- Email: [email protected]
Detection of generalized tonic-clonic seizures from ear-EEG based on EMG analysis
Seizure 59 (2018) 54–59
Contents lists available at ScienceDirect
Seizure journal homepage: www.elsevier.com/locate/yseiz
Detection of generalized tonic-clonic seizures from ear-EEG based on EMG analysis I.C. Zibrandtsena,b,* , P. Kidmosec, T.W. Kjaera,b,d a
Department of Neurophysiology, Zealand University Hospital, Roskilde, Denmark Department of Clinical Medicine, University of Copenhagen, Denmark Department of Engineering, Aarhus University, Aarhus, Denmark d Department of Neuroscience, University of Copenhagen, Denmark b c
A R T I C L E I N F O
A B S T R A C T
Article history: Received 12 February 2018 Received in revised form 30 April 2018 Accepted 3 May 2018
Purpose: Sudden unexpected death in epilepsy (SUDEP) is associated with generalized tonic-clonic seizures (GTCS) with most deaths occurring during sleep. Seizure detection devices have been suggested as a SUDEP prevention strategy. EMG-based GTCS detection can take advantage of the GTCS characteristic of sustained high-amplitude, high-frequency activity in the time-domain. Method: We present a GTCS-detection method based on median-filtered variance estimates on surface EMG measurements and describe its performance in a small exploratory proof-of-concept setting involving a group of 15 patients with 3 GTCS recorded with ear-EEG and another group of 6 patients with 11 GTCS recorded with scalp-EEG. Results: GTCS intervals were detected within 4.2-12.9 s of onset with 100% sensitivity (CI 29.2–100%) without false positives in 820.7 h of ear-EEG. The same detection method worked for the 11 GTCS from scalp EEG data with 100% sensitivity (CI 71.5–100%) and no false positives. Conclusions: Ear-EEG contains enough GTCS-specific EMG activity for GTCS detection to be feasible. EarEEG could be considered for nocturnal GTCS monitoring as a supplement to SUDEP preventive interventions. © 2018 British Epilepsy Association. Published by Elsevier Ltd. All rights reserved.
Keywords: Ear-EEG Neurophysiological monitoring GTCS SUDEP prevention Seizure detection Epilepsy
1. Introduction The incidence of sudden unexpected death in epilepsy (SUDEP) varies by type of epilepsy from 0.9–2.3 per 1000 person-years in the general epilepsy population to 1.1-5.9 per 1000 person-years in patients with chronic refractory epilepsy. Generalized tonic-clonic seizures (GTCS) is the most important risk factor with higher seizure frequency linked to greater risk of SUDEP [1]. Compared to no yearly GTCS, a seizure frequency of 1–2/year is associated with an odds-ratio (OR) of SUDEP of 5.1, increasing to 15.5 at three or more yearly GTCS [2]. Nocturnal GTCS is an independent risk factor for SUDEP (OR 2.6) and SUDEP occurs more often during the night and is mostly unwitnessed [3] and a majority are found lying dead in the prone position [4]. There is evidence that nocturnal supervision has a protective effect concerning SUDEP risk [5] and it has been suggested that medical devices for seizure detection can support SUDEP prevention [2].
* Corresponding author at: Neurophysiology Center, Department of Neurology, Zealand University Hospital, Sygehusvej 10, Roskilde, DK-4000, Denmark E-mail address: [email protected] (I.C. Zibrandtsen).
Seizure detection for GTCS using accelerometry is aimed at identifying the clonic phase and various studies have reported performance with sensitivities between 11 and 100% and false alarm rates ranging between 0.2 to 4/day [6]. Compared to accelerometry, surface EMG (sEMG) can detect the tonic phase of GTCS better and thus work with shorter detection latencies; median detection latency 17.0 s compared to 13.7s, respectively [7,8]. The sEMG detection strategy can rely on different features: Changes in median frequency, root mean square and EMG-EMG coherence can separate 63 GTCS from 20 patients from 100 simulated seizures acted by volunteers [9]. Later, an algorithm based on the number of zero-crossings achieved 100% sensitivity and 1 false positive/24 h when tested on 22 GTCS from 11 patients [8]. Beniczky et al. used a multivariate analysis of time-domain and frequency domain features to demonstrate detection of GTCS and differentiation from psychogenic convulsive seizures in recordings from a mobile device worn on the upper arm [10]. Ear-EEG a wearable EEG recording method, where electrodes are placed in and around the ear. Ear-EEG is of particular interest for long-term monitoring in epilepsy [11,12] and sleep [13,14]. EarEEG has also been used for hearing threshold estimation [15]. Physiological EMG-artifacts, particularly those elicited from jaw
https://doi.org/10.1016/j.seizure.2018.05.001 1059-1311/© 2018 British Epilepsy Association. Published by Elsevier Ltd. All rights reserved.
I.C. Zibrandtsen et al. / Seizure 59 (2018) 54–59
movements, are visible in ear-EEG measurements [16]. Thus, it seems plausible that the generalized muscle-activation pattern characteristic of GTCS could be detected using ear-EEG. The objective of this study was to investigate the feasibility of using ear-EEG for GTCS detection. 2. Methods 2.1. Measurements EEG was obtained from 15 patients admitted to the epilepsy monitoring unit (EMU) at Zealand’s University Hospital and participating in a study of ear-EEG for suspected temporal lobe seizures. Simultaneous 25 channel scalp EEG and 8-channel binaural ear-EEG of durations between 1 and 4 days were collected. Patients were admitted on a primary suspicion of temporal lobe epilepsy, but among the seizure episodes some were of GTCS type and we refer to those in the following. For this study, we are only interested in the surface EMG component of
55
the measurements. Intervals, where recording conditions were unsatisfactory because of high impedances or dislodged ear-EEG devices, were visually identified and removed. Sampling frequency for both scalp and ear-EEG was 256 Hz for five patients and 1024 Hz for the other ten patients including all those with GTCSs. For the same analysis to be performed on all data sets, we resampled the five datasets sampled at 256 Hz to 1024 Hz. Fig. 1 (panel 1 and 2) shows the ear-EEG device, how it is placed into the ear and the electrode locations. The electrodes were labelled E, I, B and A with the prefix L/R denoting left or right. References were bipolar and defined as either “intra-ear”, when created from the six possible combinations of ipsilateral ear-electrodes, or “interear”, when the reference was the same-labelled electrode in the contralateral earpiece. Full description of the measurement and study population is described in our previous study [12]. Data were exported as edf-files and imported into Matlab R2017a for further analyses. The study was done in accordance with the Helsinki Declaration and was approved by the regional ethics committee (110724815).
