EEG Feature Pre-processing for Neonatal Epileptic Seizure Detection

Abstract

Aim of our project is to further optimize neonatal seizure detection using support vector machine (SVM). First, a Kalman filter (KF) was used to filter both feature and classifier output time series in order to increase temporal precision. Second, EEG baseline feature correction (FBC) was introduced to reduce inter patient variability in feature distributions. The performance of the detection methods is evaluated on 54 multi channel routine EEG recordings from 39 both term and pre-term newborns. The area under the receiver operating characteristics curve (AUC) as well as sensitivity and specificity are used to evaluate the performance of the classification method. SVM without KF and FBC achieves an AUC of 0.767 (sensitivity 0.679, specificity 0.707). The highest AUC of 0.902 (sensitivity 0.801, specificity 0.831) is achieved on baseline corrected features with a Kalman smoother used for training data pre-processing and a KF used to filter the classifier output. Both FBC and KF significantly improve neonatal epileptic seizure detection. This paper introduces significant improvements for the state of the art SVM based neonatal epileptic seizure detection.

Appendix

Kalman Filter

It is assumed that epilepsy is observed through the continuous and noisy features z _k calculated from the EEG. By modelling the behaviour of z _k by means of a state space model the KF can be applied to obtain a filtered version $\hat{f}_{\text{k}}$ of the noiseless feature value f _k. The behaviour of z _k can be modelled as the result of a noisy measurement process:

$z_{\text{k}} = f_{\text{k}} + v_{\text{k}}$ with v _k the zero-mean measurement noise with standard deviation σ _v. Furthermore it is assumed that f _k undergoes smooth transitions between seizure and NS states. This behaviour can be enforced by a white noise acceleration model (WNA) with a nearly-constant rate of change.3 We now introduce the state vector $s_{\text{k}} = \left[ {f_{\text{k}} ,\dot{f}_{\text{k}} } \right]$ with $\dot{f}_{\text{k}}$ the rate of change of f _k. The signal z _k can now be described by the following state-space model:

$$\begin{array}{*{20}c} {s_{{{\text{k}} + 1}} = \left[ {\begin{array}{*{20}c} 1 & T \\ 0 & 1 \\ \end{array} } \right]s_{\text{k}} + w_{\text{k}} } \\ {z_{\text{k}} = \left[ {\begin{array}{*{20}c} 1 & 0 \\ \end{array} } \right]s_{\text{k}} + v_{\text{k}} } \\ \end{array} ,$$

(1)

T represents the sample interval and w _k the zero-mean process disturbance with covariance

$$Q = \left[ {\begin{array}{*{20}c} {\sigma_{\text{w}}^{2} \frac{{T^{3} }}{3}} & {\sigma_{\text{w}}^{2} \frac{{T^{2} }}{2}} \\ {\sigma_{\text{w}}^{2} \frac{{T^{2} }}{2}} & {\sigma_{\text{w}}^{2} T} \\ \end{array} } \right],$$

(2)

$\sigma_{\text{w }}$ is the assumed standard deviation of the random fluctuations of $\dot{f}_{\text{k}}$. The KF can now be applied to Eq. (1) in order to recursively calculate a filtered estimate $\hat{f}_{\text{k}}$ of f _k. Instead of the calculated feature z _k the filtered estimate of the noiseless feature $\hat{f}_{\text{k}}$ will be used in the training and classification procedure. The gain of the KF only depends on the ratio $\sigma_{\text{KF}} \triangleq \sigma_{{{\text{w}} }} /\sigma_{{{\text{v}} }}$, $\sigma_{\text{v}}$ is set to 1 and $\sigma_{\text{w }}$ is varied from 10⁻⁶ to 10⁻² (in 9 equally log-spaced steps) in order to tune the filter’s performance. Different optimal filter gains ($\sigma_{\text{w }}$) were found for the KF-test (1 × 10⁻³) and the KS-train (1 × 10⁻³) in the train phase.

The amount of filtering that is applied by the KF depends on the gain $\sigma_{\text{w }}$ which is the only parameter that has to be set by the user. The value of $\sigma_{\text{w }}$will be varied to tune filter performance.