PROSTHETIC DEVICES AND METHODS AND SYSTEMS RELATED THERETO

• US 20080177196A1
• Filed: 01/19/2007
• Published: 07/24/2008
• Est. Priority Date: 01/19/2007
• Status: Active Grant

• Current Assignee: California Institute of Technology
• Original Assignee: California Institute of Technology
• Inventors: Wolf, Michael T., Cham, Jorge G., Branchaud, Edward A., Nenadic, Zoran, Burdick, Joel W., Andersen, Richard A.
• Granted Patent: US 8,095,210 B2
• Time in Patent Office: Days
• US Class Current: 600/544
• CPC Class Codes: A61B 5/061 Determining position of a probe within the body employing means separate from the probe, e.g. sensing internal probe position employing impedance electrodes on the surface of the body A61B1/00158 takes precedence
• CPC Class Codes: A61B 5/24 Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
• CPC Class Codes: A61B 5/291 for electroencephalography [EEG]
• CPC Class Codes: A61B 5/4041 Evaluating nerves condition
• CPC Class Codes: A61B 5/4893 Nerves
• CPC Class Codes: A61B 5/7203 for noise prevention, reduction or removal
• CPC Class Codes: A61B 5/726 using Wavelet transforms
• CPC Class Codes: A61B 5/7264 Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems neural networks per se G06N3/00; expert systems per se G06N5/00
• CPC Class Codes: B33Y 80/00 Products made by additive manufacturing
• CPC Class Codes: G06F 2218/12 Classification; Matching
• CPC Class Codes: G16H 50/20 for computer-aided diagnosis, e.g. based on medical expert systems

Abstract

The present invention relates to systems and methods for controlling neural prosthetic devices and electrophysiological recording equipment, and for using the same in clinical operation. Various embodiments of the invention are directed to an algorithm for autonomously isolating and maintaining neural action potential recordings. The algorithm may be used in connection with a neural interface microdrive capable of positioning electrodes to record signals from active neurons.

Claims

1. A method of automatically positioning an electrode disposed inside a brain, the method comprising:

moving an electrode along a predetermined path of travel until a neural signal having a current signal quality metric that is greater than or equal to a minimum threshold value and comprising action potentials is detected;

after the neural signal is detected, moving the electrode to a plurality of locations along the predetermined path and after each move, recording a sample of the neural signal;

detecting action potentials in the neural signal of each of the recorded samples;

classifying the action potentials of each of the recorded samples into classes;

calculating a first signal quality metric for each of the classes;

for each of the recorded samples, identifying a dominant class;

for each of the action potentials of the dominant class of each of the recorded samples, determining a second signal quality metric;

modeling an isolation curve as a function of the second signal quality metric for each of the action potentials of the dominant class of a first portion of the recorded samples and the locations of the plurality of locations where each of the recorded samples of the first portion of the recorded samples were recorded;

determining a move distance from a current electrode position to a location along the modeled isolation curve that the modeled isolation curve indicates will improve signal quality; and

as long as the move distance is greater than a predetermined minimum move distance threshold,

moving the electrode the move distance to a new electrode position along the predetermined path and including the new electrode position in the plurality of locations,

recording a new sample of the neural signal and including the new sample in the recorded samples,

detecting action potentials in the new sample of the neural signal,

classifying the action potentials of the new sample into classes,

calculating the first signal quality metric for each of the classes,

identifying the dominant class,

for each of the action potentials of the dominant class, determining the second signal quality metric,

modeling the isolation curve as a function of the second signal quality metric of each of the action potentials in the dominant class of a second portion of the recorded samples and the locations of the plurality of locations where each of the recorded samples of the second portion of the recorded samples were recorded, and

determining the move distance from the new electrode position to a location along the modeled isolation curve that the modeled isolation curve indicates will improve signal quality.

2. The method of claim 1, wherein classifying the action potentials of each of the recorded samples into classes comprises

representing each of the action potentials by a predetermined number of features;

classifying each of the action potentials using the predetermined number of features into one of a predetermined number of classes or an outlier class; and

determining the predetermined number of classes by calculating the Bayesian Information Criterion value for a plurality of candidate numbers of classes and setting the predetermined number of classes equal to the candidate number of classes that produced the largest Bayesian Information Criterion value.