Fig. 1. 1) Photograph of the ear-EEG prototype we used. Right ear-piece. Labels A, B, E and I denote the electrodes and arrows point to their location on the earpiece. Note that there is an air duct for normal sound conductance. 2) Sketch of how the device is placed in the ear and the electrode locations. 3) Seizure onset for one of the GTCS we recorded, 8 longitudinal scalp channels displayed first. Below are two left intra-ear channels (LE-LB and LI-LA) displayed in bright red. Homologue right sided ear channels in bright blue. Inter-ear channels LE-RE and LB-RB are shown in green. Notice the lower amplitudes for the intra-ear channels. The distinct seizure pattern can better be appreciated in the inter-ear channels.
56
I.C. Zibrandtsen et al. / Seizure 59 (2018) 54–59
2.2. Median-filtered variance feature The EMG pattern of GTCS is characterized by high-amplitude, high frequency oscillations that are distinct from physiological EMG activity even during attempts to simulate convulsive seizures [9]. Variance calculated within non-overlapping sliding windows across the time series will result in a consistent increase over consecutive intervals whereas normal physiological movements will produce more irregularly distributed high variance values. Let V denote the variance calculated within non-overlapping sliding windows consisting of a number w of consecutive samples within one EEG channel. V¼
i¼1 1X ðx xÞ2 w w i
Where xi denotes the individual samples within each instance of w and x is the mean of samples within each window. Fig. 2 illustrates the principle of a sliding window moving across the different phases of a GTCS. Pre-seizure we would expect most windows to contain low V values, but in the tonic phase most windows would have high V values and then somewhat fewer during the clonic phase before returning to generally low V values post-seizure. We tried different values of w and visually inspected the results, when w becomes large there is a decrease in the maximal amplitude across the entire recording, which is undesirable when we wish to base detection upon crossing an amplitude threshold. We found that w = 32 spanned enough samples for robust estimation before unnecessary amplitude attenuation occurred at a higher w parameter. The salient characteristic of the tonic phase is the sustained increase of V, but due to random variation some windows during the tonic phase will still contain relatively low V. More importantly, many physiological actions will also cause large amplitude increases, but these will not have sustained high variance values to the same extent. Taking the median across a
number k of consecutive V segments will filter out irregular or nonsustained elevations in V. We call the sequence of V across the time series recorded from a given electrode for Vs and the result of the ~ S (Fig. 3). For online detection we median-filtered sequence for V can only consider previous indices in Vs. Choosing a higher k will produce a more robust attenuation of non-sustained elevated V, but at a cost of a longer detection delay given by k w sr1, where ~ S is mV2 and we sr denotes the sampling rate. The unit of V normalize it by dividing it with its own standard deviation ~ S for a given channel. calculated across all previous samples of V At random intervals a number of ear-EEG electrodes will become noise-contaminated by movements or loosened electrode contact and this will typically generate high-amplitude, high frequency noise, which could be misinterpreted as GTCS-related EMG signal. From a probabilistic point of view, it is unlikely that a majority of ear-EEG channels randomly will be performing poorly at a given time, but when a GTCS occurs, we should expect it to be measurable from all ear-EEG positions at the same time. It can be assumed that a GTCS event will cause bilateral high-amplitude activity, so if only unilateral noise is detected, it can be ignored as potential seizure activity. Relying on this assumption, we select the ~ S for all left and all right intra-ear channels. smallest sample of V This quantity is then compared to the most common observation of the four inter-ear channels. We defined a threshold from examining the probability distributions of seizure-interval values compared to the non-seizure intervals (Fig. 4a). When the mode of the inter-ear channels is above the threshold, a check is performed against the left and right intra-ear channels, so that if it is above the threshold a decision of GTCS event is reached but if it is below the threshold the event is disregarded. ~ S during GTCS and it Fig. 3c shows the characteristic shape of V can be seen, that if the threshold is set higher the detection delay will increase but make separation from non-seizure easier. We used a k parameter of 25 and tested different normalized threshold values from 20 to 99.
Fig. 2. Concept of how different phases of generalized tonic-clonic seizure are summarized by variance calculated windows spanning consecutive samples (orange boxes indicate the size of each window). Pre-seizure and post-seizure most windows will contain low values. During the tonic phase somewhat more windows will contain high values compared to the subsequent clonic phase. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
I.C. Zibrandtsen et al. / Seizure 59 (2018) 54–59
57
Fig. 3. A) Generalized tonic-clonic seizure in the time-domain. Yellow triangle marks seizure onset. About 30 s before there is a brief high-amplitude artifact. B) The same interval transformed into variance values within non-overlapping windows. C) The effect of median-filtering attenuates the artifact near 30 s but only smoothes the seizure activity. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Fig. 4. A) Histograms of the cumulative probability distributions of the normalized amplitudes during seizure intervals defined as seizure onset and 60 s forward (orange) and non-seizure intervals (blue). Plot window is cropped at y 0.05 and x 10 for better visibility. The distribution of non-seizure intervals has virtually no overlap at 10 and beyond. B) Detection latencies for the 3 generalized tonic-clonic seizures as a function of different threshold values. At 20 the detection latency is 4.2s, 12.2 s and 12.9s. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
2.3. Scalp-EEG We investigated a similar GTCS detection approach but on scalp-EEG from a collection of scalp-EEGs containing 11 GTCS examples from 6 other patients not monitored with ear-EEG. Sampling rate was 1024 Hz with 25 scalp electrodes including lowrow in the 10:20 system. We chose two longitudinal channels that are relatively close to ear-EEG, T9/10-P9/10 and, following the
same assumption that GTCS should produce a clear signal ~ S for both scalp channels at bilaterally, we chose the minimum V the same time. 2.4. Statistics Detections occurring up to 1 min after the seizure annotation were considered true positives. The number of true positives divided by the total seizure count is the sensitivity. R version 3.4.3
58
I.C. Zibrandtsen et al. / Seizure 59 (2018) 54–59
function binom.test for the exact test for binomial proportions was used to calculate confidence intervals for the sensitivities. ~ S following a detection would be above Subsequent epochs of V the threshold, which would produce multiple detections within the same seizure interval. To prevent this, we implemented a 10 min pause after a detection, where no further above threshold epochs would register as detections. Any detections not within 1 min of seizure annotation are false positives. False positives are described as a rate defined by the number of false positives divided by the total monitoring time in hours. 3. Results The amount of data removed after visual assessment of measurement quality was 60.8 h in total with most of it (52.7 h) originating from one patient, where electrode care had been inadequate. This patient didn’t have any GTCSs. The rest of the data, which was removed, was spread among patients at the end of the recordings, where ear-EEG was frequently disconnected before scalp-EEG was and the recording not yet stopped. In total, 820.7 h of simultaneous scalp-EEG and ear-EEG remained. There were three GTCS from three different patients. A sample seizure is shown in Fig. 1c. In one patient, the GTCS was primarily generalized. Two patients had a focal onset impaired awareness seizure to bilateral tonic-clonic seizure. All three GTCSs were independently detected based on ear-EEG and scalp-EEG, sensitivity 100% (CI 29.2–100%). The false positive rate across all 820.7 h was 0. Fig. 4b shows the detection latencies for different threshold values for the three seizures. A higher threshold decreases the risk of false positives but at a cost of increased latency and a higher possibility of false negatives. At a threshold of 20 there were no false positives and detection latency between 4.2–12.9 s. Lower threshold settings were not tested. Instances of physiological artifacts from chewing, ambulation, movement and turning were numerous but did not produce false detections. An additional 24 h of scalp-EEG containing 11 additional GTCS from 6 patients was also tested and all 11 GTCS were detected with no false positives, sensitivity 100% (71.5-100%). If only considering one scalp channel instead of two, 10/11 GTCS were detected. The 3 GTCS in the ear-EEG data could also be detected from the scalp electrode positions without false positives. Detection latencies for scalp-EEG based detection were not calculated. 