Highlights
- •Patient specific algorithms are crucial for achieving accurate detection devices.
- •Multimodal detection systems are needed to meet the requirements of seizure detection.
- •Closed-loop systems are recommended because they provide active feedback.
- •The systems will improve as information from different patients accumulates.
- •A proposal of seizure detection devices for each seizure type is made.
Abstract
Epilepsy affects almost 1% of the population and most of the approximately 20–30% of patients with refractory epilepsy have one or more seizures per month. Seizure detection devices allow an objective assessment of seizure frequency and a treatment tailored to the individual patient. A rapid recognition and treatment of seizures through closed-loop systems could potentially decrease morbidity and mortality in epilepsy. However, no single detection device can detect all seizure types. Therefore, the choice of a seizure detection device should consider the patient-specific seizure semiologies.
This review of the literature evaluates seizure detection devices and their effectiveness for different seizure types. Our aim is to summarize current evidence, offer suggestions on how to select the most suitable seizure detection device for each patient and provide guidance to physicians, families and researchers when choosing or designing seizure detection devices. Further, this review will guide future prospective validation studies.
Keywords
1. Introduction
Epilepsy is one of the most common neurological disorders with an incidence of approximately 40–70/100,000 per year in adults [
[1]
] and 41–187/100,000 per year in children [[2]
], being particularly frequent in rural and underdeveloped areas [1
, 3
, 4
, 5
]. Almost 47% of patients will become seizure-free with the first anti-seizure medication trial and an additional 14% of patients with a second or third medicine [[6]
]. Despite optimal medication management, about 20–30% of patients with epilepsy will continue to have more than one seizure per month, 12% will have weekly seizures, and 8% daily seizures [3
, 7
, 8
]. Patients with active epilepsy have 4–5 times higher standardized mortality ratios than epilepsy patients who are seizure free, especially in the initial years after a diagnosis of epilepsy [9
, 10
, 11
, 12
, 13
]. The mortality associated with epilepsy has remained stable in the last 50 years despite the introduction of multiple new anti-seizure medications [14
, 15
]. Further, seizure unpredictability worsens the quality of life (QOL) of patients with epilepsy and their families [[16]
]. Medical treatment tailored to the individual patient's seizures might eventually decrease mortality and increase QOL, but in order to do so, both seizures and response to treatment need to be objectively quantified.Currently, seizure tracking relies on subjective patient and family recall and may be influenced by the capacity to identify seizures, the level of awareness during the event, and the ability to remember details afterward. In a series of 451 children with absence seizures, up to 30% of patients with no clinically detected seizures, even during hyperventilation, had seizures detected on 1 h-EEG recordings [
17
, 18
]. In an inpatient study the number of seizures detected with EEG was 29 times higher than the number reported by families and 7 times higher than the number clinically observed by nurses [[19]
]. In several studies patients report only approximately half of their seizures, and even less during sleep [20
, 21
, 22
]. Seizure detection devices provide more accurate seizure quantification, allowing clinicians to tailor treatment more objectively. In addition, seizure prediction devices may alert when an upcoming seizure is going to occur and may enhance patient and family confidence and improving QOL.The main seizure detection modalities are outlined in Table 1. There is extensive literature describing seizure detectors, but limited data on which are optimal for each seizure type [
[23]
]. This manuscript aims to address this gap in knowledge by providing a guide on how to select specific devices for individual seizure types.Table 1Types of seizure detection devices.
