Limits...
Wavelet-based Gaussian-mixture hidden Markov model for the detection of multistage seizure dynamics: a proof-of-concept study.

Chiu AW, Derchansky M, Cotic M, Carlen PL, Turner SO, Bardakjian BL - Biomed Eng Online (2011)

Bottom Line: Using unsupervised training method, the HMM and the clustering parameters were obtained simultaneously.Minimum redundancy maximum relevance (mRMR) analysis and Akaike Information Criterion (AICc) were applied to reduce the effect of over-fitting.The subjectivity involved in partitioning the observed data prior to training can be eliminated.

View Article: PubMed Central - HTML - PubMed

Affiliation: Biomedical Engineering Department, Louisiana Tech University, Ruston, Louisiana, USA. alanchiu@latech.edu

ABSTRACT

Background: Epilepsy is a common neurological disorder characterized by recurrent electrophysiological activities, known as seizures. Without the appropriate detection strategies, these seizure episodes can dramatically affect the quality of life for those afflicted. The rationale of this study is to develop an unsupervised algorithm for the detection of seizure states so that it may be implemented along with potential intervention strategies.

Methods: Hidden Markov model (HMM) was developed to interpret the state transitions of the in vitro rat hippocampal slice local field potentials (LFPs) during seizure episodes. It can be used to estimate the probability of state transitions and the corresponding characteristics of each state. Wavelet features were clustered and used to differentiate the electrophysiological characteristics at each corresponding HMM states. Using unsupervised training method, the HMM and the clustering parameters were obtained simultaneously. The HMM states were then assigned to the electrophysiological data using expert guided technique. Minimum redundancy maximum relevance (mRMR) analysis and Akaike Information Criterion (AICc) were applied to reduce the effect of over-fitting. The sensitivity, specificity and optimality index of chronic seizure detection were compared for various HMM topologies. The ability of distinguishing early and late tonic firing patterns prior to chronic seizures were also evaluated.

Results: Significant improvement in state detection performance was achieved when additional wavelet coefficient rates of change information were used as features. The final HMM topology obtained using mRMR and AICc was able to detect non-ictal (interictal), early and late tonic firing, chronic seizures and postictal activities. A mean sensitivity of 95.7%, mean specificity of 98.9% and optimality index of 0.995 in the detection of chronic seizures was achieved. The detection of early and late tonic firing was validated with experimental intracellular electrical recordings of seizures.

Conclusions: The HMM implementation of a seizure dynamics detector is an improvement over existing approaches using visual detection and complexity measures. The subjectivity involved in partitioning the observed data prior to training can be eliminated. It can also decipher the probabilities of seizure state transitions using the magnitude and rate of change wavelet information of the LFPs.

Show MeSH

Related in: MedlinePlus

The rescaled AICc values are plotted with respect to the number of HMM states and Gaussian mixtures. (a) The simplest model with ΔAICc < 0.25 is at Q = 3 and M = 2. In this model, the reduced feature space consists of 2-D wavelet coefficients. (b) With the 4-D wavelet and rate of change features established using mRMR, the simplest model with ΔAICc < 0.25 is found at Q = 5, M = 3.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC3094216&req=5

Figure 7: The rescaled AICc values are plotted with respect to the number of HMM states and Gaussian mixtures. (a) The simplest model with ΔAICc < 0.25 is at Q = 3 and M = 2. In this model, the reduced feature space consists of 2-D wavelet coefficients. (b) With the 4-D wavelet and rate of change features established using mRMR, the simplest model with ΔAICc < 0.25 is found at Q = 5, M = 3.