4. Discussion The detection strategy presented above worked well but for earEEG it has only been tested on three GTCS events. The parameters of the algorithm, i.e. length of the window for the variance estimate, number of windows for the median filter, and the detection threshold, were based on the same data material, so it could be subject to overfitting. Thus, the generalizability of the performance across the range of possible GTCS-variability can be questioned and rightly warrants more investigation. However, the large amount of normal and varied physiological activity tested produced no false positives, which is an important metric for seizure detection systems; a high false positive rate will likely cause “alarm fatigue” and dissatisfaction with the user. Recordings were all obtained from an EMU-environment during day and night, and although the range of physical activity is restricted, a wide variety of physiological EMG occurred. This seems especially pertinent regarding facial movements since the jaw muscles engaged during GTCS likely also would be active during chewing, but we can see that chewing artifacts do not generate false positives during any mealtime intervals contained in the dataset. Additionally, the need for a GTCS alarm is principally during
unsupervised sleep, where the possibility of physiological activity that would generate a false alarm is even smaller. However, it could be argued that the fact that there were no false positives was mostly a fortuitous consequence of the low number of GTCS events the threshold was based on. Perhaps other individuals would have significantly lower amplitudes during the tonic interval so that it would become apparent that perfect classification would no longer be possible across a wider subset of GTCS? EMG patterns from GTCSs are arguably very stereotypical and so it seems unlikely that GTCS EMG would differs substantially, but it surely warrants additional investigation. The successful detection of an additional 11 GTCSs recorded with only scalp-EEG seems to suggest so. If other GTCS have markedly lower amplitudes so that separation from non-GTCS becomes imperfect by thresholding, a strategy to amend this could be based on additions to capture the sustained increase in amplitude over a longer period, possibly accepting a “cost” in the form of a longer detection latency. The MORTEMUS study points to a specific sequence of events leading from a GTCS to neurovegetative breakdown into respiratory depression and then cardiac arrest in the postictal interval several minutes after seizure onset [17]. In animal models of SUDEP, hypoventilation, not cardiac arrhythmia, was also the stronger predictor of death [18]. However, analysis of heart variability has identified pre-ictal parasympathetic overdrive and short QT as risk factor for SUDEP in a case where a patient died despite cardiopulmonary resuscitation [19]. SUDEP is theoretically preventable through interventions to support respiratory effort. A study of the effects of early peri-ictal nursing interventions on hypoxemia suggests that earlier intervention can shorten hypoxemia [20]. Our results indicate that GTCS detection strategy can work with a detection delay of about 4–13 s, which is comparable to other similar methods; two recent multi-center trials of GTCS detection by surface EMG obtained median detection latencies of 7.4 and 9 s [21,22]. These differences are small and roughly equivalent in terms of clinical relevance. If detection is based on movements during the clonic phase, which is common for accelerometry, the delays tend to be longer and then it might become more relevant if the purpose is to alarm care-givers for earlier assistance. One might suppose that our detection strategy could generalize to other motor seizures such as tonic seizures, but comparisons between tonic and generalized tonic-clonic seizures have shown that purely tonic seizures have a markedly lower amplitude [10], so this would again require different threshold parameterization. However, SUDEP mortality is associated to GTCS, not pure tonic seizures [17]. The detection strategy works on scalp-EEG with a fourelectrode montage but would likely also work for other electrode positions on the body. Other types of sensors located elsewhere on the body have also been investigated for GTCS detection. Spectral analysis of accelerometry (ACM) from wrist-worn sensor was used to detect 10/10 GTCS from 5/12 patients monitored for 246 h with a mean false positive ratio of 2/24 h [23]. However, the location of the ear-EEG device has the advantage that it also samples EEG. During sleep, this could allow for on demand sleep staging, which has recently been shown to work on automated basis with good accuracy compared to expert rated PSG [14]. This combined capability offers a “bundled solution” to multiple clinical problems. For instance, this could have relevance for the differentiation between abnormal movements during sleep, where knowledge about the sleep stage from which abnormal movement arise can inform a diagnostic process on REM and non-REM parasomnias. Bilateral use of ear-EEG is an advantage when the clinical problem is GTCS detection as accuracy can be improved by assuming that GTCS-associated activity should be reflected in all electrode
I.C. Zibrandtsen et al. / Seizure 59 (2018) 54–59
positions whereas random noise most likely will not. Other GTCS detection proposals utilizing only a single sensor location cannot take advantage of this. In our previous study [12] we described some irritation linked to prolonged use of hard ear-EEG pieces. This is unlikely to be an issue if the devices are only put in use during sleep. In addition, a new ear-EEG design recently described [24] uses dry-electrodes and a softer material and a flexible joint between the portion resting in the concha and the portion inside the ear-canal, which may turn out the decrease or eliminate the user-discomfort we described previously. 5. Conclusion We have described a simple algorithmic for GTCS detection based on a moving-median filtered sliding window variance estimate. The performance was tested on 820.7 h of simultaneous scalp-EEG and ear-EEG containing 3 GTCS from 3 different patients with detection latencies between 4 and 13 s. Sensitivity for ear-EEG was 100% (CI 29.2–100%) and no false positives. An additional 11 GTCS events from 6 different patients measured with only scalpEEG were also tested detecting all seizures with no false positives. Parameterization was developed and tuned post-hoc and the number of true events to be detected was low. The calculated detection latencies were based on an online-simulation run on a computer, not a wearable setup as is the intended future goal. Still, the results indicate that ear-EEG measurements during GTCS contain sufficient information for a seizure detection system to be feasible. Compared to GTCS detection by surface EMG collected from a sensor placed on the extremities, ear-EEG is of special interest in cases where EEG evaluation is relevant at the same time. Declarations of interest None. Acknowledgements The project was supported by the Danish Council for Strategic Research (1311-00009B) and the Elsass Foundation. References [1] Shorvon S., Tomson T. Sudden unexpected death in epilepsy. Lancet Lond Engl 2011;378(December (9808)):2028–38. [2] Tomson T, Surges R, Delamont R, Haywood S, Hesdorffer DC. Who to target in sudden unexpected death in epilepsy prevention and how? Risk factors, biomarkers, and intervention study designs. Epilepsia 2016;57(Suppl. 1):4–16. [3] Lamberts RJ, Thijs RD, Laffan A, Langan Y, Sander JW. Sudden unexpected death in epilepsy: people with nocturnal seizures may be at highest risk. Epilepsia 2012;53(February (2)):253–7.