Electroencephalogram (EEG) Intracranial EEG Surface electromyography (sEMG) Electrodermal activity (EDA) Electrocardiography (EKG) Accelerometry (ACM) Video detection systems Mattress sensors Seizure-alert dogs Implanted advisory systems Cerebral oxygen saturation sensors Near infrared spectroscopy (NIRS) Skin temperature Respiratory monitor |
2. Seizure types
2.1 Seizure type definitions
Hughlings Jackson defined an epileptic seizure in 1870 as a “symptom … of an occasional, an excessive and a disorderly discharge of nerve tissue” [
[24]
]. The definition of epileptic seizure was reformulated by the ILAE group in 2005 to, “a transient occurrence of signs and/or symptoms due to abnormal excessive or synchronous neuronal activity in the brain” [[24]
]. Specific seizure types have varying features that may enable detection (Table 1).2.2 Different seizures are best captured by different types of sensors
Knowing the main semiological components of seizures is a basic initial step in the process of selecting the best seizure detection device for each patient. Each seizure type consists of one or more phenomena occurring simultaneously or sequentially [
[25]
]. In order to evaluate clinical features, the two main components that can be assessed are movement and physiological signals. The movement refers to specific body parts, as the limbs involved in a GTCS, or head or eye deviation. These could be detected by accelerometry, surface electromyography (sEMG), video monitoring, mattress sensors, electro-oculogram, or seizure-alert dogs. The physiological signals include heart rate, respiratory rate, sweating and temperature. These could be detected by electrocardiogram (EKG), sweating by EDA, temperature by a wristband and changes in respiratory rate with a thoracic band. Generalized tonic-clonic seizures (GTCS) may present with violent body movements and often prominent autonomic changes. Therefore, several sensor types recognize GTCS more readily than other seizure types. On the other end of the spectrum, absence are challenging to capture as they consist of a brief decrease in awareness with minimal associated movements (Table 2), and thus are often only picked up by an observer or by EEG.Table 2Definitions and main findings of the principal seizure types.
Seizure type | Main features Definition | Movement | Sweating | HR/EKG changes |
---|---|---|---|---|
Atonic | Sudden loss of muscle tone lasting 1–2 s, involving head, trunk, jaw or limb musculature | ± | * | * |
Autonomic | An alteration of the autonomic nervous system, involving cardiovascular, pupillary, gastrointestinal, sudomotor, vasomotor and thermoregulatory functions | − | + | + |
Clonic | Semirhythmic high-amplitude movements that involve the same muscle groups | + | + | + |
Myoclonic | Sudden, brief involuntary single or multiple low-amplitude contraction(s) of muscle(s) or muscle groups | + | * | − |
Epileptic spasm | Sudden flexion and/or extension of predominantly proximal muscles, more sustained than a myoclonic movement but not so prolonged as a tonic seizure | + | * | * |
Focal dyscognitive seizure | Disturbance of cognition is the most apparent feature accompanied by changes in perception, attention, emotion, memory and executive function | ± | + | + |
GTCS | Bilateral symmetric tonic contraction followed by generalized clonic movements of somatic muscles, usually accompanied by autonomic phenomena | + | + | + |
Hypermotor | Involves predominantly proximal limb or axial muscles producing irregular, sequential semipurposeful movements. | + | + | + |
Tonic | A sustained increase in muscle contraction lasting seconds to minutes | + | + | + |
*No data on this particular issue.
Source: Blume, W.T et al. Glossary of descriptive terminology for ictal semiology: report of the ILAE task force on classification and terminology. Epilepsia, 2001; 42(9): p. 1212–8.
Legend: EKG: Electrocardiogram. GTCS: Generalized tonic-clonic seizure. HR: Heart rate.
The ideal seizure detection sensor should be able to detect movement in body parts and changes in physiological signals simultaneously. The system could be even more effective if it also interacted with the patient in an active way that could allow testing for awareness. This could be done with a gaming device that keeps track of the patient's responsiveness while playing. If the patient stops playing, the device could ask the patient to follow some specific commands, and in turn the system would ideally alert a designated caregiver in the event of impaired performance. This could also be done by robot or external device that wirelessly receives the input of movement and physiological signals from the monitoring device. For example a wristband analyzes signals in real time, and when parameters are altered, it will interact with the patient to assert for consciousness. This has an advantage compared to a gaming system in that it could interact with the patient when detecting abnormal physiological parameters during the night, whereas the gaming device's functionality is limited to when the patient is awake. An integral system would hypothetically allow detection of all seizure types, including those considered subtle like absence seizures.