3. The method of claim 1, wherein classifying the action potentials of each of the recorded samples into classes comprises

representing each of the action potentials by a predetermined number of principal components;

and

classifying each of the action potentials using the predetermined number of principal components into one of a predetermined number of classes or an outlier class.

4. The method of claim 1, wherein classifying the action potentials of each of the recorded samples into classes comprises

representing each of the action potentials by a predetermined number of Haar wavelet coefficients;

and

classifying each of the action potentials using the predetermined number of Haar wavelet coefficients into one of a predetermined number of classes or an outlier class.

5. The method of claim 1, wherein classifying the action potentials of each of the recorded samples into classes comprises for each action potential, extracting a segment of the sample corresponding to the action potential from the sample, and for each sample, aligning the extracted segments.

6. The method of the previous claim, wherein aligning the extracted segments comprises selecting a first segment, selecting a second segment having a first position, determining a first correlation between the first segment and the second segment in the first position, shifting the second segment from the first position to a second position, determining a second correlation between the first segment and the second segment in the second position, determining which of the first and second correlations is greatest, and placing the second segment in the position corresponding to the greatest correlation.

7. The method of claim 1, wherein classifying the action potentials of each of the recorded samples into classes comprises

representing each of the action potentials by a predetermined number of features;

classifying each of the action potentials using the predetermined number of features into one of a predetermined number of classes or an outlier class by constructing a finite mixture model to model the probability density function of action potentials in a feature space defined by the predetermined number of features used to represent each of the action potentials, wherein the finite mixture model comprises a plurality of component probability density functions, and one of the component probability density functions of the plurality corresponds to the outlier class and the other component probability density functions correspond to the predetermined number of classes; and

assigning each of the action potentials to one of the component probability density functions based upon the probability the action potential as represented by the predetermined number of features belongs to each of the component probability density functions.

8. The method of claim 7, wherein the component probability density function corresponding to the outlier class is a uniform distribution and the component probability density functions corresponding to the predetermined number of classes are Gaussian distributions.

9. The method of claim 7, wherein the component probability density functions corresponding to the predetermined number of classes each have a set of parameter values and the sets of parameter values are determined by maximizing the likelihood the action potentials as represented by the predetermined number of features belong to the finite mixture model constructed using the sets of parameter values.

10. The method of claim 1, wherein classifying the action potentials of each of the recorded samples into classes comprises

representing each of the action potentials by a predetermined number of features;

classifying each of the action potentials using the predetermined number of features into one of a predetermined number of classes or an outlier class; and

determining the predetermined number of classes comprising;

determining a plurality of candidate predetermined number of classes;

for each of the candidate predetermined number of classes, constructing a finite mixture model to model the probability density function of action potentials in a feature space defined by the predetermined number of features used to represent each of the action potentials, wherein the finite mixture model comprises a plurality of component probability density functions, and one of the component probability density functions of the plurality corresponds to the outlier class and the other component probability density functions correspond to the candidate predetermined number of classes;

for each of the finite mixture models constructed, determining a likelihood the action potentials as represented by the predetermined number of features belong to the finite mixture model constructed; and

setting the predetermined number of classes equal to the candidate predetermined number of classes used to construct the finite mixture model corresponding to the greatest likelihood.

11. The method of claim 1, wherein the modeled isolation curve comprises a sum of a predetermined number of basis functions having the form u⁽ⁱ⁻

1) wherein u is the position of the electrode when the sample was recorded and i takes the value of a natural number in a set from one to the predetermined number of basis functions.

12. The method of claim 1, wherein the modeled isolation curve comprises a sum of a predetermined number of basis functions, each of the basis functions having a form selected by an operator.

13. The method of claim 1, wherein the modeled isolation curve comprises a sum of a predetermined number of basis functions and the predetermined number of basis functions is determined using a Bayesian Model Section.

14. The method of claim 1, wherein the modeled isolation curve comprises a sum of a predetermined number of basis functions, each of the basis functions comprises an expansion coefficient, and modeling the isolation curve of the neuron comprises determining the expansion coefficient of each of the basis functions, and determining the expansion coefficient of each of the basis functions comprises using a least squares optimization to determine a value for each of the expansion coefficients.

15. The method of claim 1, further comprising, comparing the move distance to a maximum move threshold, and if the move distance is greater than the maximum move threshold, setting the move distance equal to predetermined maximum move distance.