59
[4] Kloster R, Engelskjon T. Sudden unexpected death in epilepsy (SUDEP): a clinical perspective and a search for risk factors. J Neurol Neurosurg Psychiatry 1999;67(October (4)):439–44. [5] Langan Y, Nashef L, Sander JW. Case-control study of SUDEP. Neurology 2005;64(April (7)):1131–3. [6] van Andel J, Thijs RD, de Weerd A, Arends J, Leijten F. Non-EEG based ambulatory seizure detection designed for home use: What is available and how will it influence epilepsy care? Epilepsy Behav EB 2016;57(Pt A):82–9. [7] Kramer U, Kipervasser S, Shlitner A, Kuzniecky R. A novel portable seizure detection alarm system: preliminary results. J Clin Neurophysiol Off Publ Am Electroencephalogr Soc 2011;28(February (1)):36–8. [8] Conradsen I, Beniczky S, Hoppe K, Wolf P, Sorensen HBD. Automated algorithm for generalized tonic-clonic epileptic seizure onset detection based on sEMG zero-crossing rate. IEEE Trans Biomed Eng 2012;59(February (2)):579–85. [9] Conradsen I, Wolf P, Sams T, Sorensen HBD, Beniczky S. Patterns of muscle activation during generalized tonic and tonic-clonic epileptic seizures. Epilepsia 2011;52(November (11)):2125–32. [10] Beniczky S, Conradsen I, Pressler R, Wolf P. Quantitative analysis of surface electromyography: biomarkers for convulsive seizures. Clin Neurophysiol. 2016;127(August (8)):2900–7. [11] Looney D, Kidmose P, Park C, Ungstrup M, Rank M, Rosenkranz K, et al. The inthe-ear recording concept: user-centered and wearable brain monitoring. IEEE Pulse 2012;3(December (6)):32–42. [12] Zibrandtsen IC, Kidmose P, Christensen CB, Kjaer TW. Ear-EEG detects ictal and interictal abnormalities in focal and generalized epilepsy — A comparison with scalp EEG monitoring. Clin Neurophysiol Off J Int Fed Clin Neurophysiol 2017;128(October (12)):2454–61. [13] Zibrandtsen I, Kidmose P, Otto M, Ibsen J, Kjaer TW. Case comparison of sleep features from ear-EEG and scalp-EEG. Sleep Sci 2016;9(April (2)):69–72. [14] Mikkelsen KB, Villadsen DB, Otto M, Kidmose P. Automatic sleep staging using ear-EEG. Biomed Eng Online 2017;16(September (1)):111. [15] Bech Christensen C, Harte JM, Lunner T, Kidmose P. Ear-EEG based objective hearing threshold estimation evaluated on normal hearing subjects. IEEE Trans Biomed Eng 2017(August). [16] Kappel SL, Looney D, Mandic DP, Kidmose P. Physiological artifacts in scalp EEG and ear-EEG. Biomed Eng Online [Internet] 201716(August). . Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5553928/. [17] Ryvlin P, Nashef L, Lhatoo SD, Bateman LM, Bird J, Bleasel A, et al. Incidence and mechanisms of cardiorespiratory arrests in epilepsy monitoring units (MORTEMUS): a retrospective study. Lancet Neurol 2013;12(October (10)):966–77. [18] Johnston SC, Horn JK, Valente J, Simon RP. The role of hypoventilation in a sheep model of epileptic sudden death. Ann Neurol 1995;37(April (4)):531–7. [19] Jeppesen J, Fuglsang-Frederiksen A, Brugada R, Pedersen B, Rubboli G, Johansen P, et al. Heart rate variability analysis indicates preictal parasympathetic overdrive preceding seizure-induced cardiac dysrhythmias leading to sudden unexpected death in a patient with epilepsy. Epilepsia 2014;55(July (7)):e67–71. [20] Seyal M, Bateman LM, Li C-S. Impact of periictal interventions on respiratory dysfunction, postictal EEG suppression, and postictal immobility. Epilepsia 2013;54(February (2)):377–82. [21] Halford JJ, Sperling MR, Nair DR, Dlugos DJ, Tatum WO, Harvey J, et al. Detection of generalized tonic-clonic seizures using surface electromyographic monitoring. Epilepsia 2017;58(November (11)):1861–9. [22] Beniczky S, Conradsen I, Henning O, Fabricius M, Wolf P. Automated real-time detection of tonic-clonic seizures using a wearable EMG device. Neurology 2018;90(January (5)):e428–34. [23] Joo HS, Han S-H, Lee J, Jang DP, Kang JK, Woo J. Spectral analysis of acceleration data for detection of generalized tonic-clonic seizures. Sensors [Internet] 201717(February (3)). . Available from: https://www.ncbi.nlm.nih.gov/pmc/ articles/PMC5375767/. [24] Kappel SL, Rank ML, Toft HO, Andersen M, Kidmose P. Dry-contact Electrode Ear-EEG IEEE TBME. 2018.