3. Types of seizure detection sensors
3.1 Unimodal detection devices
3.1.1 Electroencephalogram
Video-EEG has long been considered the gold standard for the diagnosis of seizures. Several groups have developed algorithms for automatic seizure detection based on EEG [
26
, 27
, 28
, 29
, 30
, 31
]. Most of them tested their algorithms with data from the Freiburg Seizure Prediction and the European Epilepsy Databases [26
, 27
, 30
, 31
, 32
, 33
, 34
]. Most researchers used data from 2 to 6 electrodes trying to recreate the conditions for ambulatory performance. The Freiburg database uses three electrodes close to the focal area and three remote from it [26
, 27
, 32
, 34
].Researchers first differentiated the preictal period from the ictal period – which can be challenging as the transition is sometimes subtle [
33
, 35
]. The preictal period is seizure and patient specific, and its identification requires a training period of the algorithm for the individual patient [27
, 29
, 30
, 33
, 35
]. Investigators developed an algorithm based on the parameters that yielded the best performance [29
, 30
, 32
, 35
].Seizures detected with this approach were focal dyscognitive seizures, focal without dyscognitive changes, secondarily generalized seizures, and absence seizures [
27
, 28
, 33
, 34
, 35
, 36
, 37
]. Intracranial EEG yielded a sensitivity of 80.5–98.8% and a false detection rate (FDR) of 0.054–1/h, within a prediction horizon of 30–60 min [26
, 27
, 28
, 29
, 30
, 31
, 34
, 35
, 38
]. The data from scalp EEG had a sensitivity of 74–99% and an FDR of 0.28–1/h, with an anticipation time of 16 min [28
, 32
, 36
, 39
, 40
]. A device for detection of absence seizures achieved a sensitivity of 95%, with an FDR of 0.037/h, when tested in an ambulatory setting in one patient [[41]
]. The electrodes were placed in F7-Fp1, as this yielded the best results in a previous study [[37]
].Patients must wear scalp electrodes and remain attached to EEG equipment during monitoring, which increases artifact, is impractical, and potentially leads to stigmatization and discomfort [
[42]
]. The current devices for ambulatory EEG could be used as a model, as they are light weight and patients can wear them attached to a belt or over the shoulder [[43]
]. Several researchers are working on the development of wireless EEG modalities, with few and small electrodes [44
, 45
, 46
, 47
, 48
, 49
]. A group used four electrodes installed in a 20 cm area and demonstrated that this is feasible and comparable to a 10–20 EEG system [[44]
]. The Emotiv EPOC is a low-cost wireless system with 14 electrodes and a battery life of 12 h enabling the device to be worn during the day [50
, 51
]. The B-Alert is a wireless EEG head unit with 21 channels (20 electrodes), in which the data is transmitted to the receiver [[50]
]. These two systems were compared and B-Alert (in two different versions) was the most similar to a research-grade system and it detected all of the EEG expected features [50
, 51
]. Another wireless system was developed for ambulances and emergency departments with a short preparation time because it is a cap with six channels [[52]
]. There is also a portable, wireless system that uploads the data to a smartphone [[47]
]. It has two electrodes, one over the area of interest and the other over the mastoid, and it was able to record a seizure [[47]
]. Finally a waterproof, 2 electrodes (1-channel) EEG monitoring device called the EEG Patch TM is able to track seizures for 7 days and allows home monitoring [[53]
]. The experience with these devices detecting seizures and their sensitivity compared to standard EEG is not yet published for some of them. They represent promising systems that when linked with an automated detection algorithm could allow for accurate seizure detection.Alternatively, intracranial EEG is an invasive procedure that requires close supervision and is associated with risk of infection. In the future, devices with only a few and small electrodes will increase portability, and machine learning algorithms will further automatize the seizure identification process. Large studies targeting homogeneous seizure types with machine learning algorithms will clarify the role of EEG as a potential portable seizure detection device.
