Early detection of epileptic seizures based on parameter identification of neural mass model
Introduction
Epilepsy is defined by Fisher et al. [1] as “a disorder of the brain characterised by an enduring predisposition to generate epileptic seizures and by the neurobiologic, cognitive, psychological, and social consequences of this condition”. Epilepsy has different etiologies, may affect both children and adults, and has an active incidence of 4–8/1000. Seizure detection algorithms are often used during presurgical evaluations for identification as well as for retrospective analysis of seizures [2]. For those algorithms, detection is based on the analysis of the entire seizure, often lasting over a minute [3], leading to relatively high detection accuracy. Various other applications also require that seizure detection be performed at an early stage (i.e. close to the onset). Algorithms answering those needs, called early seizure detection algorithms, face a greater challenge as the occurrence of a seizure has to be evaluated shortly after it begins [2]. However, they open new horizons for various clinical applications.
If reliable, such an algorithm could be implemented in an automatic warning system that alerts the patient of the occurrence of a seizure at an early stage. In addition, this system could greatly benefit medical staff, allowing them to assess which specific functions may be impaired by a seizure and helping them to localise the source of the seizure activity [4]. Furthermore, electrical stimulation, when applied in an appropriate manner at seizure onset, has been reported to suppress spontaneous epileptiform activity [5], [6], [7], [8], [9]. Using an early seizure detection algorithm, an ambitious system could therefore work in a closed loop to detect and terminate electrographic seizures [10], [11]. Recently, such a system has been positively evaluated in a clinical setting [12].
Various seizure detection algorithms have previously been proposed. Most of them are based on a feature extraction used as input data by different types of classifiers, such as fuzzy-logic [13], [14], neural network [13], [15], [16], support vector machines [17], [18], or linear separation based on feature-channel matrix [19]. The methods previously chosen to extract useful features from the recording include wavelet transforms [13], [18], [20], differential operators [21], eigen-decomposition [22] and Gabor functions [16]. However, performances are far from ideal [4].
Physiologically based models are attractive for seizure detection, as their parameters can be explicitly related to neurological mechanisms. Therefore, changes in underlying physiological processes (that are potentially unidentified) may be related to parameter shifts over time [23]. However, to our knowledge, only one model-based seizure detection algorithm has been proposed so far [24]. It is based on a model adapted from Lopes da Silva et al. [25] and applied to scalp EEG of two newborn babies. It was not intended for early seizure detection.
EEG modelling has been addressed in numerous works. One strategy is to use coarse-grained cortical units to compose cortical tissue models (see Coombes [26] for review). Compared to fine-grained models, this approach has the advantage of a low-dimensional parameter space, avoiding the need for detailed knowledge of single neuron properties. One of the best-known neural mass models is that proposed by Jansen and Rit [27], which is based on the original work of Lopes da Silva et al. [25] and further studied by Wendling et al. [28] in the framework of the analysis of epileptic signals.
This paper presents an early seizure detection algorithm that relies on parameter identification of a model based on the work of Jansen and Rit [27]. To our knowledge, it is the first time a neural mass model is used for early seizure detection. Seizure occurrence is detected by analysing the shifts over time of key model parameters based on physiological mechanisms. In this regard, a link was made between physiologically based models and seizure detection algorithms, which allows relating explicitly neurological mechanisms to the occurrence of a seizure. The ability of the algorithm to detect seizures was evaluated against the manual scoring by a human expert on Intracranial EEG (IEEG) samples from 16 patients.
Section snippets
EEG recordings
A group of 16 patients with refractory focal epilepsy who underwent IEEG between May 2005 and February 2010 was retrospectively selected from the database of Erasme University Hospital. All patients underwent IEEG with 500 Hz sampling frequency and 16-bit resolution. Table 1 summarises the details of the IEEG data used in this study for each patient. In total, 73.52 h of IEEG data, containing 82 seizures, were analysed. Five patients (A, C, E, F, and O) were diagnosed with Temporal Lobe Epilepsy
Results
Individual results are given in Table 1. The two patients in the training set (patients A and B) respectively showed a sensibility of 100.0%, a specificity of 99.95% and 99.79%, an FPR of 0.32 and 1.43 per hour, and a delay of 9.9 s and 13.7 s.
In the test set, sensitivities, specificities and FPR differed among patients. The sensitivity ranged from 0% to 100% (56.4%±42.7%), the specificity ranged from 98.82% to 100% (99.84%±0.32%) and the FPR ranged from 0 to 7.5 per hour (1.01±2.02 per hour)
Discussion
We have shown that a model-based approach may be used for seizure detection by analysing parameter shifts over time. Results were better for TLE patients than for the other patient groups. When tested on TLE patients, the sensitivity and FPR were 95.0% and 0.20 per hour, whereas for the other patients sensitivity and FPR were 40.9% and 1.33 per hour. The ROC curves show that the parameters used for seizure detection decision were well chosen and could be adapted for higher sensitivity or lower FPR
Summary
Physiologically based models are attractive for seizure detection, because their parameters can be explicitly related to neurological mechanisms. Therefore, changes in underlying physiological processes (that are possibly unidentified) may be related to parameter shifts over time [23]. EEG modelling has been addressed previously. One strategy is to use coarse-grained cortical units to compose cortical tissue models (see Coombes [26] for review). Compared to fine-grained models, this approach
Conflict of interest statement
None Declared.