Contents lists available at ScienceDirect
Seizure journal homepage: www.elsevier.com/locate/yseiz
Detection of generalized tonic-clonic seizures from ear-EEG based on EMG analysis I.C. Zibrandtsena,b,* , P. Kidmosec, T.W. Kjaera,b,d a
Department of Neurophysiology, Zealand University Hospital, Roskilde, Denmark Department of Clinical Medicine, University of Copenhagen, Denmark Department of Engineering, Aarhus University, Aarhus, Denmark d Department of Neuroscience, University of Copenhagen, Denmark b c
A R T I C L E I N F O
A B S T R A C T
Article history: Received 12 February 2018 Received in revised form 30 April 2018 Accepted 3 May 2018
Purpose: Sudden unexpected death in epilepsy (SUDEP) is associated with generalized tonic-clonic seizures (GTCS) with most deaths occurring during sleep. Seizure detection devices have been suggested as a SUDEP prevention strategy. EMG-based GTCS detection can take advantage of the GTCS characteristic of sustained high-amplitude, high-frequency activity in the time-domain. Method: We present a GTCS-detection method based on median-filtered variance estimates on surface EMG measurements and describe its performance in a small exploratory proof-of-concept setting involving a group of 15 patients with 3 GTCS recorded with ear-EEG and another group of 6 patients with 11 GTCS recorded with scalp-EEG. Results: GTCS intervals were detected within 4.2-12.9 s of onset with 100% sensitivity (CI 29.2–100%) without false positives in 820.7 h of ear-EEG. The same detection method worked for the 11 GTCS from scalp EEG data with 100% sensitivity (CI 71.5–100%) and no false positives. Conclusions: Ear-EEG contains enough GTCS-specific EMG activity for GTCS detection to be feasible. EarEEG could be considered for nocturnal GTCS monitoring as a supplement to SUDEP preventive interventions. © 2018 British Epilepsy Association. Published by Elsevier Ltd. All rights reserved.
Keywords: Ear-EEG Neurophysiological monitoring GTCS SUDEP prevention Seizure detection Epilepsy
1. Introduction The incidence of sudden unexpected death in epilepsy (SUDEP) varies by type of epilepsy from 0.9–2.3 per 1000 person-years in the general epilepsy population to 1.1-5.9 per 1000 person-years in patients with chronic refractory epilepsy. Generalized tonic-clonic seizures (GTCS) is the most important risk factor with higher seizure frequency linked to greater risk of SUDEP [1]. Compared to no yearly GTCS, a seizure frequency of 1–2/year is associated with an odds-ratio (OR) of SUDEP of 5.1, increasing to 15.5 at three or more yearly GTCS [2]. Nocturnal GTCS is an independent risk factor for SUDEP (OR 2.6) and SUDEP occurs more often during the night and is mostly unwitnessed [3] and a majority are found lying dead in the prone position [4]. There is evidence that nocturnal supervision has a protective effect concerning SUDEP risk [5] and it has been suggested that medical devices for seizure detection can support SUDEP prevention [2].
* Corresponding author at: Neurophysiology Center, Department of Neurology, Zealand University Hospital, Sygehusvej 10, Roskilde, DK-4000, Denmark E-mail address: [email protected] (I.C. Zibrandtsen).
Seizure detection for GTCS using accelerometry is aimed at identifying the clonic phase and various studies have reported performance with sensitivities between 11 and 100% and false alarm rates ranging between 0.2 to 4/day [6]. Compared to accelerometry, surface EMG (sEMG) can detect the tonic phase of GTCS better and thus work with shorter detection latencies; median detection latency 17.0 s compared to 13.7s, respectively [7,8]. The sEMG detection strategy can rely on different features: Changes in median frequency, root mean square and EMG-EMG coherence can separate 63 GTCS from 20 patients from 100 simulated seizures acted by volunteers [9]. Later, an algorithm based on the number of zero-crossings achieved 100% sensitivity and 1 false positive/24 h when tested on 22 GTCS from 11 patients [8]. Beniczky et al. used a multivariate analysis of time-domain and frequency domain features to demonstrate detection of GTCS and differentiation from psychogenic convulsive seizures in recordings from a mobile device worn on the upper arm [10]. Ear-EEG a wearable EEG recording method, where electrodes are placed in and around the ear. Ear-EEG is of particular interest for long-term monitoring in epilepsy [11,12] and sleep [13,14]. EarEEG has also been used for hearing threshold estimation [15]. Physiological EMG-artifacts, particularly those elicited from jaw
https://doi.org/10.1016/j.seizure.2018.05.001 1059-1311/© 2018 British Epilepsy Association. Published by Elsevier Ltd. All rights reserved.
I.C. Zibrandtsen et al. / Seizure 59 (2018) 54–59
movements, are visible in ear-EEG measurements [16]. Thus, it seems plausible that the generalized muscle-activation pattern characteristic of GTCS could be detected using ear-EEG. The objective of this study was to investigate the feasibility of using ear-EEG for GTCS detection. 2. Methods 2.1. Measurements EEG was obtained from 15 patients admitted to the epilepsy monitoring unit (EMU) at Zealand’s University Hospital and participating in a study of ear-EEG for suspected temporal lobe seizures. Simultaneous 25 channel scalp EEG and 8-channel binaural ear-EEG of durations between 1 and 4 days were collected. Patients were admitted on a primary suspicion of temporal lobe epilepsy, but among the seizure episodes some were of GTCS type and we refer to those in the following. For this study, we are only interested in the surface EMG component of
55
the measurements. Intervals, where recording conditions were unsatisfactory because of high impedances or dislodged ear-EEG devices, were visually identified and removed. Sampling frequency for both scalp and ear-EEG was 256 Hz for five patients and 1024 Hz for the other ten patients including all those with GTCSs. For the same analysis to be performed on all data sets, we resampled the five datasets sampled at 256 Hz to 1024 Hz. Fig. 1 (panel 1 and 2) shows the ear-EEG device, how it is placed into the ear and the electrode locations. The electrodes were labelled E, I, B and A with the prefix L/R denoting left or right. References were bipolar and defined as either “intra-ear”, when created from the six possible combinations of ipsilateral ear-electrodes, or “interear”, when the reference was the same-labelled electrode in the contralateral earpiece. Full description of the measurement and study population is described in our previous study [12]. Data were exported as edf-files and imported into Matlab R2017a for further analyses. The study was done in accordance with the Helsinki Declaration and was approved by the regional ethics committee (110724815).