3.1.2 Surface electromyography (sEMG)
Most seizures have a motor component; therefore the analysis of muscle activity with sEMG is a viable option for seizure detection [
[54]
]. sEMG helps differentiate epileptic seizures from non-epileptic seizures: epileptic seizures have a large proportion of EMG activity in the frequency band above 100–150 Hz [55
, 56
]. sEMG detects muscle activity with as few as one channel; deltoid and anterior tibialis muscles are the preferred placement sites [55
, 56
, 57
].Tonic stiffening consists of an intense muscle contraction, which allows for early GTCS detection [
54
, 55
, 57
]. In a study, the efficacy of deltoid muscle detection site was better than for the anterior tibialis muscle, with a sensitivity of 100%, mean detection latency (time from beginning of a seizure to its detection) of 13.7 s, and FDR of one alarm per 24 h [[55]
]. When using only the nocturnal data, the FDR improved to one every 10 nights [[55]
]. sEMG in the anterior tibialis muscle had a sensitivity of 57%, but the FDR improved to one false alarm in 12 days with a latency of 25 s [[57]
]. The sEMG on tonic seizures, recorded at the deltoid muscle, had a sensitivity of 53–63% and a FDR of 1.49–4.03 [[54]
]. A recent study placed the sEMG electrodes in the biceps and triceps detecting 95% of GTCs but none of the other seizures (myoclonic, tonic, absence, and focal seizures with or without loss of consciousness) [[58]
].Disadvantages of sEMG sensors include discomfort when strongly fixed to the skin and potential for detachment [
[57]
]. Better results may be achieved when certain parameters are tailored to the individual patient, especially for tonic seizures [[54]
]. In summary, sEMG reliably detects GTCS and tonic seizures and can potentially detect other seizure types with a motor component.3.1.3 Electrodermal activity (EDA)
Modulation in skin conductance is referred to as EDA, and it reflects the activity of the sympathetic branch of the autonomic nervous system [
59
, 60
]. Sweat increases the conductance of an applied current [[59]
]. The device applies direct current to the stratum corneum beneath the electrodes, and measures the EDA in the ventral side of distal forearms [59
, 60
]. The rationale for the use of this device is that epileptic seizures transiently increase EDA [[60]
].In a study with seven patients the EDA was significantly elevated immediately after the onset of each EEG seizure, including GTCS and focal dyscognitive seizures [
[60]
]. But the change in EDA was higher and remained elevated for a longer period in GTCS compared to focal dyscognitive seizures [[60]
]. In a study including 11 patients, 100% of GTCS had a greater than 2 standard deviation (SD) increase in EDA, but only 86% of focal dyscognitive seizures had it, with a median latency for all seizures of 33 s [[61]
].Studies in adult and pediatric patients have demonstrated a strong correlation between the duration of postictal generalized EEG suppression (PGES) and the degree of EDA response. Children seem to have lower PGES duration, which could explain lower sudden unexplained death in epilepsy (SUDEP) rates, although they have more sympathetic activation and diminished vagal tone when compared to adults [
62
, 63
]. EDA recordings may also be able to help us better understand the pathophysiology of SUDEP [61
, 62
].Disadvantages include that EDA recording is susceptible to motion and pressure artifacts and it could be uncomfortable or obtrusive. The EDA in the ventral side of the distal forearm is well tolerated, even for long periods [
[59]
]. Large studies on continuous ambulatory autonomic monitoring will provide insights to optimize this promising modality.3.1.4 Electrocardiogram (EKG)
Cardiovascular changes are relatively easy to measure, and in patients with epilepsy they are particularly relevant as they may be linked to SUDEP [
64
, 65
]. EKG can be recorded from a single channel and has a higher signal to noise ratio than EEG [[66]
]. Multiple studies have aimed to characterize heart rate changes before, during, and after seizures [64
, 65
, 66
, 67
, 68
, 69
, 70
]. Heart rate changes could be explained by increased motor activity, release of catecholamines, sympathetic and parasympathetic shifts, activation of limbic structures, increased neuronal firing, or a combination of these and other unknown factors [64
, 66
, 69
, 71