16. The method of claim 1, wherein the action potentials of each of the recorded samples are classified into classes using unsupervised clustering.

17. The method of claim 1, wherein the modeled isolation curve is modeled using unsupervised regression analysis.

18. A method of monitoring a neural signal using an electrode disposed inside a brain comprising:

moving the electrode inside the brain to a plurality of locations and at each location, collecting a sample of the neural signal to obtain a plurality of samples of the neural signal;

collecting a sample of the neural signal to obtain a new sample of the neural signal;

detecting action potentials in the new neural signal, each action potential having an amplitude;

using unsupervised clustering to classify the action potentials into clusters;

identifying a dominant cluster;

calculating a current signal quality metric as a function of the amplitude of the action potentials of the dominant cluster; and

as long as the current signal quality metric is less than a minimum signal quality threshold,

modifying the plurality of samples to include the new sample, using a portion of the plurality of samples to create a model of an isolation curve of the neural signal,

determining a move distance to a new location that the model indicates will improve signal quality,

moving the electrode the move distance,

after moving the electrode, re-sampling the neural signal to replace the new sample,

detecting action potentials in the new neural signal, each action potential having an amplitude,

using unsupervised clustering to classifying the action potentials into clusters,

re-identifying the dominant cluster, and

recalculating the current signal quality metric as a function of the amplitude of the action potentials of the dominant cluster.

19. The method of claim 18, wherein using unsupervised clustering to classify the action potentials into clusters comprises

representing each of the action potentials by a predetermined number of features;

clustering each of the action potentials using the predetermined number of features into one of a predetermined number of clusters or an outlier cluster; and

determining the predetermined number of clusters by calculating the Bayesian Information Criterion value for a plurality of candidate numbers of clusters and setting the predetermined number of clusters equal to the candidate number of clusters that produced the largest Bayesian Information Criterion value.

20. The method of claim 18, wherein using unsupervised clustering to classify the action potentials into clusters comprises

representing each of the action potentials by a predetermined number of principal components;

and

clustering each of the action potentials using the predetermined number of principal components into one of a predetermined number of clusters or an outlier cluster.

21. The method of claim 18, wherein using unsupervised clustering to classify the action potentials into clusters comprises

representing each of the action potentials by a predetermined number of Haar wavelet coefficients;

and

clustering each of the action potentials using the predetermined number of Haar wavelet coefficients into one of a predetermined number of clusters or an outlier cluster.

22. The method of claim 18, wherein using unsupervised clustering to classify the action potentials into clusters comprises

representing each of the action potentials by a predetermined number of features;

and

clustering each of the action potentials using the predetermined number of features into one of a predetermined number of clusters or an outlier cluster by

constructing a finite mixture model to model the probability density function of action potentials in a feature space defined by the predetermined number of features used to represent each of the action potentials, wherein the finite mixture model comprises a plurality of component probability density functions, and one of the component probability density functions of the plurality corresponds to the outlier class and the other component probability density functions correspond to the predetermined number of classes; and

assigning each of the action potentials to one of the component probability density functions based upon the probability the action potential as represented by the predetermined number of features belongs to each of the component probability density functions.

23. The method of claim 22, wherein the component probability density function corresponding to the outlier cluster is a uniform distribution and the component probability density functions corresponding to the predetermined number of clusters are Gaussian distributions.

24. The method of claim 18, wherein using unsupervised clustering to classify the action potentials into clusters comprises

representing each of the action potentials by a predetermined number of features;

clustering each of the action potentials using the predetermined number of features into one of a predetermined number of clusters or an outlier cluster; and

determining the predetermined number of clusters comprising;

determining a plurality of candidate predetermined number of clusters;

for each of the candidate predetermined number of clusters, constructing a finite mixture model to model the probability density function of action potentials in a feature space defined by the predetermined number of features used to represent each of the action potentials, wherein the finite mixture model comprises a plurality of component probability density functions, and one of the component probability density functions of the plurality corresponds to the outlier cluster and the other component probability density functions correspond to the candidate predetermined number of clusters;

for each of the finite mixture models constructed, determining a likelihood the action potentials as represented by the predetermined number of features belong to the finite mixture model constructed; and

setting the predetermined number of classes equal to the candidate predetermined number of classes used to construct the finite mixture model corresponding to the greatest likelihood.

