Epilepsy Prediction from EEG and ECOG Data Nathan

Epilepsy Prediction from EEG and ECOG Data Nathan Intrator Computer Science Tel-Aviv University cs. tau. ac. il/~nin Collaborators TAU Hospital: Talma Hendler, Itzhak Fried, Miri Noifeld, TAU: Eshel Ben Jacob, Ilana Podipsky, Andrey Zhdanov Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Outline � � � � � The Epilepsy Problem, Clinical Terms, and need for prediction Sensing, eeg, ecog, depth electrodes Animal models Wavelets boaz@eng. tau. ac. il Eshel Vagus nerve Heart/EEG, HRV, HS Complex Network Theory bocaletti Da Silva / Cerotti, Correlation, My contribution – level sets Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Seizure prediction by nonlinear time series analysis of brain electrical activity Ilana Podlipsky Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Epilepsy � Epilepsy › Synchronous firing of neurons which create high amplitude electrical discharges; this ‘storm’ inhibits other neural signals from getting through and disables function areas of the brain � Statistics › Everyone's brain has the ability to produce a seizure under the right conditions › 1 in 20 will have an epileptic seizure at some time in their life � Treatment › Once diagnosed with epilepsy, people are generally given anti-epileptic medication. With the appropriate treatment, up to 70% of people could be seizure free. � Characteristics / symptoms › Seizures (40 different types) › ‘Aura’, a sensory hallucination, often precludes a seizure � EEG › Recording of neural activity of targeted neurons / neural regions in brain › Outputs brainwaves with associated rhythms and frequencies Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Types of Epilepsy � Partial Seizures (Most Common) Video › Simple partialynchronous firing of neurons which create high amplitude electrical discharges; this ‘storm’ inhibits other neural Complex partial › Statistics › Everyone's brain has the ability to produce a seizure under the right conditions › 1 in 20 will have an epileptic seizure at some time in their life � Absence › Once diagnosed with epilepsy, people are generally given anti-epileptic medication. With the appropriate treatment, up to 70% of people could be seizure free. � Characteristics / symptoms › Seizures (40 different types) › ‘Aura’, a sensory hallucination, often precludes a seizure � EEG › Recording of neural activity of targeted neurons / neural regions in brain › Outputs brainwaves with associated rhythms and frequencies Epilepsy. com

Epilepsy � Epilepsy › Synchronous firing of neurons which create high amplitude electrical discharges; this ‘storm’ inhibits other neural signals from getting through and disables function areas of the brain � Statistics › Everyone's brain has the ability to produce a seizure under the right conditions › 1 in 20 will have an epileptic seizure at some time in their life � Treatment › Once diagnosed with epilepsy, people are generally given anti-epileptic medication. With the appropriate treatment, up to 70% of people could be seizure free. � Characteristics / symptoms › Seizures (40 different types) › ‘Aura’, a sensory hallucination, often precludes a seizure � EEG › Recording of neural activity of targeted neurons / neural regions in brain › Outputs brainwaves with associated rhythms and frequencies Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Examples of Seizure Morphologies Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Complex Network Theory Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Prof Paul Gompers Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Vagus nerve stimulation (VNS) A lead is attached to the left vagus nerve in the lower part of the neck. • It delivers mild electrical stimulations on demand Deep brain stimulation targets the thalamus (which relays pain, temperature, and touch sensations to the brain). Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Humans Rats Results of HRV Prediction � Successful forecasting � Tachycardia period � success rate � 86% � |∆RRI| Vs. RRI � forecasting times � 1. 5 -11 min. � Successful forecasting � Bradycardia period � success rate � 82% � |∆RRI| Vs. RRI � forecasting times � 2. 5 -9 min.