Acknowledgement
This work has been funded by the “Fonds pour la Formation à la Recherche dans l’Industrie et dans l’Agriculture” (F.R.I.A.). The authors would like to thank Ir. Rudy Ercek and Pr. Jean Schoentgen from the Laboratory of Image, Signal and Telecommunications, and Dr. Emilio Araujo Mino from the School of Medicine of the University of New Mexico for their help on this paper.
References (50)
- et al.
Improving early seizure detection
Epilepsy Behav.
(2011) - et al.
An automatic warning system for epileptic seizures recorded on intracerebral EEGs
Clin. Neurophysiol.
(2005) - et al.
Effects of applied currents on epileptiform bursts in vitro
Exp. Neurol.
(1991) - et al.
Suppression of spontaneous epileptiform activity with applied currents
Brain Res.
(1991) - et al.
Effects of applied currents on spontaneous epileptiform activity induced by low calcium in the rat hippocampus
Brain Res.
(1998) Application of adaptive neuro-fuzzy inference system for epileptic seizure detection using wavelet feature extraction
Comput. Biol. Med.
(2007)- et al.
A fuzzy rule-based system for epileptic seizure detection in intracranial EEG
Clin. Neurophysiol.
(2009) - et al.
A novel real-time patient-specific seizure diagnosis algorithm based on analysis of EEG and ECG signals using spectral and spatial features and improved particle swarm optimization classifier
Comput. Biol. Med.
(2012) - et al.
Detection of seizures in EEG using subband nonlinear parameters and genetic algorithm
Comput. Biol. Med.
(2010) - et al.
Incorporating structural information from the multichannel EEG improves patient-specific seizure detection
Clin. Neurophysiol.
(2012)
An automatic patient-specific seizure onset detection method in intracranial EEG based on incremental nonlinear dimensionality reduction
Comput. Biol. Med.
Differential operator in seizure detection
Comput. Biol. Med.
Comparison of subspace-based methods with AR parametric methods in epileptic seizure detection
Comput. Biol. Med.
Characterising the dynamics of EEG waveforms as the path through parameter space of a neural mass model: application to epilepsy seizure evolution
Neuroimage
Large-scale neural dynamics: simple and complex
Neuroimage
Actions of noradrenaline on neocortical neurons in vitro
Brain Res.
Full-band EEG (FbEEG): an emerging standard in electroencephalography
Clin. Neurophysiol.
Spike detection algorithm automatically adapted to individual patients applied to spike-and-wave percentage quantification
Neurophysiol. Clin.
Automatic recognition of epileptic seizures in the EEG
Electroencephalogr. Clin. Neurophysiol.
Computerized seizure detection of complex partial seizures
Electroencephalogr. Clin. Neurophysiol.
Performance assessment for EEG-based neonatal seizure detectors
Clin. Neurophysiol.
Computerized epileptiform transient detection in the scalp electroencephalogram: obstacles to progress and the example of computerized ECG interpretation
Clin. Neurophysiol.
Spike detection: a review and comparison of algorithms
Clin. Neurophysiol.
Bifurcation analysis of neural mass models: impact of extrinsic inputs and dendritic time constants
Neuroimage
Self-organised transients in a neural mass model of epileptogenic tissue dynamics
Neuroimage
Cited by (20)
Automated epileptic seizure detection based on break of excitation/inhibition balance
2019, Computers in Biology and MedicineCitation Excerpt :Second, the implicit assumption underlying the model-based seizure detection approach is that the dynamics of ictogenesis can be captured by smooth variations of several system parameters, which was true in TLE, but not in FLE [46]. It was proposed that TLE seizures were caused by a deformation of the attractor leading to a gradual evolution onto the ictal state, thus seizures can be detected by analyzing the gradual change in dynamics [18]. This contrast with FLE seizures that were reported to be caused by a perturbation in a bistable state without change in parameters [47] and therefore might be rather difficult to detect [46].
Tracking slow modulations in synaptic gain using dynamic causal modelling: Validation in epilepsy
2015, NeuroImageCitation Excerpt :A recent study (Nevado-Holgado et al., 2012) characterised the evolution of an absence seizure as a path through the parameter space of a neural mass model. In another approach (Hocepied et al., 2013) a similar scheme was proposed for early seizure detection. In both cases, the authors suggest that tracking a set of parameters over time can disclose the nature of ictogenesis.
Functional brain connectivity from EEG in epilepsy: Seizure prediction and epileptogenic focus localization
2014, Progress in NeurobiologyCitation Excerpt :They studied the parameters trajectory during 1 Hz electrical stimulation in 20 patients and were able to reveal which structures expressed a strong modulatory input to the epileptic focus. In a more recent paper, effective connectivity was used by Hocepied et al. (2013) to predict seizures based on a neural mass model. Using a physiologically based model, key parameters to underlying neurological mechanisms can be estimated.
Revisiting Seizure Prediction with Path Signatures: the Limitations of System Identification
2024, 2024 Australian and New Zealand Control Conference, ANZCC 2024Computational Evidence for a Competitive Thalamocortical Model of Spikes and Spindle Activity in Rolandic Epilepsy
2021, Frontiers in Computational Neuroscience