25. The method of claim 18, wherein the model of the isolation curve comprises a sum of a predetermined number of basis functions having the form u⁽ⁱ⁻

1) wherein u is the position of the electrode when the sample was recorded and i takes the value of a natural number in a set from one to the predetermined number of basis functions.

26. The method of claim 25, wherein the predetermined number of basis functions is determined using a Bayesian Model Section.

27. The method of claim 18, wherein the model of the isolation curve is created using unsupervised regression analysis.

28. A system comprising:

a microdrive comprising an electrode disposed within a brain, the electrode being configured to detect a neural signal of the brain, and a piezoelectric actuator configured to move the electrode along a predetermined path of travel in response to a command signal;

a memory comprising instructions; and

a processor coupled to the memory, coupled to the electrode, and configured to execute the instructions and provide the command signal to the electrode,

wherein the instructions comprise instructions directing the processor to generate the command signal, the instructions comprising;

instructions for generating the command signal to move the electrode a predetermined search distance, after the electrode is moved detect action potentials in the neural signal detected by the electrode, and until action potentials are not detected repeatedly generate the command signal to move the electrode the predetermined search distance,

instructions for generating a command signal to move the electrode a second predetermined distance a predetermined number of times, and after each move, record the neural signal, and

instructions for using the predetermined number of recorded neural signals detected by the electrode to model an isolation curve, use the isolation curve to determine a move distance, and if the move distance is greater than a minimum move threshold, generate a command signal to move the electrode the move distance.

29. A computer readable-medium having instructions executable by a processor coupled to a microdrive comprising a movable electrode, the instructions comprising:

instructions for detecting action potentials comprising;

instructions for determining a search distance,

instructions for moving the electrode of the microdrive the search distance,

after the move, instructions for recording the neural signal,

instructions for determining whether action potentials are present in the recorded neural signal, and

if action potentials are not present in the recorded neural signal,

instructions directing the processor to execute the instructions for detecting action potentials;

instructions for moving the electrode to a plurality of electrode positions and after each move, recording the neural signal to create a plurality of neural signal recordings each having a corresponding recorded electrode position;

instructions for modeling an isolation curve of a neuron having an order using the neural signal recordings and corresponding recorded electrode positions and if the order is greater than a predetermined order threshold, instructions for using the model of the isolation curve to determine a move distance to a new location that the model of the isolation curve indicates will improve signal quality; and

instructions for iteratively moving the electrode comprising;

comparing the move distance to a move threshold value, if the move distance is greater than the move threshold value, moving the electrode of the microdrive the move distance to a new electrode position,

after the move, recording a new neural signal and the new electrode position,

determining a signal quality metric using the new neural signal, and

comparing the signal quality metric to a minimum signal quality threshold, if the signal quality metric is below the minimum signal quality threshold, modifying the plurality of neural signal recordings to include the new neural signal, modifying the electrode positions corresponding to the plurality of neural signal recordings to include the new electrode position, and executing the instructions for modeling the isolation curve of a neuron and the instructions for iteratively moving the electrode.

30. The computer readable-medium of claim 29, comprising instructions for comparing the signal quality metric to a maximum signal quality threshold and if the signal quality metric exceeds the maximum signal quality threshold, determining a back away distance and moving the electrode the back away distance.

31. The computer readable-medium of claim 29, comprising instructions for comparing the signal quality metric to a minimum track signal quality threshold and if the signal quality metric is less than the minimum track signal quality threshold, executing the instructions for detecting action potentials.

32. The computer readable-medium of claim 29, comprising instructions for maintaining neural signal quality comprising instructions for periodically recording a sample neural signal at a sample electrode position, determining a signal quality metric using the sample neural signal, comparing the signal quality metric to the minimum signal quality threshold, if the signal quality metric is below the minimum signal quality threshold, modifying the plurality of neural signal recordings to include the sample neural signal, modifying the electrode positions corresponding with the plurality of neural signal recordings to include the sample electrode position, and executing the instructions for modeling the isolation curve of a neuron and the instructions for iteratively moving the electrode.

33. The computer readable-medium of claim 29, comprising instructions for maintaining neural signal quality comprising instructions for periodically recording a sample neural signal, determining a signal quality metric using the sample neural signal, comparing the signal quality metric to a minimum track signal quality threshold and if the signal quality metric is less than the minimum track signal quality threshold, executing the instructions for detecting action potentials.