Automated seizure detection in the newborn

Seizures in newborn babies are commonly caused by problems such as lack of oxygen, haemorrhage, meningitis, infection and strokes. Seizures are often missed because babies do not exhibit obvious clinical change during the seizures.

The best tool for diagnosing seizures is the EEG, a test which uses electrodes placed on the surface of the scalp to measure the electrical activity of the brain. Newborn EEG seizure and nonseizure waveforms are shown in Figure 1.

Figure 1: Newborn EEG seizure and nonseizure patterns.  

EEG monitoring however requires highly specialised technical and medical personnel to acquire and interpret the results. Most neonatal units in Europe lack this expertise and seizures often go undiagnosed. There is an urgent need to develop an automated (via the computer) seizure detection system that is simple to operate, easy to interpret, and that provides reliable accurate information as soon as possible. An outline of the automated EEG seizure problem is shown in Figure 2.

Figure 2: The automated seizure detection problem.

Early detection of seizures will allow prompt and effective treatment, which should translate into better long term neurological outcome for the smallest and most vulnerable members of the population. 

We are currently developing several automated techniques based on measurements of the time, frequency, nonstationary and nonlinear characteristics of the EEG. The process of seizure detection can be broken down according to Figure 3.

Figure 3: Automated seizure detection algorithms.

We have used filtering, independent component analysis, and principle component analysis to condense the information in the 12 channels of EEG into more manageable quantities. We also have an artifact removal system that removes (or quarantines) EEG data that is generated by large patient movements, EEG lead disconnections and amplifier saturations. We then segment each channel of the EEG recording into overlapping 8 second epochs for processing.

We automatically extract several features from each channel of an EEG recording that are passed to the classifier to determine the presence or absence of seizure. We have used several parametric and nonparametric features derived from each epoch of EEG. These features include;

Parametric

• Linear AR Models
• Gaussian Process Models

Nonparametric

• Time domain features
• Frequency domain features
• Wavelet features
• Entropy features

 The response of a subset of these features as the EEG changes from nonseizure to seizure is shown in Figure 4.

Figure 4: The response of several features to the presence of seizure.

We have used discriminant (linear, quadratic, and regularised) classifiers, support vector machines and Gaussian mixture models in classification. These classifiers take in the features generated and convert them into a probability that seizure is present. An example of this probability function for the EEG trace shown in Figure 4 is shown in Figure 5.

Figure 5: The output of the classifier.

The postprocessing step involves interpreting the classifier output and making a decision. In our case we filter the classifier output and then use a threshold to determine if there is seizure on any EEG channel.

There is much debate as to how to best present the results of seizure detection algorithms with contingency tables, receiver operator characteristics and precision/recall curves all being employed. There is also debate as to exactly how to assess the algorithm with event related assessment (if any part of a seizure is detected then the seizure is assumed detected) being used as well as direct, epoch by epoch or second by second, assessment. A summary of these quantities based on the results of our latest generation detection algorithm are shown below.

Figure 6: The various assessments of algorithm performance.

Further information regarding our seizure detection algorithms can be found in,

1. B.R. Greene, W.P. Marnane, G. Lightbody, R.B. Reilly and G.B. Boylan, "Classifier models and architectures for EEG-based neonatal seizure detection", Physiological Measurement, vol. 29, pp. 1157-1178, 2008.
2. E.M. Thomas, B.R. Greene, G. Lightbody, W.P. Marnane and G.B. Boylan, "Seizure detection in neonates: improved classification through supervised adaptation", Proceedings of the 30th Annual International IEEE EMBS Conference, Vancouver, Canada, pp. 903-906, 20-24 August 2008.
3. O.M. Doyle, B.R. Greene, W. Marnane, G. Lightbody, and G.B. Boylan, "Characterisation of heart rate changes and their correlation with EEG during neonatal seizures". Proceedings of the 30th Annual International IEEE EMBS Conference, Vancouver, Canada, pp. 4984-4987, 20-24 August 2008.
4. B.R. Greene, S. Faul, W.P. Marnane, G. Lightbody, I. Korotchikova, and G.B. Boylan, "A comparison of quantitative EEG features for neonatal seizure detection", Clinical Neurophysiology, vol. 119, pp. 1248-1261, 2008.
5. B.R. Greene, G.B. Boylan, W.P. Marnane, G. Lightbody, and S. Connolly, "Automated single channel seizure detection in the neonate". Proceedings of the 30th Annual International IEEE EMBS Conference, Vancouver, Canada, pp. 915-918, 20-24 August 2008.
6. S. Faul, G. Gregorcic, G. Boylan, W. Marnane, G. Lightbody, and S. Connolly, "Gaussian process modeling of EEG for the detection of neonatal seizures", IEEE Transactions on Biomedical Engineering, vol. 54, no. 12, pp. 2151-2162, December 2007.
7. B.R. Greene, G.B. Boylan, R.B. Reilly, P. de Chazal, and S. Connolly, "Combination of EEG and ECG for improved automatic neonatal seizure detection", Clinical Neurophysiology, vol. 118, pp.1348-1359, 2007.
8. B.R. Greene, P. de Chazal, G.B. Boylan, S. Connolly, and R.B. Reilly, "Electrocardiogram based neonatal seizure detection", IEEE Transactions on Biomedical Engineering, vol. 54, no. 4, pp. 673-682, April 2007.
9. B.R. Greene, R.B. Reilly, G. Boylan, P. de Chazal, and S. Connolly, "Multi-channel EEG based neonatal seizure detection", Proceedings of the 28th IEEE EMBS Annual International Conference, New York City, USA, pp. 4679-4684, 30 August-3 September 2006.
10. G.B. Boylan and J.M. Rennie, "Automated neonatal seizure detection", Clinical Neurophysiology, vol. 117, no. 7, pp. 1412-1413, July 2006.
11. S. Faul, L. Marnane, G. Lightbody, G. Boylan, and S. Connolly, "A method for the blind source separation of sources for use as the first stage of a neonatal seizure detection system", Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, Philadelphia, USA, pp. 409-412, 18-23 March 2005.