Unified Probabilistic Framework For Predicting And Detecting Seizure Onsets In The Brain And Multitherapeutic Device

• US 20070276279A1
• Filed: 08/13/2007
• Published: 11/29/2007
• Est. Priority Date: 10/20/2000
• Status: Abandoned Application

• Current Assignee: Trustees Of The University Of Pennsylvania (University of Pennsylvania)
• Original Assignee: Trustees Of The University Of Pennsylvania (University of Pennsylvania)
• Inventors: Litt, Brian, Esteller, Rosana, Vachtsevanos, George, Echauz, Javier
• Application Number:
  US11/837,801
• Publication Number:
  US 20070276279A1
• Time in Patent Office: Days
• US Class Current: 600/544
• CPC Class Codes: A61B 5/076 Permanent implantations telemetry using implanted circuitry A61B5/0031; implanted stimulators for therapy A61N1/3605, A61N1/362, A61N1/372
• CPC Class Codes: A61B 5/369 Electroencephalography [EEG] devices for psychotechnics A61B5/16
• CPC Class Codes: A61B 5/37 Intracranial electroencephalography [IC-EEG], e.g. electrocorticography [ECoG]
• CPC Class Codes: A61B 5/372 Analysis of electroencephalograms
• CPC Class Codes: A61B 5/4094 Diagnosing or monitoring seizure diseases, e.g. epilepsy
• CPC Class Codes: A61B 5/7267 involving training the classification device
• CPC Class Codes: A61B 5/7275 Determining trends in physiological measurement data; Predicting development of a medical condition based on physiological measurements, e.g. determining a risk factor
• CPC Class Codes: A61N 1/36017 with leads or electrodes penetrating the skin
• CPC Class Codes: A61N 1/36064 Epilepsy
• CPC Class Codes: A61N 1/36082 Cognitive or psychiatric applications, e.g. dementia or Alzheimer's disease
• CPC Class Codes: A61N 1/36135 using physiological parameters
• CPC Class Codes: G06F 18/217 Validation; Performance evaluation; Active pattern learning techniques
• CPC Class Codes: G06F 18/2413 based on distances to training or reference patterns
• CPC Class Codes: G06F 18/24155 Bayesian classification
• CPC Class Codes: G16H 20/30 relating to physical therapies or activities, e.g. physiotherapy, acupressure or exercising
• CPC Class Codes: G16H 50/20 for computer-aided diagnosis, e.g. based on medical expert systems
• CPC Class Codes: G16H 50/30 for calculating health indices; for individual health risk assessment
• CPC Class Codes: G16H 50/50 for simulation or modelling of medical disorders

Abstract

A method and an apparatus for predicting and detecting epileptic seizure onsets within a unified multiresolution probabilistic framework, enabling a portion of the device to automatically deliver a progression of multiple therapies, ranging from benign to aggressive as the probabilities of seizure warrant. Based on novel computational intelligence algorithms, a realistic posterior probability function P(S_T|x) representing the probability of one or more seizures starting within the next T minutes, given observations x derived from IEEG or other signals, is periodically synthesized for a plurality of prediction time horizons. When coupled with optimally determined thresholds for alarm or therapy activation, probabilities defined in this manner provide anticipatory time-localization of events in a synergistic logarithmic-like array of time resolutions, thus effectively circumventing the performance vs. prediction-horizon tradeoff of single-resolution systems. The longer and shorter prediction time scales are made to correspond to benign and aggressive therapies respectively. The imminence of seizure events serves to modulate the dosage and other parameters of treatment during open-loop or feedback control of seizures once activation is triggered. Fast seizure onset detection is unified within the framework as a degenerate form of prediction at the shortest, or even negative, time horizon. The device is required to learn in order to find the probabilistic prediction and control strategies that will increase the patient'"'"'s quality of life over time. A quality-of-life index (QOLI) is used as an overall guide in the optimization of patient-specific signal features, the multitherapy activation decision logic, and to document if patients are actually improving.

Claims

1. A method for periodic learning to improve and maintain the performance of a device implanted in an individual subject to seizures to provide treatment therapies, comprising the steps of:

marking a time of unequivocal electrographic onset (UEO) in each recorded seizure over a fixed period of time;

creating a plurality of sets of learning data based on the UEOs by clipping a plurality of intracranial electroencephalographic (IEEG) epochs immediately preceding seizures and labeling the clipped epochs as preseizure raw data;

clipping and labeling randomly chosen, non-overlapping IEEG data as nonpreseizure raw data;

generating a time series of each feature in a feature library from the preseizure and nonpreseizure raw data;

searching for an optimal feature vector in the feature library to minimize a classifier-based performance metric;

synthesizing a posterior probability estimator for the optimal feature vector; and

coupling an optimal therapy activation threshold to the probability estimator.