]. Another aspect that must be taken into consideration is that heart rate also depends on the state of vigilance so sensors using heart rate to detect seizures might be affected by this [72
, 73
]. The pattern of heart rate changes seems to be patient-specific, reflecting the individual spread and evolution of seizure activity and warrant the development of patient-tailored detection algorithms [64
, 69
, 71
, 74
].EKG has been used to detect focal seizures, secondarily generalized seizures and GTCS [
[75]
]. EKG abnormalities have been linearly correlated to electrocorticogram (ECoG) seizure severity, proving the feasibility of EKG as a seizure detection device [[67]
]. Ictal tachycardia was more prominent when arising from the right hemisphere [65
, 74
]. In contrast, short myoclonic seizures often did not produce heart changes [[64]
]. In a large study, 73% of focal seizures had heart rate increase, and in 23% it preceded EEG onset [[68]
]. In another study in children, ictal tachycardia was present in 70% of focal seizures (temporal or frontal onset), but not in generalized seizures [[75]
]. One interesting study found that heart rate increased from the preictal to the ictal period in 74% of the patients and it was higher in seizures with secondary generalization than in complex partial (focal dyscognitive) seizures [[76]
]. One of the few studies that included generalized seizures detected heart rate changes in 35% of seizures, either GTCS or secondarily generalized seizures [[65]
]. The sensitivity of the automated detection algorithms was 90–98% in seizures with heart rate changes, with a greater than 50% positive predictive value [64
, 77
]. The latency was between 0.8 s and 10 min for all seizure types [66
, 69
, 71
, 74
, 77
].Regarding ambulatory monitoring, a group developed a wireless device for real time detection of seizures that allows unobtrusive monitoring with good mobility. This may be effective for detection of tonic-clonic, tonic, clonic and hypermotor seizures [
[78]
]. The monitor is composed of a wireless sensor board, ultra-low power EKG sensor readout, accelerometer, and micro-secure digital-card. The detection algorithm has a sensitivity of 99.8% and positive predictive value of 99.8% [[79]
]. The device with two electrodes was placed on the left arm of four healthy volunteers who considered it comfortable and reported no inconvenience during sleep [[78]
]. A smartphone application monitors heart rate remotely based on skin color changes [[80]
]. Since cardiac pulse leads to subtle changes in skin color, a photoplethysmographic signal can be measured recording the face with the front facing-camera of a smartphone. Heart rates were measured for a minute with good accuracy in adults who held the smartphone 30 centimeters away from their face without moving [[80]
]. This technique could ultimately prove useful in patients with seizures having a motor component. Another option utilizes cardiac-based activation vagus nerve stimulation as part of a commercially available closed-loop system [[81]
]. This cardiac-based seizure detection algorithm had a sensitivity of 80%. The effect on seizure frequency was moderate but there was significant improvement in QOL [[81]
].Some disadvantages of EKG as a seizure detection signal include the low specificity of changes in heart rate [
64
, 82
], low stability of electrodes, and discomfort with long-term use, but these could avoided with wireless, video, or VNS activated devices. The findings above and the development of an ambulatory device demonstrate that automated EKG seizure detection is possible, particularly when parameters are tailored to the individual patient.3.1.5 Accelerometry (ACM)
ACM has been used for motor seizures as it detects changes in velocity and direction [
19
, 83
, 84
, 85
]. The signal is recorded by means of a three-axis motion/accelerometer sensor, a microprocessor, and a small, rechargeable battery, usually placed on a limb [83
, 84
, 86
, 87
, 88
]. The main challenge is to differentiate seizures from normal, daily, repetitive movements [19
, 89
]. Some systems have a cancel button and this gives the opportunity to indicate that a movement was a false alarm, avoiding a false-positive alert to the caregiver [[84]
].This modality was able to detect focal seizures with minimal motor component, GTCS, secondarily generalized seizures, myoclonic, clonic, tonic and hypermotor seizures [
