] T 2< d 8n

45 modeling time series of length N hours and hours of observations d channels 154 X N 1 =(x 1, x 2,, x N ) x n =[x n (1),x n (2),,x n (d)] T 2< d 8n

46 modeling time series of length N hours and hours of observations d channels 154 exploit natural sparsity!

47 modeling time series of length N hours and hours of observations d channels 154 exploit natural sparsity! spectral? temporal? spatial?

48 spectral decomposition time-frequency analysis Hz 4-8 Hz > 30 band band band band band Frequency (Hz) TP Time (secs) Time (secs)

49 temporal segmentation Bayesian change detection run length Data Samples - x (n) n Run-Length True Run-Length arg max r P (r (n) = r x 1:n ) n

50 temporal segmentation Bayesian change detection run length seizure markers inter-ictal spikes Data Samples - x (n) n high frequency oscillations Run-Length True Run-Length arg max r P (r (n) = r x 1:n ) n

51 {z } {z } spatial connectivity graphical model connectivity causality TP4 TP3 TP2 TP1 Frequency Time

52 Granger causality one time series forecasting another economics transportation

53 Granger causality one time series forecasting another economics transportation Norbert Wiener (1956) Clive Granger (1969) Hans Marko (1973) James Massey (1990)

54 a little background

55 a little background mutual information!! I(X; Y )= Z f XY log f XY f X f Y average information about X provided by Y X Y

56 a little background mutual information!! I(X; Y )= Z f XY log f XY f X f Y not directional no temporal causality X Y

57 a little background directed information and causality!! I(X N 1! Y N 1 )= NX n=1 I(X n 1 ; Y n Y n 1 1 ) directional X N 1 (X 1,X 2,...,X N ) Y N 1 (Y 1,Y 2,...,Y N )

58 a little background mutual information of time series!! I(X N 1 ; Y N 1 )= NX n=1 I(X N 1 ; Y n Y n 1 1 ) causality? X N 1 (X 1,X 2,...,X N ) Y N 1 (Y 1,Y 2,...,Y N )

59 a little background I(X N 1! Y N 1 )= NX n=1 I(X n 1 ; Y n Y n 1 1 ) mutual information of time series!! I(X N 1 ; Y N 1 )= NX n=1 I(X N 1 ; Y n Y n 1 1 ) causality? X N 1 (X 1,X 2,...,X N ) Y N 1 (Y 1,Y 2,...,Y N )

60 insight time series Does knowing time series X help with prediction of time series Y?

61 insight time series Does knowing time series X help with prediction of time series Y? Does knowing time series Y help with prediction of time series X? if directionality is not clear.

62 insight time series (specific underlying assumption!!) causality! X n = PX p=1 U p X n p + QX q=0 V q Y n q + z n Y n FIR filter + IIR filter X n noise z n

63 insight time series (specific underlying assumption!!) causality!! X n = PX p=1 U p X n p + QX q=0 V q Y n q + z n what is I(X N 1! Y N 1 )= NX n=1 I(X n 1 ; Y n Y n 1 1 )

64 2 examples

65 example 1!! X n = Y n 1 + z n with i.i.d. Y n Gaussian(0, 2 Y ) z n Gaussian(0, 2 z)

66 example 1 independent!! X n = Y n 1 + z n with i.i.d. Y n Gaussian(0, 2 Y ) z n Gaussian(0, 2 z)

67 example 1 Y n Gaussian(0, z n Gaussian(0, 2 Y ) 2 z)!! X n = Y n 1 + z n then I(Y! X) = 1 2 log(1 + 2 Y 2 z ) I(X! Y )=0

68 example 2!! X n = Y n + z n with i.i.d. Y n Gaussian(0, 2 Y ) z n Gaussian(0, 2 z)

69 example 2!! independent X n = Y n + z n with i.i.d. Y n Gaussian(0, 2 Y ) z n Gaussian(0, 2 z)

70 example 2 Y n Gaussian(0, z n Gaussian(0, 2 Y ) 2 z)!! X n = Y n + z n then I(Y! X) =I(X! Y )=I(X; Y )= 1 2 log(1 + 2 Y 2 z )