Mentions: Another method for selecting the appropriate HMM topology was investigated by reducing the feature space dimension and the number of free parameters in the model. Minimum redundancy maximum relevance (mRMR) technique was used to select a subset of relevant wavelet features. It can alleviate the effect of over-fitting caused by the curse of dimensionality and improve the model's ability to generalize. In conjunction with the AICc, which helped balance the LL against the number of model parameters, suitable optimum labelled HMMAIC can be found. A summary of the mRMR analysis is shown in Table 4. The wavelet coefficients associated with the alpha (8 - 15 Hz) and beta (15 - 40 Hz) bands had the largest mutual information quotient (VF/Wc = 0.497 and 0.436 respectively). The wavelet coefficients at these two frequency bands then constituted the reduced features space. For the feature space consisted of the wavelet coefficients (c) only, the simplest 2-D HMM topology with ΔAICc < 0.25 was found at Q = 3, M = 2 (Figure 7a). The state transition diagram of the corresponding HMM is shown in Figure 8a. It contained three bidirectionally connected states. When the posterior probabilities of this model were matched against the LFP after assigning the model states (Figure 8b), no distinction between the tonic firing and the postictal events was found. There were also cases in which the state transitions jumped from interictal to chronic seizure directly and back. ON average, the performance of this HMM was slightly better than the HMMopt7D, with TP = 80.9 ± 34.4%, TN = 94.8 ± 15.6%, ΔT = 3.79 ± 8.67 s and O = 0.813 ± 0.247, even though student T-test analysis did not reveal any statistically significant improvement. When the wavelet rate of change information (Δc) was added as feature, a 4-D reduced feature space was created. The simplest topology with ΔAICc < 0.25 for this feature set was found at Q = 5 and M = 3 (Figure 7b). Figure 9a shows the state transition diagram for this HMM (called HMMAIC). Similar to the HMMopt14D, the state progression was mainly unidirectional with a non-zero state transition probability from the early tonic state back to the interictal state. This suggests that it is possible to have seizure permissive early tonic activity that can be reverted back to the interictal state. The marginal posterior probability γi(t) for each state i of HMMAIC is plotted against a sample test data in Figure 9b. Even though no significant improvement over HMMopt14D was revealed using the student T-test, the HMMAIC gives the best overall performance out of all the HMMs created in this study with TP = 95.7 ± 14.0%, TN = 98.9 ± 6.5%, ΔT = -2.03 ± 7.10 s and O = 0.995 ± 0.129. A summary of the performance measures for the HMMAIC using mRMR and ΔAICc is presented in Table 5.


Wavelet-based Gaussian-mixture hidden Markov model for the detection of multistage seizure dynamics: a proof-of-concept study.

Chiu AW, Derchansky M, Cotic M, Carlen PL, Turner SO, Bardakjian BL - Biomed Eng Online (2011)

The rescaled AICc values are plotted with respect to the number of HMM states and Gaussian mixtures. (a) The simplest model with ΔAICc < 0.25 is at Q = 3 and M = 2. In this model, the reduced feature space consists of 2-D wavelet coefficients. (b) With the 4-D wavelet and rate of change features established using mRMR, the simplest model with ΔAICc < 0.25 is found at Q = 5, M = 3.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC3094216&req=5