Fig. 1. 1) Photograph of the ear-EEG prototype we used. Right ear-piece. Labels A, B, E and I denote the electrodes and arrows point to their location on the earpiece. Note that there is an air duct for normal sound conductance. 2) Sketch of how the device is placed in the ear and the electrode locations. 3) Seizure onset for one of the GTCS we recorded, 8 longitudinal scalp channels displayed first. Below are two left intra-ear channels (LE-LB and LI-LA) displayed in bright red. Homologue right sided ear channels in bright blue. Inter-ear channels LE-RE and LB-RB are shown in green. Notice the lower amplitudes for the intra-ear channels. The distinct seizure pattern can better be appreciated in the inter-ear channels.
56
I.C. Zibrandtsen et al. / Seizure 59 (2018) 54–59
2.2. Median-filtered variance feature The EMG pattern of GTCS is characterized by high-amplitude, high frequency oscillations that are distinct from physiological EMG activity even during attempts to simulate convulsive seizures [9]. Variance calculated within non-overlapping sliding windows across the time series will result in a consistent increase over consecutive intervals whereas normal physiological movements will produce more irregularly distributed high variance values. Let V denote the variance calculated within non-overlapping sliding windows consisting of a number w of consecutive samples within one EEG channel. V¼
i¼1 1X ðx xÞ2 w w i
Where xi denotes the individual samples within each instance of w and x is the mean of samples within each window. Fig. 2 illustrates the principle of a sliding window moving across the different phases of a GTCS. Pre-seizure we would expect most windows to contain low V values, but in the tonic phase most windows would have high V values and then somewhat fewer during the clonic phase before returning to generally low V values post-seizure. We tried different values of w and visually inspected the results, when w becomes large there is a decrease in the maximal amplitude across the entire recording, which is undesirable when we wish to base detection upon crossing an amplitude threshold. We found that w = 32 spanned enough samples for robust estimation before unnecessary amplitude attenuation occurred at a higher w parameter. The salient characteristic of the tonic phase is the sustained increase of V, but due to random variation some windows during the tonic phase will still contain relatively low V. More importantly, many physiological actions will also cause large amplitude increases, but these will not have sustained high variance values to the same extent. Taking the median across a
number k of consecutive V segments will filter out irregular or nonsustained elevations in V. We call the sequence of V across the time series recorded from a given electrode for Vs and the result of the ~ S (Fig. 3). For online detection we median-filtered sequence for V can only consider previous indices in Vs. Choosing a higher k will produce a more robust attenuation of non-sustained elevated V, but at a cost of a longer detection delay given by k w sr1, where ~ S is mV2 and we sr denotes the sampling rate. The unit of V normalize it by dividing it with its own standard deviation ~ S for a given channel. calculated across all previous samples of V At random intervals a number of ear-EEG electrodes will become noise-contaminated by movements or loosened electrode contact and this will typically generate high-amplitude, high frequency noise, which could be misinterpreted as GTCS-related EMG signal. From a probabilistic point of view, it is unlikely that a majority of ear-EEG channels randomly will be performing poorly at a given time, but when a GTCS occurs, we should expect it to be measurable from all ear-EEG positions at the same time. It can be assumed that a GTCS event will cause bilateral high-amplitude activity, so if only unilateral noise is detected, it can be ignored as potential seizure activity. Relying on this assumption, we select the ~ S for all left and all right intra-ear channels. smallest sample of V This quantity is then compared to the most common observation of the four inter-ear channels. We defined a threshold from examining the probability distributions of seizure-interval values compared to the non-seizure intervals (Fig. 4a). When the mode of the inter-ear channels is above the threshold, a check is performed against the left and right intra-ear channels, so that if it is above the threshold a decision of GTCS event is reached but if it is below the threshold the event is disregarded. ~ S during GTCS and it Fig. 3c shows the characteristic shape of V can be seen, that if the threshold is set higher the detection delay will increase but make separation from non-seizure easier. We used a k parameter of 25 and tested different normalized threshold values from 20 to 99.
Fig. 2. Concept of how different phases of generalized tonic-clonic seizure are summarized by variance calculated windows spanning consecutive samples (orange boxes indicate the size of each window). Pre-seizure and post-seizure most windows will contain low values. During the tonic phase somewhat more windows will contain high values compared to the subsequent clonic phase. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
I.C. Zibrandtsen et al. / Seizure 59 (2018) 54–59
57
Fig. 3. A) Generalized tonic-clonic seizure in the time-domain. Yellow triangle marks seizure onset. About 30 s before there is a brief high-amplitude artifact. B) The same interval transformed into variance values within non-overlapping windows. C) The effect of median-filtering attenuates the artifact near 30 s but only smoothes the seizure activity. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Fig. 4. A) Histograms of the cumulative probability distributions of the normalized amplitudes during seizure intervals defined as seizure onset and 60 s forward (orange) and non-seizure intervals (blue). Plot window is cropped at y 0.05 and x 10 for better visibility. The distribution of non-seizure intervals has virtually no overlap at 10 and beyond. B) Detection latencies for the 3 generalized tonic-clonic seizures as a function of different threshold values. At 20 the detection latency is 4.2s, 12.2 s and 12.9s. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
2.3. Scalp-EEG We investigated a similar GTCS detection approach but on scalp-EEG from a collection of scalp-EEGs containing 11 GTCS examples from 6 other patients not monitored with ear-EEG. Sampling rate was 1024 Hz with 25 scalp electrodes including lowrow in the 10:20 system. We chose two longitudinal channels that are relatively close to ear-EEG, T9/10-P9/10 and, following the
same assumption that GTCS should produce a clear signal ~ S for both scalp channels at bilaterally, we chose the minimum V the same time. 2.4. Statistics Detections occurring up to 1 min after the seizure annotation were considered true positives. The number of true positives divided by the total seizure count is the sensitivity. R version 3.4.3
58
I.C. Zibrandtsen et al. / Seizure 59 (2018) 54–59
function binom.test for the exact test for binomial proportions was used to calculate confidence intervals for the sensitivities. ~ S following a detection would be above Subsequent epochs of V the threshold, which would produce multiple detections within the same seizure interval. To prevent this, we implemented a 10 min pause after a detection, where no further above threshold epochs would register as detections. Any detections not within 1 min of seizure annotation are false positives. False positives are described as a rate defined by the number of false positives divided by the total monitoring time in hours. 3. Results The amount of data removed after visual assessment of measurement quality was 60.8 h in total with most of it (52.7 h) originating from one patient, where electrode care had been inadequate. This patient didn’t have any GTCSs. The rest of the data, which was removed, was spread among patients at the end of the recordings, where ear-EEG was frequently disconnected before scalp-EEG was and the recording not yet stopped. In total, 820.7 h of simultaneous scalp-EEG and ear-EEG remained. There were three GTCS from three different patients. A sample seizure is shown in Fig. 1c. In one patient, the GTCS was primarily generalized. Two patients had a focal onset impaired awareness seizure to bilateral tonic-clonic seizure. All three GTCSs were independently detected based on ear-EEG and scalp-EEG, sensitivity 100% (CI 29.2–100%). The false positive rate across all 820.7 h was 0. Fig. 4b shows the detection latencies for different threshold values for the three seizures. A higher threshold decreases the risk of false positives but at a cost of increased latency and a higher possibility of false negatives. At a threshold of 20 there were no false positives and detection latency between 4.2–12.9 s. Lower threshold settings were not tested. Instances of physiological artifacts from chewing, ambulation, movement and turning were numerous but did not produce false detections. An additional 24 h of scalp-EEG containing 11 additional GTCS from 6 patients was also tested and all 11 GTCS were detected with no false positives, sensitivity 100% (71.5-100%). If only considering one scalp channel instead of two, 10/11 GTCS were detected. The 3 GTCS in the ear-EEG data could also be detected from the scalp electrode positions without false positives. Detection latencies for scalp-EEG based detection were not calculated. 