Fyodor Dostoyevsky (1821 -1881) Most known epileptic novelist � Gave vivid accounts of apparent temporal lobe seizures in his novel “The Idiot” � � Describes an aura he used to get before the onset of a seizure Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Vagus Nerve Stimulation Longest nerve in the body; � sweat, blood pressure, and heart activity (heart rate) � Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

The Vagus Nerve Longest nerve in the body; � Originates in the Brainstem � Goes all the way to the stomach, passing through essential organs (Vocal cords, heart, lungs, intestines) � Also controls sweat, blood pressure, and heart activity (e. g. , heart rate) � Yaari & Beck, 202; Lopes da Silva et al. , 2003;

The Vagus Nerve (cont) Modulates the SYMPATHETIC and PARASYMPATHETIC system � Goes all the way to the stomach, passing through essential organs (Vocal cords, heart, lungs, intestines) � Also controls sweat, blood pressure, and heart activity (heart rate) � Yaari & Beck, 202; Lopes da Silva et al. , 2003;

The problem ~30% of epileptics left untreated and victim of violent seizures � Injuries resulting from epilepsy is most often caused by convulsive seizures � If a ‘lead-time’ could be provided by a seizure detection system, physical injury would be greatly reduced and quality of life increased � Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Distinctive Features of Epilepsy � The epileptogenic process is characterized by abnormal synchronous burst discharges in neuronal cell assemblies recordable during and in between seizures (Matsumoto & Ajmone‐Marsan 1964 a, Matsumoto & Ajmone Marsan 1964 b; Babb et al. 1987). � The transition to a seizure is caused by an increasing spatial and temporal non-linear summation of the activity of discharging neurons (Calvin 1971; Calvin et al. 1973). � Due to the typically unpredictable occurrence of seizures it remains difficult to investigate the rules governing the initiation of seizure activity in humans. Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Brain as a Dynamic System � A dynamical system consists of › State › Dynamics � State – the information necessary at any time instant to describe the future evolution of a system � Dynamics – defines how the state evolves over time Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Attractors and Dimensions � Attractor › Set of states towards which the system evolves – Characterizes the long term behavior of the system � Dimension of a system › Describes the amount of information required to specify a point on the attractor - the long term behavior of a system � More complex behavior – more information is required to describe this behavior – higher dimension of the system Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Brain as a Dynamic System � The application of theory of non-linear dynamics offers information about the dynamics of the neuronal networks. � Several authors have shown that EEG/ECo. G signals exhibit chaotic behavior (Basar, 1990; Frank et al, 1990; Pijn et al, 1991). � The correlation dimension D 2 (Grassberger and Procaccia 1983), provides good information about EEG complexity and chaotic behavior. (Mayer-Kress and Layne (1987) ) Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Dynamics of Epileptic EEG � The spatio-temporal dynamics of the epileptogenic focus is characterized by temporary transitions from high-to low-dimensional system states (dimension reductions) (Lehnertz & Elger 1995, 1997). � These dimension reductions allow the lateralization and possibly localization of the epileptogenic focus (Lehnertz & Elger 1995, 1997). Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Seizure prediction by non-linear time series analysis of brain electrical activity Christian E. Elger, Klaus Lehnertz (1998) � Do prolonged and pronounced transitions from high - to low dimensional system states characterize a pre-seizure phase? � The identification of this phase would enable new diagnostic and therapeutic possibilities in the field of epileptology. Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Methods Electrocorticograms (ECo. G) and stereoelectroencephalograms (SEEG) of 16 patients � 68 EEG epochs were analyzed. › Fifty‐two data sets of state 1; mean duration: 19. 5 ± 6. 9 min; ran – 40 min; minimum distance to any seizure: 24 h. › 16 data sets of state 2; mean duration before the electrographic seizure onset: 15. 1 ± 5. 8 min; range: 10– 30 min; seizure onset was defined as earliest signs of ictal ECo. G/SEEG patterns). � Seizure prediction by non‐linear time series analysis of brain electrical activity Christian E. Elger, Klaus Lehnertz Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Methods � A moving window dimension analysis was applied: 1. Data sets were segmented into half-overlapping digitally lowpass filtered consecutive epochs of 30 s duration. 2. Calculation of the modified correlation integral - the mean probability that the states at two different times are close. 