Figure 7: The rescaled AICc values are plotted with respect to the number of HMM states and Gaussian mixtures. (a) The simplest model with ΔAICc < 0.25 is at Q = 3 and M = 2. In this model, the reduced feature space consists of 2-D wavelet coefficients. (b) With the 4-D wavelet and rate of change features established using mRMR, the simplest model with ΔAICc < 0.25 is found at Q = 5, M = 3.
Mentions: Another method for selecting the appropriate HMM topology was investigated by reducing the feature space dimension and the number of free parameters in the model. Minimum redundancy maximum relevance (mRMR) technique was used to select a subset of relevant wavelet features. It can alleviate the effect of over-fitting caused by the curse of dimensionality and improve the model's ability to generalize. In conjunction with the AICc, which helped balance the LL against the number of model parameters, suitable optimum labelled HMMAIC can be found. A summary of the mRMR analysis is shown in Table 4. The wavelet coefficients associated with the alpha (8 - 15 Hz) and beta (15 - 40 Hz) bands had the largest mutual information quotient (VF/Wc = 0.497 and 0.436 respectively). The wavelet coefficients at these two frequency bands then constituted the reduced features space. For the feature space consisted of the wavelet coefficients (c) only, the simplest 2-D HMM topology with ΔAICc < 0.25 was found at Q = 3, M = 2 (Figure 7a). The state transition diagram of the corresponding HMM is shown in Figure 8a. It contained three bidirectionally connected states. When the posterior probabilities of this model were matched against the LFP after assigning the model states (Figure 8b), no distinction between the tonic firing and the postictal events was found. There were also cases in which the state transitions jumped from interictal to chronic seizure directly and back. ON average, the performance of this HMM was slightly better than the HMMopt7D, with TP = 80.9 ± 34.4%, TN = 94.8 ± 15.6%, ΔT = 3.79 ± 8.67 s and O = 0.813 ± 0.247, even though student T-test analysis did not reveal any statistically significant improvement. When the wavelet rate of change information (Δc) was added as feature, a 4-D reduced feature space was created. The simplest topology with ΔAICc < 0.25 for this feature set was found at Q = 5 and M = 3 (Figure 7b). Figure 9a shows the state transition diagram for this HMM (called HMMAIC). Similar to the HMMopt14D, the state progression was mainly unidirectional with a non-zero state transition probability from the early tonic state back to the interictal state. This suggests that it is possible to have seizure permissive early tonic activity that can be reverted back to the interictal state. The marginal posterior probability γi(t) for each state i of HMMAIC is plotted against a sample test data in Figure 9b. Even though no significant improvement over HMMopt14D was revealed using the student T-test, the HMMAIC gives the best overall performance out of all the HMMs created in this study with TP = 95.7 ± 14.0%, TN = 98.9 ± 6.5%, ΔT = -2.03 ± 7.10 s and O = 0.995 ± 0.129. A summary of the performance measures for the HMMAIC using mRMR and ΔAICc is presented in Table 5.

Bottom Line: Using unsupervised training method, the HMM and the clustering parameters were obtained simultaneously.Minimum redundancy maximum relevance (mRMR) analysis and Akaike Information Criterion (AICc) were applied to reduce the effect of over-fitting.The subjectivity involved in partitioning the observed data prior to training can be eliminated.

View Article: PubMed Central - HTML - PubMed

Affiliation: Biomedical Engineering Department, Louisiana Tech University, Ruston, Louisiana, USA. alanchiu@latech.edu

ABSTRACT

Background: Epilepsy is a common neurological disorder characterized by recurrent electrophysiological activities, known as seizures. Without the appropriate detection strategies, these seizure episodes can dramatically affect the quality of life for those afflicted. The rationale of this study is to develop an unsupervised algorithm for the detection of seizure states so that it may be implemented along with potential intervention strategies.

Methods: Hidden Markov model (HMM) was developed to interpret the state transitions of the in vitro rat hippocampal slice local field potentials (LFPs) during seizure episodes. It can be used to estimate the probability of state transitions and the corresponding characteristics of each state. Wavelet features were clustered and used to differentiate the electrophysiological characteristics at each corresponding HMM states. Using unsupervised training method, the HMM and the clustering parameters were obtained simultaneously. The HMM states were then assigned to the electrophysiological data using expert guided technique. Minimum redundancy maximum relevance (mRMR) analysis and Akaike Information Criterion (AICc) were applied to reduce the effect of over-fitting. The sensitivity, specificity and optimality index of chronic seizure detection were compared for various HMM topologies. The ability of distinguishing early and late tonic firing patterns prior to chronic seizures were also evaluated.

Results: Significant improvement in state detection performance was achieved when additional wavelet coefficient rates of change information were used as features. The final HMM topology obtained using mRMR and AICc was able to detect non-ictal (interictal), early and late tonic firing, chronic seizures and postictal activities. A mean sensitivity of 95.7%, mean specificity of 98.9% and optimality index of 0.995 in the detection of chronic seizures was achieved. The detection of early and late tonic firing was validated with experimental intracellular electrical recordings of seizures.

Conclusions: The HMM implementation of a seizure dynamics detector is an improvement over existing approaches using visual detection and complexity measures. The subjectivity involved in partitioning the observed data prior to training can be eliminated. It can also decipher the probabilities of seizure state transitions using the magnitude and rate of change wavelet information of the LFPs.

Show MeSH
Related in: MedlinePlus