4. Discussion The detection strategy presented above worked well but for earEEG it has only been tested on three GTCS events. The parameters of the algorithm, i.e. length of the window for the variance estimate, number of windows for the median filter, and the detection threshold, were based on the same data material, so it could be subject to overfitting. Thus, the generalizability of the performance across the range of possible GTCS-variability can be questioned and rightly warrants more investigation. However, the large amount of normal and varied physiological activity tested produced no false positives, which is an important metric for seizure detection systems; a high false positive rate will likely cause “alarm fatigue” and dissatisfaction with the user. Recordings were all obtained from an EMU-environment during day and night, and although the range of physical activity is restricted, a wide variety of physiological EMG occurred. This seems especially pertinent regarding facial movements since the jaw muscles engaged during GTCS likely also would be active during chewing, but we can see that chewing artifacts do not generate false positives during any mealtime intervals contained in the dataset. Additionally, the need for a GTCS alarm is principally during
unsupervised sleep, where the possibility of physiological activity that would generate a false alarm is even smaller. However, it could be argued that the fact that there were no false positives was mostly a fortuitous consequence of the low number of GTCS events the threshold was based on. Perhaps other individuals would have significantly lower amplitudes during the tonic interval so that it would become apparent that perfect classification would no longer be possible across a wider subset of GTCS? EMG patterns from GTCSs are arguably very stereotypical and so it seems unlikely that GTCS EMG would differs substantially, but it surely warrants additional investigation. The successful detection of an additional 11 GTCSs recorded with only scalp-EEG seems to suggest so. If other GTCS have markedly lower amplitudes so that separation from non-GTCS becomes imperfect by thresholding, a strategy to amend this could be based on additions to capture the sustained increase in amplitude over a longer period, possibly accepting a “cost” in the form of a longer detection latency. The MORTEMUS study points to a specific sequence of events leading from a GTCS to neurovegetative breakdown into respiratory depression and then cardiac arrest in the postictal interval several minutes after seizure onset [17]. In animal models of SUDEP, hypoventilation, not cardiac arrhythmia, was also the stronger predictor of death [18]. However, analysis of heart variability has identified pre-ictal parasympathetic overdrive and short QT as risk factor for SUDEP in a case where a patient died despite cardiopulmonary resuscitation [19]. SUDEP is theoretically preventable through interventions to support respiratory effort. A study of the effects of early peri-ictal nursing interventions on hypoxemia suggests that earlier intervention can shorten hypoxemia [20]. Our results indicate that GTCS detection strategy can work with a detection delay of about 4–13 s, which is comparable to other similar methods; two recent multi-center trials of GTCS detection by surface EMG obtained median detection latencies of 7.4 and 9 s [21,22]. These differences are small and roughly equivalent in terms of clinical relevance. If detection is based on movements during the clonic phase, which is common for accelerometry, the delays tend to be longer and then it might become more relevant if the purpose is to alarm care-givers for earlier assistance. One might suppose that our detection strategy could generalize to other motor seizures such as tonic seizures, but comparisons between tonic and generalized tonic-clonic seizures have shown that purely tonic seizures have a markedly lower amplitude [10], so this would again require different threshold parameterization. However, SUDEP mortality is associated to GTCS, not pure tonic seizures [17]. The detection strategy works on scalp-EEG with a fourelectrode montage but would likely also work for other electrode positions on the body. Other types of sensors located elsewhere on the body have also been investigated for GTCS detection. Spectral analysis of accelerometry (ACM) from wrist-worn sensor was used to detect 10/10 GTCS from 5/12 patients monitored for 246 h with a mean false positive ratio of 2/24 h [23]. However, the location of the ear-EEG device has the advantage that it also samples EEG. During sleep, this could allow for on demand sleep staging, which has recently been shown to work on automated basis with good accuracy compared to expert rated PSG [14]. This combined capability offers a “bundled solution” to multiple clinical problems. For instance, this could have relevance for the differentiation between abnormal movements during sleep, where knowledge about the sleep stage from which abnormal movement arise can inform a diagnostic process on REM and non-REM parasomnias. Bilateral use of ear-EEG is an advantage when the clinical problem is GTCS detection as accuracy can be improved by assuming that GTCS-associated activity should be reflected in all electrode
I.C. Zibrandtsen et al. / Seizure 59 (2018) 54–59
positions whereas random noise most likely will not. Other GTCS detection proposals utilizing only a single sensor location cannot take advantage of this. In our previous study [12] we described some irritation linked to prolonged use of hard ear-EEG pieces. This is unlikely to be an issue if the devices are only put in use during sleep. In addition, a new ear-EEG design recently described [24] uses dry-electrodes and a softer material and a flexible joint between the portion resting in the concha and the portion inside the ear-canal, which may turn out the decrease or eliminate the user-discomfort we described previously. 5. Conclusion We have described a simple algorithmic for GTCS detection based on a moving-median filtered sliding window variance estimate. The performance was tested on 820.7 h of simultaneous scalp-EEG and ear-EEG containing 3 GTCS from 3 different patients with detection latencies between 4 and 13 s. Sensitivity for ear-EEG was 100% (CI 29.2–100%) and no false positives. An additional 11 GTCS events from 6 different patients measured with only scalpEEG were also tested detecting all seizures with no false positives. Parameterization was developed and tuned post-hoc and the number of true events to be detected was low. The calculated detection latencies were based on an online-simulation run on a computer, not a wearable setup as is the intended future goal. Still, the results indicate that ear-EEG measurements during GTCS contain sufficient information for a seizure detection system to be feasible. Compared to GTCS detection by surface EMG collected from a sensor placed on the extremities, ear-EEG is of special interest in cases where EEG evaluation is relevant at the same time. Declarations of interest None. Acknowledgements The project was supported by the Danish Council for Strategic Research (1311-00009B) and the Elsass Foundation. References [1] Shorvon S., Tomson T. Sudden unexpected death in epilepsy. Lancet Lond Engl 2011;378(December (9808)):2028–38. [2] Tomson T, Surges R, Delamont R, Haywood S, Hesdorffer DC. Who to target in sudden unexpected death in epilepsy prevention and how? Risk factors, biomarkers, and intervention study designs. Epilepsia 2016;57(Suppl. 1):4–16. [3] Lamberts RJ, Thijs RD, Laffan A, Langan Y, Sander JW. Sudden unexpected death in epilepsy: people with nocturnal seizures may be at highest risk. Epilepsia 2012;53(February (2)):253–7.
