Prediction of epileptic seizures with convolutional neural networks and functional near-infrared spectroscopy signals
Graphical abstract
Introduction
Prediction of epileptic seizures refers to the analysis of brain signals, extracted from an epileptic patient, to detect the future occurrence of a seizure. There are four states within a brain signal, which is extracted from an epileptic patient. The ictal state is the signal fragment where an epileptic seizure occurs; the pre-ictal state is the segment that happens before the onset of a seizure; post-ictal is the state that follows a seizure, and inter-ictal is the fragment between the end of post-ictal and the start of pre-ictal. In this work, we merge post-ictal and inter-ictal states into one state, inter-ictal. Furthermore, we discard the ictal state since detecting an epileptic seizure, when it is happening, is of no interest for seizure prediction. Thus, we only consider two states, inter-ictal and pre-ictal. Epileptic seizure prediction is achieved when the system detects the pre-ictal state [1].Real-time epileptic seizure prediction mainly consists of two stages: (1) extraction of a tensor from the signal of interest, at any time position; and (2) classification of the tensor as inter-ictal or pre-ictal. Epileptic seizures are predicted when pre-ictal segments are correctly detected. Detection of the ictal state is of no interest since the goal is detecting a seizure before it happens by identifying the occurrence of the pre-ictal state. Thus, ictal fragments are of no use. There have been different efforts to predict seizures and most of them are based on the analysis of electroencephalography (EEG) signals [2]; however, recent publications have suggested that fNIRS signals, a relatively new modality, could be used to predict seizures [3,4]. fNIRS (functional near-infrared spectroscopy) is an optical technique, which conveys monitoring information about brain activity [5]; specifically, two physiological parameters or measurements, the relative level of oxygenated hemoglobin (HbO) and the relative level of deoxygenated hemoglobin (HbR). fNIRS signals are generated by injecting infrared lights of different wavelengths (650 nm–1000 nm) into the scalp, followed by registration of reflected light through optodes. Blood hemoglobin has different levels of light absorption and reflection depending on the level of blood oxygen due to brain activity [5,6]. fNIRS recordings arise from multiple channels. At each channel, two values are registered, HbO and HbR.fNIRS is a non-invasive neuroimaging technique that is effective for the monitoring of cortical hemodynamic changes. Studies have shown that fNIRS can be used to assess cortical hemodynamic changes associated with seizures, helping in the detection of seizures and the assessment of their impact on brain oxygenation [4]. Optical sensors (such as ratiometric nano quantum dot fluorescence resonance energy transfer sensors) were used to quantitatively measure oxygen dynamics from single-cell microdomains during oxygen-deprivation episodes (hypoxic episodes) and during induced seizure-like events in rats. A substantial decrease in the oxygenation of pyramidal cells was found through the concomitant measurement of blood oxygenation and electrical neural activity. This decrease in oxygen levels took place up to a few seconds before ictal activity could be revealed in electrophysiological recordings. Hence, an internally generated deficiency in oxygen is a crucial and decisive factor for hemodynamics related to epileptiform patterns [7]. The use of fNIRS in seizure prediction is based on scientific evidence that BOLD (Blood oxygen-level dependent) activity may result from neuronal changes occurring several seconds before a surface EEG event [8], and that vasoconstriction of blood vessels occurs a few seconds before the electrical onset [9]. A suitable format for fNIRS data consists of a two- or three-dimensional grid, which is the result of recording signals from multiple channels (rows), at multiple time positions (columns), by using two different measured values (feature maps or planes), HbO and HbR. Motivations for applying Convolutional Neural Networks (CNNs) to the problem of predicting epileptic seizures are (1) a large number of fNIRS channels, (2) the three-dimensional grid topology of fNIRS data, and (3) the straight application of raw fNIRS data to a CNN, where appropriate features are progressively extracted at convolutional layers for further classification. The use of fNIRS recordings for automated epileptic seizure detection requires more explorative studies. There are not public datasets related to fNIRS recordings on epileptic patients. This work is one of the firsts where fNIRS and deep learning are used to tackle the problem of seizure prediction. Detection of epileptic seizures in advance would allow epileptic patients or their caretakers to take precautions before the occurrence of an epileptic seizure, therefore mitigating risks and potential harm associated with the event, for instance in the intensive care unit, where electroencephalography (EEG) is typically not monitored. This document is organized as follows: Section 2 presents previous work. Section 3 provides the Convolutional Neural Network framework for the classification of fNIRS tensors, which includes a description of the architecture parameters and the implementation of the learning algorithm. Section 4 gives details of the methodology followed in this research. The results of this work are presented in Section 5 and discussed in Section 6. Finally, conclusions are given in Section 7.
Section snippets
Related work
Predicting epileptic seizures is the core of monitoring and warning systems so that precautions are taken before a seizure occurs [2]. One of the main problems for patients with epilepsy is the unpredictability of seizures [10]. Seizure detection and prediction have been mostly based on EEG recordings [1,2,[11], [12], [13], [14], [15]] given its capability to monitor electrical activity from the brain at a high temporal resolution [16]. EEG is the fastest and cheapest non-invasive method for
Implementation of a convolutional neural network
This section provides the framework for the complete implementation of a convolutional neural network (CNN) along with the corresponding learning algorithm. This section introduces the notation that is used to describe the selection of the most appropriate CNN model for seizure prediction. This notation is later used in the Results section, which shows the outcome of a set of experiments aimed at selecting the CNN model parameter values.