19
, 23
, 83
, 84
, 85
, 87
, 90
, 91
]. Clonic seizures present with a burst-like pattern, which was better identified and discriminated from other movements [19
, 90
]. Tonic seizures are block-shaped because the acceleration is almost constant. They resemble slow normal movements, which makes them harder to identify [19
, 85
]. Focal dyscognitive seizures without motor phenomena and absence seizures were not detected [19
, 83
, 84
]. Sensitivity ranges between 16 and 100%; one study had a FDR of 0.2/day [83
, 84
, 86
, 87
, 89
, 90
, 91
, 92
]. Seizures were detected 9–60 s after seizure onset [84
, 87
]. The same accuracy for nocturnal and daytime seizures was achieved [19
, 83
, 84
]. One study validated the system in a home environment detecting 78.5% of the seizures reported by parents, with 0.6 false alarms per night [[92]
].Some disadvantages include that the system is restricted to seizures with a motor component, and that seizures are not detected when there is an obstacle to free limb movement [
[83]
]. More studies in ambulatory settings are needed, as most studies have been in epilepsy monitoring units where movements might be limited [83
, 87
]. This modality has good sensitivity with good night detection rates, and most patients and families found the device user-friendly [83
, 84
, 90
].3.1.6 Video detection systems
Automatic video detection systems use velocity, area, duration, rotation, oscillation, angular speed, and/or displacement (motion trajectory) to detect seizures [
93
, 94
, 95
, 96
, 97
]. The underlying concept is to detect complex motor patterns by automatic interpretation of video data [[93]
]. The systems have been classified as marker-based or marker-free, depending on whether the cameras track detectable markers placed in relevant places [[93]
].Seizure types that can be captured by video include focal, hypermotor, myoclonic, and clonic [
93
, 94
, 95
, 98
]. Myoclonic seizures are detected with good sensitivity and specificity with a marker-based system using spatio-temporal interest points [[95]
]. Reference markers could be placed on the head, trunk and extremities to asses for movement when evaluated with infrared light by a video system. This was done with frontal and temporal lobe seizures to evaluate for the lateralizing value of ictal head turning [99
, 100
]. The overall sensitivity varies from 75 to 100%, positive predictive value over 85%, and specificity of 53–93% [93
, 94
, 95
, 96
]. Reported latencies range from 4.6 to 21.4 s [[94]
].Marker-based devices present with the shortcoming that sensors can be uncomfortable or dislocate over time [
94
, 97
]. Marker-free systems detect only seizures with a motor component, and they are more limited to the area covered by video: the patient must be visible and properly placed [[93]
]. Seizure detection based on video is feasible, but it recognizes mainly seizures with large movements.3.1.7 Mattress sensors
Mattress sensors were developed because most cases of SUDEP occur in patients with GTCS while unsupervised in bed at night [
101
, 102
, 103
, 104
, 105
]. These devices consist of a sensor placed under the patient's mattress and connected to a monitor. The sensor alerts the family when it detects a stimulus above the set threshold [102
, 103
, 105
]. The MP5 bed seizure monitor detects movement and noise frequency and intensity, with a minimum patient weight of 25 kg [[102]
]. The ST-2 bed alarm mainly detects abnormal bed motion in patients weighting more than 6.4 kg [[103]
]. The Emfit monitor detects micro-movements, without a minimum weight [[105]
].For GTCS during sleep the MP5 sensitivity was 62.5% and specificity 90% [
[102]
]. In another study including multiple seizure types, the device only detected one GTCS during sleep (16.7%) [[103]
]. The ST-2 did not detect any of 20 nocturnal seizures. It only detected one focal dyscognitive seizure with motor phenomena of the 26 seizures recorded in awake patients, with an overall sensitivity of 2.2%. These two devices were designed to detect GTCS, but their sensitivity is suboptimal for seizures overall [[103]
]. The Emfit sensitivity was 30% during daytime and 85%during sleep for GTCS [[105]
].Most of these devices have the disadvantages of having a weight restriction, detecting only seizures with rhythmic movements, and a low sensitivity [
