Methods for Prediction and Early Detection of Neurological Events

• US 20130261490A1
• Filed: 12/05/2011
• Published: 10/03/2013
• Est. Priority Date: 12/05/2010
• Status: Active Grant

• Current Assignee: Brown University, US Department of Veterans Affairs (Government of the United States of America), The General Hospital Corporation (Mass General Brigham, Inc.)
• Original Assignee: Brown University, US Department of Veterans Affairs (Government of the United States of America), The General Hospital Corporation (Mass General Brigham, Inc.)
• Inventors: Donoghue, John P., Truccolo, Wilson, Hochberg, Leigh R., Cash, Sydney S.
• Granted Patent: US 10,448,877 B2
• Time in Patent Office: Days
• US Class Current: 600/544
• CPC Class Codes: A61B 5/316 Modalities, i.e. specific diagnostic methods
• CPC Class Codes: A61B 5/369 Electroencephalography [EEG] devices for psychotechnics A61B5/16
• CPC Class Codes: A61B 5/374 Detecting the frequency distribution of signals, e.g. detecting delta, theta, alpha, beta or gamma waves
• CPC Class Codes: A61B 5/375 using biofeedback
• CPC Class Codes: A61B 5/4094 Diagnosing or monitoring seizure diseases, e.g. epilepsy
• CPC Class Codes: A61B 5/7282 Event detection, e.g. detecting unique waveforms indicative of a medical condition A61B5/4094, A61B5/4818, A61B5/349, A61B5/0823 take precedence

Abstract

Several methods for prediction and detection of neurological events are proposed based on spatiotemporal patterns in recorded neural signals. The methods are illustrated with examples from neural data recorded from human and non-human primates.

Claims

1. A method for continuously predicting or detecting a future occurrence of a neurological event in a subject, the method comprising:

recording continuous signals generated from single neurons or other cells in the brain of the subject, detecting electrical signals generated from the single cells and measuring spiking activity of at least one recorded single neuron or other cell and characterizing the measured spiking activity of each recorded single neuron or other single cell as a collection of individual neural point process sample paths;

estimating a sample path probability distribution of a collection of sample paths of length T computed in the time interval (t−

T−

W,t−

T], for one or more neurons or cells, such that t is the current time, and W equals a time period extending into the past of specified duration such that W>

T>

0 wherein sample paths are overlapping or non-overlapping in time; and

determining, for each neuron or cell, whether a sample path measured in the time interval (t−

T,t] for the given neuron or cell falls outside a given confidence interval of the current sample path distribution, wherein the occurrence of the neurological event is predicted or detected by whether the measured sample path falls outside of the given confidence interval.

2. The method according to claim 1, further comprising after determining:

assigning weight to a neuron based on the importance or reliability of that neuron for the prediction or detection of the neurological event;

wherein the occurrence of the neurological event is predicted or detected if the weighted sum of the neurons, whose sample paths were computed on the interval (t−

T,t], falls outside a given confidence interval of their corresponding sample path probability distributions computed on the interval (t−

T−

W,t−

T], and is above a given threshold.

6. The method according to claim 1 further comprising predicting or detecting that the neurological event will occur if the relative predictive power calculated within the time interval (t−

T,t] falls outside of a confidence interval of a prior determined relative predictive power.

7. The method according to claim 6, wherein the probability is conditional on past spiking histories.

10. The method according to claim 6 wherein the relative predictive power is calculated according to the formula:

RPP_i=2(AUC_i−

AUC_i*),

wherein AUC_i* is the area under the receiver operating characteristic curve corresponding to a chance level predictor for a particular neuron or other cell and model, estimated via random permutation methods.

11. The method according to claim 10, further comprising:

fitting conditional intensity function models to each recorded neuron or other cell based on data collected during a first time interval, (t−

T,t], and fitting multiple conditional intensity function models to each recorded neuron or other cell based on data collected during overlapping or non-overlapping time windows from a second time interval (t−

T−

W,t−

T];

determining a relative predictive power for all of the recorded neurons or other cells during the first time interval under a cross-validation scheme;

determining a probability distribution of relative predictive powers from each of the recorded neurons or other cells based on a plurality of RPPs computed from the multiple models fitted to data from the second time interval under a cross-validation scheme;

wherein the neurological event is predicted or detected to occur or is detected if the RPPs of at least a fraction of neurons or other cells, computed based on data from the interval (t−

T,t], fall outside a confidence interval of their corresponding RPP probability distribution.

12. The method according to claim 10 the method further comprising:

fitting the conditional intensity function models to each recorded neuron or other cell based on data collected during a first time interval, (t−

T,t], and fitting the multiple conditional intensity function models to each recorded neuron or other cell based on data collected during overlapping or non-overlapping time windows from a second time interval (t−

T−

W,t−

T];

determining a relative predictive power for all of the recorded neurons or other cells during the first time interval;

determining a probability distribution of relative predictive powers from each of the recorded neurons or other cells based on the RPPs computed from the multiple models fitted to data from the second time interval;

assigning weight to a neuron based on the importance or reliability of that neuron for predicting a neurological event;

wherein the occurrence of the neurological event is predicted or detected if the weighed sum of neurons, whose RPPs computed from data in (t−

T,t] falls outside a confidence interval of their corresponding RPP probability distributions, is above a threshold.