Dataset
There are not public datasets with fNIRS recordings from patients with epileptic seizures. The dataset for this work was the result of recording fNIRS signals from patients with focal refractory epilepsy. These recordings were performed and analyzed by epileptologists from Hôpital Notre-Dame du Centre Hospitalier de l'Université de Montréal. Forty-nine patients were subjects for the generation of the dataset; however, only five patients presented epileptic seizures while recording their brain
Results
Fig. 4 shows the ictal, inter-ictal and pre-ictal states within the recordings of (1) a single channel of EEG (specified in micro-volts [μV]); (2) three states within HbO and HbR recordings (specified in changes in hemoglobin concentration [C]). Finally, parts B), C), D), E) and F) show the specific placement of hundreds of optodes on five different patients, where blue lines represent sources and red lines represent detectors at different optode positions.
For each patient, experiments were
Discussion
According to Table 2, epileptic seizure prediction, based on the application of CNN to fNIRS recordings, is effective since performance metrics reach values above 95%. The performance of seizure prediction, when HbO and HbR are jointly used, is higher than that obtained by using only single reading, HbO or HbR. The HbO-HbR (fNIRS) integration allows higher prediction rates since this combination allows convolutional layers to extract more relevant features, which provides spatial and temporal
Conclusions
The problem of predicting epileptic seizures, using deep learning and fNIRS recordings, was studied. It is shown that the application of CNN to fNIRS data is suitable given the nature of fNIRS, which can be modeled as two- or three-dimensional tensors of high dimensionality. Evidence of the effectiveness of the proposed method is obtained through assessment of different performance metrics such as accuracy ranging between 96.9% and 100%, sensitivity between 95.24% and 100%, and specificity
Acknowledgement
This work was supported in part by a Mexico-Quebec Mobility program (project No. 265578). Data recordings were obtained from Hôpital Notre-Dame du Centre Hospitalier de l'Université de Montréal. Edgar Guevara acknowledges support from “Cátedras CONACYT” project 528.
References (42)
- et al.
BOLD changes occur prior to epileptic spikes seen on scalp EEG
Neuroimage
(2007) - et al.
Advantages and limitations of brain imaging methods in the research of absence epilepsy in humans and animal models
J. Neurosci. Methods
(2013) - et al.
Predicting epileptic seizures in advance
PLoS One
(2014) - et al.
Seizure detection, seizure prediction, and closed-loop warning systems in epilepsy
Epilepsy Behav.
(2014) - et al.
Toward new paradigms of seizure detection
Epilepsy Behav.
(2013) - et al.
Epileptic seizure detection in fNIRS signals using a supervised classifier
Proc. fNIRS
(2014) NIRS in clinical neurology a promising tool?
Neuroimage
(2014)- et al.
Nonlinear hemodynamic responses in human epilepsy: a multimodal analysis with fNIRS-EEG and fMRI-EEG
J. Neurosci. Methods
(2012) - et al.
Oxygen and seizure dynamics: I. experiments
J. Neurophysiol.
(2014) - et al.
Preictal and ictal neurovascular and metabolic coupling surrounding a seizure focus
J. Neurosci.
(2011)
Idiopathic generalized epilepsy with generalized tonic-clonic seizures (IGE-GTC): a population-based cohort with more than 20 year follow up for medical and social outcome
Epilepsy Behav.
Identification of epilepsy seizures using multi-resolution analysis and artificial neural networks
Recent Adv. Hybrid Approaches Des. Intell. Syst.
Review: a survey of performance and techniques for automatic epilepsy detection
J. Med. Biol. Eng.
Seizure prediction: the long and winding road
Brain J. Neurol.
Automated EEG analysis of epilepsy: a review
Knowl. Based Syst.
EEG-NIRS in epilepsy in children and neonates
Neurophysiol. Clin. Clin. Neurophysiol.
EEG in the diagnosis, classification, and management of patients with epilepsy
J. Neurol. Neurosurg. Psychiatry
Non-invasive pre-surgical investigation of a 10 year-old epileptic boy using simultaneous EEG-NIRS
Seizure J. Br. Epilepsy Assoc.
Using fiberless, wearable fNIRS to monitor brain activity in real-world cognitive tasks
J. Vis. Exp.
Cited by (44)
An exhaustive comparison of distance measures in the classification of time series with 1NN method
2024, Journal of Computational ScienceAssessment of flourishing levels of individuals by using resting-state fNIRS with different functional connectivity measures
2021, Biomedical Signal Processing and ControlCitation Excerpt :In contrast to the conventional approach using ΔHbO for FC analysis, we also used ΔHb due to previous evidence suggesting both hemoglobin concentrations should be analyzed for sparse networks [58]. Previous fNIRS-based ML studies focusing on classification of various psychiatric disorders, also use ΔHb in addition to ΔHbO [29,30,59–64]. In some cases, ΔHb based features may also show higher accuracy results compared to ΔHbO based features [59].
A systematic review on hybrid EEG/fNIRS in brain-computer interface
2021, Biomedical Signal Processing and ControlCitation Excerpt :Multimodal EEG-fNIRS data can provide superior performance metrics, and different classification procedures have been applied to combined EEG-fNIRS BCI [94,110–112,114–121,123–126,133–143]. Specifically, by employing support vector machine (SVM) [110–115] or linear discriminant analysis (LDA) [94,116–120] (which are as the classic machine learning methods), the BCI signals have been classified by EEG power spectral densities and fNIRS signal amplitudes. Furthermore, linear discriminant analysis ensemble classifiers have been employed in [127], to increase the bitrate as well as the classification accuracy for fNIRS-BCI datasets.