3. Estimate of the correlation dimension D 2 for each epoch. Seizure prediction by non‐linear time series analysis of brain electrical activity Christian E. Elger, Klaus Lehnertz Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Calculation of correlation dimension � Digital low-pass filtering (cut-off frequency 40 Hz) � Construction of m-dimensional vectors Xm(i) (i = 1, N; m = 1, . . . , 30) from the initial ECo. G samples v(i) (i = 1, N) of a given electrode using the method of delays (Takens, 1981): Seizure prediction by non‐linear time series analysis of brain electrical activity Christian E. Elger, Klaus Lehnertz Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Correlation Integral � For a stepwise decreasing radius r of a hypersphere centered at each vector Xm(i) for increasing m the correlation integral Cm(r) was calculated as (Grassberger and Procaccia, 1983): � Counts the number of pairs of points with distance less then r. � For small r: Cm(r) ≈ r. D � D 2 = slope of 2 (in a linear region) Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Calculation of Correlation Dimension � The correlation dimension D 2 is obtained by: D 2=slope of for decreasing r in a linear region � Alternatively: � If no linear region is found D 2 = 10 Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Results For each selected electrode of the ECo. G sets, a time profile of the estimated D 2, values was constructed. � The seizure (S) exhibits lowest dimension values. � Seizure prediction by non‐linear time series analysis of brain electrical activity Christian E. Elger, Klaus Lehnertz Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Results � For both states maximum dimension reductions were always found inside the epileptogenic focus regardless of spike activity. � During state 2, maximum dimension reductions were always observed in time windows immediately preceding seizures. � In state 1: › Dimension reductions with a mean of 1. 0; range 0. 5 -2. 5. › Mean duration of 5. 25 min; range 1. 00– 10. 75 min. � In state 2: › Dimension reduction mean 2. 0; range: 1. 0– 3. 5. › Mean duration 11. 50 min; range: 4. 25– 25. 00 min. � Highly significant differences between maximum state 1 and pre-seizure state dimension reductions (Dr: Z = – 3. 41, P = 0. 0006; Tr: Z = – 3. 52, P = 0. 0004). Seizure prediction by non‐linear time series analysis of brain electrical activity Christian E. Elger, Klaus Lehnertz Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Discussion � A reduced dimensionality of brain activity, as soon as it is of sufficient size and duration, precisely defines states which proceed to a seizure. � I was demonstrated that the features of the pre-seizure state differ clearly from the one found during seizure. � Pronounced dimension reductions of pre-seizure electrical brain activity are restricted to the area of the epileptogenic focus, they can reflect increasing degree of synchronicity of pathologically discharging neurons. Seizure prediction by non‐linear time series analysis of brain electrical activity Christian E. Elger, Klaus Lehnertz Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Discussion � Correlation Dimension measure as presented here is subjective. � Highly sensitive to noise. � Subject specific. Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Introduction cont. qthe brain-heart q. Vagus Nerve axis q. The existence of pre- ictal phase Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Introduction cont. This study q Forecasting seizures q Partial complex – humans q Generalized - rats q Novel method for HRV analysis q Ph. D. D. H. Kerem q Ph. D. A. B. Geva q Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Known Methods Spectral analysis of the time series of R-R intervals q non-linear dynamics q shortcoming q inability to account for non-stationary states and transients q Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Known Methods cont. q Time-varying power spectral density estimation q Attractors and correlation dimensions Karhunen-Love transform-based signal analysis method q Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Fuzzy clustering approach q comet or torpedo-shaped q unsupervised method advantage Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Chosen method q. EEG-contained information of HRV. q (GEVA and KEREM, 1998) qan unsupervised method designed to deal merging and overlapping states qability with to spot and classify Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Humans Rats Data resources q Humans q 21 patients records, archived records q The recording machinery q simultaneous EEG and video recording q ECG channel q visual inspection by an EEG expert q The actual database q Rats q Hyperbaric-oxygen q ECG and EEG filtering and recording q Rats effects q Time period analyzing q Control rats Vs. research rats Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Method cont. q. Choice of analysis parameters q |∆RRI| Vs. RRI q embedding dimension N q For this experiment – q. Both features q. N = 3 q number of clusters Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Method cont. Forecasting criteria q Appearance q Disappearance q Dominant q False negative - False positive q Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Humans Rats Results � Successful forecasting � Tachycardia period � success rate � 86% � |∆RRI| Vs. RRI � forecasting times � 1. 5 -11 min. � Successful forecasting � Bradycardia period � success rate � 82% � |∆RRI| Vs. RRI � forecasting times � 2. 5 -9 min. Yaari & Beck, 2002; Lopes da Silva et al. , 2003;

Humans failures �false negative �One case �false positive �Two cases � Longer records � prediction Rats Results cont. � prediction failures �false negative �none �false positive �Two cases � Ignoring changes shown in control rats Yaari & Beck, 2002; Lopes da Silva et al. , 2003;