13. The method according to claim 10 further comprising:

fitting conditional intensity function models to each recorded neuron or other cell based on data collected during a first time interval, (t−

T,t], and fitting multiple conditional intensity function models to each recorded neuron or other cell based on data collected during overlapping or non-overlapping time windows from a second time interval (t−

T−

W,t−

T];

determining a probability distribution of relative predictive powers computed under a cross-validation scheme from all of a plurality of the recorded neurons or other cells during the first time interval;

determining a probability distribution of relative predictive powers computed under a cross-validation scheme from all of the recorded neurons or other cells during the second time interval;

wherein the neurological event is predicted to occur or is detected if the two probability distributions are determined to be statistically different.

14. The method according to claim 10, further comprising:

fitting the point process models to each recorded neuron or other cell based on data collected during a current first time interval, (t−

T, t] and a second time interval (t−

T−

W,t−

T], where t is the current time and W is a specified amount of time greater than T; and

,

determining a probability distribution of relative predictive powers for all of a plurality of the recorded neurons or other cells during the first time interval and determining a probability distribution of relative predictive powers for all of the recorded neurons or other cells during the second time interval, wherein the occurrence of the neurological event is predicted or detected if the two probability distributions are determined to be statistically different.

15. The method according to claim 11, wherein the time interval (t−

T,t] is at least one window selected from a single window, overlapping windows, and non-overlapping windows.

3. A method for continuously predicting or detecting an occurrence of a neurological event in a subject'"'"'s body, the method comprising:

recording signals generated from single neurons or other cells in the brain of the subject and detecting electrical signals generated from the single neurons or other cells;

measuring spiking activity of at least one recorded single neuron or other cell, wherein the spiking activity of each neuron or other cell is represented as a spike train;

estimating a conditional intensity function model of the spike train for each neuron or other cell;

calculating a probability of a given neuron or other cell spiking at a given time using the estimated conditional intensity function model;

computing a receiver operating characteristic curve for each neuron or other cell from its corresponding spike train and calculated probability; and

deriving a relative predictive power measure from the receiver operating characteristic curve, wherein the occurrence of the neurological event is determined from the measured relative predictive power.

4. The method according to claim 3, further comprising after measuring spiking activity, fitting a history point process model to each recorded neuron or cell using the spike train.

5. The method according to claim 3, further comprising after computing the receiver operating characteristic curve, measuring the area under the curve.

8. The method according to claim 4 wherein the step of fitting a history point process model comprises using the following equation:

9. The method according to claim 3, wherein the step of calculating the probability comprises using the following equation:

Pr(x_i,t=1|H_t)=λ

_i(t|H_t)Δ

+o(Δ

)≈

λ

_i(t|H_t)Δ

wherein x_i,tε

{0,1} denotes the state (0=no spike, 1=spike) of the ith neuron or cell at time t, H_t is the spiking history up to and not including, time t of all recorded neurons or other cells, λ

_i(t|H_t) is the conditional intensity function model of the ith neuron or cell, conditioned on the recorded spiking histories, Δ

is the bin size for the discretization of time, and o(Δ

) relates to a probability of observing more than one spike in a specified time interval.

30. The method according to claim 8, further comprising choosing a temporal basis function.

31. The method according to claim 30, wherein the temporal basis function comprises a raised cosine function.

16. A method for continuously predicting or detecting an occurrence of a neurological event in a patient, the method comprising:

recording signals generated from a plurality of single neurons or other cells in the brain of the patient and detecting electrical signals generated from the single neurons or other cells;

measuring spiking activity of a corresponding plurality of recorded single neurons or other cells, wherein the spiking activity of each neuron or other cell is represented as a spike train;

estimating a conditional intensity function model of the spike train for each neuron or other cell, and deriving a graphical model from estimated conditional intensity functions for the set of recorded neurons or other cells during the time interval (t−

T, t], wherein t is the current time and T>

0 such that a parameter is determined from the graphical model; and

,

comparing the parameter to a probability distribution of the same parameter determined during multiple windows in the time interval (t−

T−

W,t−

T], wherein W>

T and occurrence of a neurological event is predicted or detected from the comparison.

17. The method according to claim 16, wherein the parameter comprises graph density.

18. The method according to claim 17, wherein the graph density comprises dividing number of directed edges by total possible number of edges.

27. The method according to claim 17 wherein the neurological event is an epileptic seizure.

28. The method according to claim 17 wherein the method further comprises providing a diagnosis based on the prediction or detection.

29. The method according to claim 17, wherein the method further comprises providing a prognosis based on the diagnosis.

19. A method for continuously predicting or detecting an occurrence of a neurological event in a subject, the method comprising:

recording signals generated from single neurons or other cells in the brain of the subject, detecting electrical signals generated from the single neurons or other cells, and measuring spiking activity of at least one recorded single neuron or other cell;

measuring an electric field potential of at least one recorded single neuron or other cell;

measuring a neural electric field potential and estimating a pairwise spike-field spectral coherence between the spike train and the field potential at a given frequency f, for each pair of recorded neuron (or other cell) and field potential, wherein the spectral coherence is determined according to

20. The method according to claim 19 wherein the electric field potential comprises multi-unit activity.

21. The method according to claim 19 wherein the second probability distribution is based on a single time window collected during the time interval (t−

T−

W,t−

T].

22. The method according to claim 21 wherein the comparison comprises utilizing a statistical test.

23. The method according to claim 22 wherein the occurrence of the neurological event is predicted or detected to occur if the first and the second distributions are determined to be statistically different.

24. The method according to claim 23 wherein the statistical test comprises a two-sample Kolmogorov-Smirnov test.

25. The method according to claim 19 wherein the second probability distribution is based on a plurality of non-overlapping windows collected during the time interval (t−

T−

W,t−

T].

26. The method according to claim 19 wherein the occurrence of a neurological event is predicted or detected if a currently determined Spike-Field spectral coherence falls outside of a confidence interval of the second probability distribution.