103
, 105
]. Mattress devices allow for adjustment of the parameters to compensate for individual differences in movement during sleep [[103]
]. Individual calibration and testing over a couple of nights in a home setting are recommended [[102]
]. None of these systems performed as well as video EEG in detecting nocturnal seizures but this is a step in the right direction for a much needed device.3.1.8 Seizure-alert dogs
Seizure-alert dogs are those that display some attention-getting behavior prior to human appreciation of an impending seizure event [
106
, 107
, 108
]. While the trigger to which these animals respond is not completely understood, it is believed that they alert to subtle human behavioral changes [106
, 107
, 109
]. Hypothetically they may be responding to changes in human respiratory or heart rate or even olfactory phenomena, perhaps pheromone production [106
, 108
]. They are able to alert from 30 s to 45 min before seizure onset [108
, 109
, 110
].Seizure-alert dogs have been reported to detect atonic, focal dyscognitive, and GTC seizures [
106
, 110
, 111
]. One of the few seizures detected by a seizure-alert dog with concomitant video EEG monitoring was focal dyscognitive [[111]
]. One study reported a median sensitivity estimate of 80% and specificity of 100% [[107]
]. Another study found a 43% mean reduction in seizure frequency with the use of seizure-alert dogs, hypothetically due to diminished stress [[109]
].There are few reports of EEG or other monitoring concomitant to the dogs’ behavior and most articles used diaries or questionnaires as the gold-standard, raising some skepticism on the reliability of the information [
111
, 112
]. Other difficulties are that seizure-alert dogs also respond to psychogenic seizures, as reported on at least seven well documented cases, and they are not able to monitor patients during their own sleep [82
, 111
, 113
]. Companionship seems to be one of the primary benefits and this could lead to the increase in QOL [110
, 111
, 112
]. To date, no rigorous data confirms whether seizure prediction by seizure-alert dogs is better than chance [[106]
].3.1.9 Cerebral oxygen saturation sensors
Increased cerebral blood flood precedes the onset of a clinical seizure in temporal lobe epilepsy by approximately 20 min [
[114]
]. This finding was confirmed by single-proton emission computed tomography (SPECT) [[115]
]. One study placed transcutaneous regional cerebral oxygen saturation (rSO2) sensors on each side of the forehead in five patients with GTCS, and compared this to video-EEG. The mean rSO2 value significantly increased in the preictal period by at least 3 SDs in 4 out of the 7 registered GTCS, on average 18 min prior to EEG seizure onset [[116]
].Larger studies will clarify the relevance of this approach in other seizure types. These findings suggest that cerebral oxygen saturation sensors could have a role as automatic prediction devices, at least in GTCS, with a prolonged time of detection before seizure onset [
[116]
].3.1.10 Near-infrared spectroscopy (NIRS)
Wavelengths used in NIRS measure cerebral oxygen saturation by using the specific absorption properties of tissues in the near infrared range [
117
, 118
]. The spectrophotometer, typically placed on the forehead [117
, 118
, 119
], emits light into the tissue from the surface of the scalp, and then collects it from a detector close to the emitter [[118]
].Seizures analyzed with this approach are mainly focal, including focal dyscognitive, focal without impairment of consciousness, and focal with secondary generalization, but also absence seizures [
117
, 119
]. One study in temporal lobe seizures found an increase in rSO2 in the preictal period, followed by a decline around seizure onset, and then a postictal increase [[117]
]. The mean preictal rSO2 increase was 7% above baseline, and 5 min before seizure onset [[117]
]. A study in children with generalized spike and wave discharges had similar results [[118]
]. The authors revealed that frontal generalized spike and wave discharges were associated with moderate oxygenation 10 s before the discharge, followed by a strong deoxygenation, then a second increase in oxygenation, and a return to baseline [[118]
]. In another study, the hemodynamic changes in the frontal lobe during seizures were only slightly higher than during non-seizure activities [[119]
