A Machine learning approach to detect seizures in patients with drug-resistant epilepsy

Luigi Pavone 1, Jaime Delgado Saa 2

1   IRCCS Neuromed, Via Atinense 18, 86077 Pozzilli (IS); Questo indirizzo email è protetto dagli spambots. È necessario abilitare JavaScript per vederlo.

2   Universidad del Norte, Department of electrical and electronics engineering, Colombia

  1. Introduction

Epilepsy is a brain disorder that affects over 40 million people worldwide, it is identified as the world’s second most common brain disorder after stroke [1], and can produce significant morbidity or death when uncontrolled [2]. Epilepsy is a neurological disease characterized by seizures, involving abnormal, rhythmic discharges of cortical neurons.

ElectroEncephaloGraphic (EEG) signal, invasive ElectroCorticoGram (ECoG) and intracranial EEG (iEEG) signals (all referred here as EEG) are important clinical tools for diagnosing and monitoring epilepsy. Seizures are manifested in EEG as paroxysmal events characterized by stereotyped repetitive waveforms that advance in amplitude and frequency before decaying ultimately [3]. Continuous monitoring of EEG signal is required for the detection of clinical seizures, which is made through visual inspection by neurologists. Developing an automated and computerized method to detect seizures from EEG recordings could be a handy tool for clinicians, speeding up the process of seizure detection on EEG and also providing them valuable data for epilepsy management.

Furthermore, the availability of such systems to early detect seizures can improve the quality of life in epileptic patients, embedding them in a closed-loop stimulation system, which electrically stimulates the brain to suppress epileptic activity at seizure onset [4,5].

In this study, we presented a seizure detection system based on Log-Energy Entropy (LogEn), computed on band-passed intracranial EEG recordings. The extracted features are input to a Support Vector Machine (SVM) for classification. The proposed algorithms was tested in 20 patients with excellent results.

  1. Experimental Section

The proposed algorithm for detecting seizures in epileptic EEG signals consists of preprocessing, feature extraction and classification blocks. We tested our algorithm on Freiburg Seizure Prediction EEG (FSPEEG) Database [6,7]. The FSPEEG database was proposed in early 2000 as an EEG database available for downloading to researchers working primarily on seizure detection and prediction.

The database contains intracerebral EEG (iEEG) recordings collected from 21 patients with medically intractable focal epilepsy.  For each patient, there are six EEG traces, identified by certified epileptologists by visual inspection. The first three where selected near to the region where the seizure occurred; the remaining three were selected in areas distal to the seizure focus.

Each EEG (iEEG) channel was first decomposed in three different frequency bands, 1-7 Hz, 8-12 Hz and 13-30 Hz, which are frequency bands correspond to well-known oscillatory phenomena observed in the brain.

In this way, the signal from each electrode was split into three different signals, and for each of those signals, the LogEn was computed using entropy-based wavelet decomposition method presented by Coifman and Wickerhauser [8].

The features extracted from the EEG signals were used to train a classifier based on SVM [9,10], to identify recordings with seizures based on these features.

  1. Results

The performance of the proposed method was evaluated in terms of sensitivity, specificity, error rate and accuracy. We analyzed in total 80 seizures in 20 patients from Freiburg dataset and at least 50 minutes preictal EEG for each seizure, achieving 91,16% of sensitivity and 97,19% of classification accuracy. The high accuracy and sensitivity of our method, together with its low computational cost, makes it feasible to include it in a real-time EEG monitoring system for epileptic seizure detection.

  1. Conclusions

Our method show high classification accuracy, correctly identifying epileptic signals. This result indicates that it may have a high translational value for detection of incoming seizures in drug-resistant patients, with different seizure origin, and different seizure types. Besides, the described algorithm may be used to deliver a fast onset treatment that would stop the seizure, for example in a closed-loop system with brain electrical stimulation via an implantable device. Furthermore, as the band-pass filtering, the log-energy entropy and the SVM implementation are not computationally complex, this algorithm is suitable to be implemented in a closed-loop stimulation device.


  1. Mormann F, Andrzejak R, Elger C, Lehnertz K. Seizure prediction: the long and wind-ing road. Brain 2007;130(2):314–33
  2. Langan Y., ”Sudden unexpected death in epilepsy (SUDEP): risk factors and case control studies.” Seizure 2000;9:179–83
  3. Karayiannis N.B., Mukherjee A., Glover J.R., Ktonas P.Y., Frost J.D., Hrachovy R.A., Mizrahi E.M. "Detection of Pseudo-sinusoidal Epileptic Seizure Segments in the Neonatal EEGby Cascading a Rule-Based Algorithm With a Neural Network", IEEE Transaction On Biomedical Engineering, Vol. 53, No. 4, pp. 633-641, 2006
  4. Gotman, Automatic seizure detection: Improvements and evaluation, Electroencephalogr. Clin. Neurophysiol. 76(4) (1990) 317–324
  5. Gotman, Automatic detection of seizures and spikes, J.Clin. Neurophysiol. 16(2) (1999) 130–140
  6. Aschenbrenner-Scheibe R, Maiwald T, Winterhalder M, Voss HU, Timmer J, Schulze-Bonhage A. How well can epileptic seizures be predicted? an evaluation of a nonlinear method. Brain. 2003;126(12):2616–2626. [PubMed]
  7. Maiwald T, Winterhalder M, Aschenbrenner-Scheibe R, Voss HU, Schulze-Bonhage A, Timmer J. Comparison of three nonlinear seizure prediction methods by means of the seizure prediction characteristic. Physica D. 2004;194(3-4):357–368.
  8. Coifman R.R., Wickerhauser M.V. (1992), "Entropy-based Algorithms for best basis selection" IEEE Trans. on Inf. Theory, vol. 38, 2, pp. 713–718.
  9. Vapnik V. The nature of statistical learning theory. Springer; 1995
  10. Vapnik, V. and A. Lerner, 1963. Pattern recognition using generalized portrait method. Automation and Remote Control. 24, 774-780.