59
[4] Kloster R, Engelskjon T. Sudden unexpected death in epilepsy (SUDEP): a clinical perspective and a search for risk factors. J Neurol Neurosurg Psychiatry 1999;67(October (4)):439–44. [5] Langan Y, Nashef L, Sander JW. Case-control study of SUDEP. Neurology 2005;64(April (7)):1131–3. [6] van Andel J, Thijs RD, de Weerd A, Arends J, Leijten F. Non-EEG based ambulatory seizure detection designed for home use: What is available and how will it influence epilepsy care? Epilepsy Behav EB 2016;57(Pt A):82–9. [7] Kramer U, Kipervasser S, Shlitner A, Kuzniecky R. A novel portable seizure detection alarm system: preliminary results. J Clin Neurophysiol Off Publ Am Electroencephalogr Soc 2011;28(February (1)):36–8. [8] Conradsen I, Beniczky S, Hoppe K, Wolf P, Sorensen HBD. Automated algorithm for generalized tonic-clonic epileptic seizure onset detection based on sEMG zero-crossing rate. IEEE Trans Biomed Eng 2012;59(February (2)):579–85. [9] Conradsen I, Wolf P, Sams T, Sorensen HBD, Beniczky S. Patterns of muscle activation during generalized tonic and tonic-clonic epileptic seizures. Epilepsia 2011;52(November (11)):2125–32. [10] Beniczky S, Conradsen I, Pressler R, Wolf P. Quantitative analysis of surface electromyography: biomarkers for convulsive seizures. Clin Neurophysiol. 2016;127(August (8)):2900–7. [11] Looney D, Kidmose P, Park C, Ungstrup M, Rank M, Rosenkranz K, et al. The inthe-ear recording concept: user-centered and wearable brain monitoring. IEEE Pulse 2012;3(December (6)):32–42. [12] Zibrandtsen IC, Kidmose P, Christensen CB, Kjaer TW. Ear-EEG detects ictal and interictal abnormalities in focal and generalized epilepsy — A comparison with scalp EEG monitoring. Clin Neurophysiol Off J Int Fed Clin Neurophysiol 2017;128(October (12)):2454–61. [13] Zibrandtsen I, Kidmose P, Otto M, Ibsen J, Kjaer TW. Case comparison of sleep features from ear-EEG and scalp-EEG. Sleep Sci 2016;9(April (2)):69–72. [14] Mikkelsen KB, Villadsen DB, Otto M, Kidmose P. Automatic sleep staging using ear-EEG. Biomed Eng Online 2017;16(September (1)):111. [15] Bech Christensen C, Harte JM, Lunner T, Kidmose P. Ear-EEG based objective hearing threshold estimation evaluated on normal hearing subjects. IEEE Trans Biomed Eng 2017(August). [16] Kappel SL, Looney D, Mandic DP, Kidmose P. Physiological artifacts in scalp EEG and ear-EEG. Biomed Eng Online [Internet] 201716(August). . Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5553928/. [17] Ryvlin P, Nashef L, Lhatoo SD, Bateman LM, Bird J, Bleasel A, et al. Incidence and mechanisms of cardiorespiratory arrests in epilepsy monitoring units (MORTEMUS): a retrospective study. Lancet Neurol 2013;12(October (10)):966–77. [18] Johnston SC, Horn JK, Valente J, Simon RP. The role of hypoventilation in a sheep model of epileptic sudden death. Ann Neurol 1995;37(April (4)):531–7. [19] Jeppesen J, Fuglsang-Frederiksen A, Brugada R, Pedersen B, Rubboli G, Johansen P, et al. Heart rate variability analysis indicates preictal parasympathetic overdrive preceding seizure-induced cardiac dysrhythmias leading to sudden unexpected death in a patient with epilepsy. Epilepsia 2014;55(July (7)):e67–71. [20] Seyal M, Bateman LM, Li C-S. Impact of periictal interventions on respiratory dysfunction, postictal EEG suppression, and postictal immobility. Epilepsia 2013;54(February (2)):377–82. [21] Halford JJ, Sperling MR, Nair DR, Dlugos DJ, Tatum WO, Harvey J, et al. Detection of generalized tonic-clonic seizures using surface electromyographic monitoring. Epilepsia 2017;58(November (11)):1861–9. [22] Beniczky S, Conradsen I, Henning O, Fabricius M, Wolf P. Automated real-time detection of tonic-clonic seizures using a wearable EMG device. Neurology 2018;90(January (5)):e428–34. [23] Joo HS, Han S-H, Lee J, Jang DP, Kang JK, Woo J. Spectral analysis of acceleration data for detection of generalized tonic-clonic seizures. Sensors [Internet] 201717(February (3)). . Available from: https://www.ncbi.nlm.nih.gov/pmc/ articles/PMC5375767/. [24] Kappel SL, Rank ML, Toft HO, Andersen M, Kidmose P. Dry-contact Electrode Ear-EEG IEEE TBME. 2018.