2. The method for periodic learning of claim 1 further comprising joining and weighting the preseizure and nonpreseizure raw data sets with corresponding data sets for a plurality of preceding fixed periods of time so that the preseizure and nonpreseizure raw data sets have an incremental effect on learning.

3. The method for periodic learning of claim 2 wherein the plurality of preseizure and nonpreseizure data sets corresponds to a multiple period rectangular moving window of learning data.

4. The method for periodic learning of claim 1 wherein searching for an optimal feature vector further comprises:

generating trajectories of candidate feature vectors by time synchronizing the time series of each feature in the feature library;

synthesizing nonparametric classifiers that memorize a training data set to simulate therapy activation decision rules and accelerate feature optimization;

correcting a discriminant function of the nonparametric classifiers to determine an optimal decision rule;

measuring an overall risk over a validation set of data that is not directly used for synthesizing the nonparametric classifiers; and

determining a next candidate feature vector using a heuristic search criterion.

5. The method for periodic learning of claim 1 wherein the step of synthesizing a posterior probability estimator further comprises:

training a wavelet neural network to calculate posterior probability estimators;

minimizing an expected value of a squared error loss function on a validation set of data; and

selecting a wavelet neural network that minimizes the error over the validation set of data.

6. The method for periodic learning of claim 1 further comprising ranking the feature vectors using a plurality of sets of classifier-based performance metrics.

7. The method for periodic learning of claim 6 wherein a first set of classifier-based performance metrics includes a probability of a true positive, a probability of a false negative, a probability of a true negative, a probability of a false positive, a probability of correct classification, a probability of error, a positive predictive value and a class balance value.

8. The method for periodic learning of claim 6 wherein a second set of classifier-based performance metrics includes a prior probability of preseizure and a prior probability of nonpreseizure.

9. The method for periodic learning of claim 6 wherein a third set of classifier-based performance metrics includes a false positives per hour measure, an average detection delay measure, an error risk measure and an overall risk measure.

10. The method for periodic learning of claim 9 wherein the error risk measure is determined by applying penalty factors to the probability of a false negative and the probability of a false positive.

11. The method for periodic learning of claim 9 wherein the overall risk measure penalizes all therapy activations.

12. The method for periodic learning of claim 9 wherein the overall risk measure is determined by applying penalty factors to the probability of a false negative, the probability of a false positive, the probability of a true positive and the probability of a true negative.

13. The method for periodic learning of claim 1 wherein searching for an optimal feature vector further comprises using an optimality criterion to determine an optimal decision rule.

14. The method for periodic learning of claim 4 wherein the nonparametric classifiers that memorize the training data set to simulate therapy activation decision rules and accelerate feature optimization include k-nearest neighbors (kNNs), probabilistic neural networks (PNNs), and hybrids.

15. The method for periodic learning of claim 4 wherein correcting a discriminant function of the nonparametric classifier includes applying a correction factor to the discriminant function to correct for prior probability mismatches between the a priori probability of seizure estimated from data and the true prior probability of seizure.

16. The method for periodic learning of claim 1 wherein searching for an optimal feature vector includes performing a forward sequential search through the feature library.

17. A system for periodic learning to improve and maintain the performance of a device implanted in an individual subject to seizures in providing treatment therapies comprising:

a signal acquisition component to condition and digitize a plurality of intracranial electroencephalographic (IEEG) signals received from a transducer implanted in the individual;

a storage medium for storing the plurality of digitized IEEG signals as IEEG data, and for storing a library of feature vectors generated from the IEEG data;

a processor for executing a learning and training component including;

a module for marking a time of unequivocal electrographic onset (UEO) in each recorded seizure over a fixed period of time;

a module for creating learning sets of data based on the UEOs by clipping all the IEEG epochs immediately preceding seizures and labeling the clipped epochs as preseizure raw data;

a module for clipping and labeling randomly chosen, non-overlapping IEEG data as nonpreseizure raw data;

a module for generating a time series of each feature in a feature library from the preseizure and nonpreseizure raw data;

a module for searching for an optimal feature vector in the feature library to minimize a classifier-based performance metric;

a module for synthesizing a posterior probability estimator for the optimal feature vector; and

a module for coupling an optimal therapy activation threshold to the probability estimator.