71 insight time series X n = PX QX U p X n p + V q Y n p=1 q=0 q + z n Gaussian moving average autoregressive I(X N 1! Y N 1 )= NX n=1 I(X n 1 ; Y n Y n 1 1 ) Granger causality = directed information

72 back to real data! subdural recording from epileptic patients left hippocampus region 151 channels sampling rate 1khz focus 6 channels 100 seconds

73 real data LAH6 LAH5 LAH4 LAH3 LAH2 LAH Time (s) 23

74 real data Gaussian? linear moving average autoregressive model? LAH6 LAH5 LAH4 LAH3 LAH2 LAH Time (s) 23

75 real data model order in the range P, Q 2 [75, 125] electrodes labeled LAH1-LAH X (6) n = PX p=1 U p X (6) n p + QX q=0 V q X (5) n q + z n

76 real data directed information {I(X (i)! X (j) )} 6 6 = B A

77 strong connections directed information

78 strong connections directed information? I(X (3)! X (6) )= ? 4 5 6

79 strong connections conditional directed information I(X (3)! X (6) X (5) )= ? 4 5 6

80 strong connections conditional directed information I(X (3)! X (6) X (5) )=

81 strong connections eliminating indirect influence

82 a broader view all the electrodes not sparse RAH2 RAH15 RAH14 RPH5 RAH1 RAH16 RAH4 RAH3 RAH5 RPH3 RPH1 RAH9 RAH7 RPH6 RPH2 RAH10 RAH8 RPH8 RPH7 RAH6 RPH4 RAH13 RAH12 RAH11 RPH9

83 a broader view eliminating indirect influences RAH2 RAH15 RAH14 RPH5 RAH1 RAH16 RAH4 RAH3 RAH5 RPH3 RPH1 RAH9 RAH7 RPH6 RPH2 RAH10 RAH8 RPH8 RPH7 RAH6 RPH4 RAH13 RAH12 RAH11 RPH9

84 a broader view not sparse temporal variations RAH2 RAH15 RAH14 RPH5 RAH1 RAH16 RAH4 RAH3 RAH5 RPH3 RPH1 RAH9 RAH7 RPH6 RPH2 RAH10 RAH8 RPH8 RPH7 RAH6 RPH4 RAH13 RAH12 RAH11 RPH9

85 a broader view eliminating indirect influences RAH2 RAH15 RAH14 RPH5 RAH1 RAH16 RAH4 RAH3 RAH5 RPH3 RPH1 RAH9 RAH7 RPH6 RPH2 RAH10 RAH8 RPH8 RPH7 RAH6 RPH4 RAH13 RAH12 RAH11 RPH9

86 a broader view not sparse spectral variations? RAH2 RAH15 RAH14 RPH5 RAH1 RAH16 RAH4 RAH3 RAH5 RPH3 RPH1 RAH9 RAH7 RPH6 RPH2 RAH10 RAH8 RPH8 RPH7 RAH6 RPH4 RAH13 RAH12 RAH11 RPH9

87 a broader view not sparse spectral variations RAH2 RAH15 RAH Hz band RPH5 RAH1 RAH16 RAH4 RAH3 RAH5 RPH3 RPH1 RAH9 RAH7 RPH6 RPH2 RAH10 RAH8 RPH8 RPH7 RAH6 RPH4 RAH13 RAH12 RAH11 RPH9

88 a broader view eliminating indirect influences RAH2 RAH15 RAH14 RPH5 RAH1 RAH16 RAH4 RAH3 RAH5 RPH3 RPH1 RAH9 RAH7 RPH6 RPH2 RAH10 RAH8 RPH8 RPH7 RAH6 RPH4 RAH13 RAH12 RAH11 RPH9

89 final thoughts closed loop stimulation markers spectral temporal spatial

90 final thoughts closed loop stimulation markers real time low frequency stimulation to depress the excitable state CCEPs ECoG Spectral Decomposition Temporal Decomposition Sparse Effective Connectivity Stimulation Locations Low-frequency Stimulation Stimulation Timing Stimulation Protocol

91 final thoughts develop protocols temporal markers directed connectivity build the system clinical trial!!!!!