18. The system for periodic learning of claim 17 wherein the learning and training component further comprises:

a module for generating trajectories of candidate feature vectors by time synchronizing the time series of all features in the feature library;

a module for synthesizing nonparametric classifiers that memorize a training data set to simulate therapy activation decision rules and accelerate feature optimization;

a module for correcting a discriminant function of the nonparametric classifiers to determine an optimal decision rule;

a module for measuring an overall risk over a validation set of data that is not directly used for synthesizing the nonparametric classifiers; and

a module for determining a next candidate feature vector using a heuristic search criterion.

19. The system for periodic learning of claim 17 wherein the learning and training component further comprises:

a module for training a wavelet neural network to calculate posterior probability estimators;

a module for minimizing an expected value of a squared error loss function on a validation set of data; and

a module for selecting a wavelet neural network that minimizes the error over the validation set of data.

20. The system for periodic learning of claim 17 wherein the learning and training component further comprises a module for ranking the feature vectors using a plurality of sets of classifier-based performance metrics.

21. The system for periodic learning of claim 20 wherein a first set of classifier-based performance metrics includes a probability of a true positive, a probability of a false negative, a probability of a true negative, a probability of a false positive, a probability of correct classification, a probability of error, a positive predictive value, and a class balance value.

22. The system for periodic learning of claim 20 wherein a second set of classifier-based performance metrics includes a prior probability of preseizure and a prior probability of nonpreseizure.

23. The system for periodic learning of claim 20 wherein a third set of classifier-based performance metrics includes a false positives per hour measure, an average detection delay measure, an error risk measure, and an overall risk measure.

24. The system for periodic learning of claim 23 wherein the overall risk measure is determined by applying penalty factors to the probability of a false negative, the probability of a false positive, the probability of a true positive, and the probability of a true negative.

25. The system for periodic learning of claim 17 wherein the learning and training module further comprises a module for determining an optimal decision rule based on a selected optimality criterion.

26. A computer readable medium containing instructions for causing a computer system to improve and maintain the performance of a device implanted in an individual subject to seizures to provide treatment therapies, the computer readable medium comprising:

program instructions that mark a time of unequivocal electrographic onset (UEO) in each recorded seizure over a fixed period of time;

program instructions that create a plurality of sets of learning data based on the UEOs by clipping a plurality of intracranial electroencephalographic (IEEG) epochs immediately preceding seizures and labeling the clipped epochs as preseizure raw data;

program instructions that clip and label randomly chosen, non-overlapping IEEG data as nonpreseizure raw data;

program instructions that generate a time series of each feature in a feature library from the preseizure and nonpreseizure raw data;

program instructions that search for an optimal feature vector in the feature library to minimize a classifier-based performance metric;

program instructions that synthesize a posterior probability estimator for the optimal feature vector; and

program instructions that couple an optimal therapy activation threshold to the probability estimator.

27. The computer readable medium containing instructions for causing a computer system to improve and maintain the performance of an implanted device of claim 26 wherein the program instructions that search for an optimal feature vector further comprise:

program instructions that generate trajectories of candidate feature vectors by time synchronizing the time series of each feature in the feature library;

program instructions that synthesize nonparametric classifiers that memorize a training data set to simulate therapy activation decision rules and accelerate feature optimization;

program instructions that correct a discriminant function of the nonparametric classifiers to determine an optimal decision rule;

program instructions that measure an overall risk over a validation set of data that is not directly used for synthesizing the nonparametric classifiers; and

program instructions that determine a next candidate feature vector using a heuristic search criterion.

28. The computer readable medium containing instructions for causing a computer system to improve and maintain the performance of an implanted device of claim 26 wherein the program instructions that synthesize a posterior probability estimator further comprise:

program instructions that train a wavelet neural network to calculate posterior probability estimators;

program instructions that minimize an expected value of a squared error loss function on a validation set of data; and

program instructions that select a wavelet neural network that minimizes the error over the validation set of data.

29. The computer readable medium containing instructions for causing a computer system to improve and maintain the performance of an implanted device of claim 26 further comprising program instructions that rank the feature vectors using a plurality of sets of classifier-based performance metrics.

30. The computer readable medium containing instructions for causing a computer system to improve and maintain the performance of an implanted device of claim 26 wherein the program instructions that search for an optimal feature vector further comprise program instructions that use an optimality criterion to determine an optimal decision rule.