].One of the disadvantages of this system is that the sensor pad might be difficult to wear from a cosmetic point of view, since it must be surrounded by a black cloth so that no external light interferes with the NIRS signal [
[119]
]. In addition, one of its biggest challenges is developing a generic algorithm, since there is a huge diversity in hemodynamic changes amongst patients [[119]
]. The findings suggest that NIRS might be a noninvasive detection system, at least for focal seizures, with a patient specific algorithm [[117]
].3.1.11 Implanted advisory system
This system was developed to predict and quantify seizures in adults with refractory focal seizures [
[120]
]. Two silicon leads, each with eight contacts, were placed over the quadrant suspected to contain the epileptogenic zone [[120]
]. A hand-held device analyzed the recorded EEG based on a patient specific algorithm, developed during the data collection phase. This device transmitted audible and visual signals showing the likelihood of seizure occurrence minutes to hours before it actually occurred [[120]
].The mean sensitivity after four months of implantation was 66% (10 patients). The usefulness of seizure prediction was inconclusive, and the variability in warning times and difficulties adapting to the system seemed to play a role [
[120]
].The main difficulty with this system is that it requires an invasive procedure that could have complications [
[120]
]. In one year after implantation, four patients (27%) had serious adverse effects, including device migration and infection. This study demonstrated for the first time that a seizure detection device predicted seizures better than chance [[120]
].3.1.12 Skin temperature
The ratio of mean temperature during sleep and awake periods could be related to seizures over the following days as suggested by recent work with wristband data in children with epilepsy. This work is in progress but if confirmed it suggests that skin temperature could also be analyzed for seizure detection [
[121]
].3.1.13 Respiratory monitor
Cardiac and respiratory abnormalities have been suggested as a cause for SUDEP [
122
, 123
, 124
]. In a study with 26 children (101 seizures) an elastic belt was used to monitor chest and abdominal excursion by respiratory inductance plethysmography [[122]
]. Of the seizures 39% were associated with ictal central apnea, 34% with ictal tachypnea, and 13% with ictal bradypnea. No patients had ictal obstructive sleep apnea [[122]
]. Ictal apnea was more frequent in temporal lobe seizures when compared to frontal seizures (OR 8.04, p = 0.0005). A thoracic band for detecting cardiac and respiratory changes could be coupled with a movement detection modality, potentially improving the overall sensitivity.3.2 Multimodal detection devices
Seizure detection is more accurate if it combines more than one modality, as multimodal systems have shown increased sensitivity and lower FDR [
125
, 126
, 127
, 128
].3.2.1 EDA and ACM
The combination of EDA and ACM might improve detection of motor seizures and those with autonomic involvement. A system with EDA and ACM biosensors-placed on the ventral forearms of six patients with GTCS, during awake and sleep states – yielded 94% sensitivity and one false alarm per 24 h. The FDR was lower than for ACM alone [
[129]
]. The mean latency from clinical onset to detection was 31 s [[129]
]. This emerging modality may benefit from more training in different settings, and preliminary results are promising.3.2.2 sEMG and ACM
This combination yielded better results than each signal individually, and it also enhances detection of seizures with motor and autonomic system involvement. The modality was proposed because ACM seems to be more sensitive in detecting the clonic phase, and sEMG the tonic phase of seizures [
[126]
]. In a study of children with tonic-clonic seizures, the combination of two ACM and two sEMG sensors achieved a sensitivity of 91%, with a FDR of 0.5/12 h and a latency of 10.5 s [[126]
]. The FDR was higher than previous studies on adults, which probably reflects the fact that children usually move more than adults at baseline [[126]
]. The ACM sensors yielded best results on the left wrist (non-dominant hand) and the right ankle [[126]
]. The use of isolated sEMG yielded a slightly lower sensitivity (82% vs 86%) when compared to isolated ACM, but FDR and latency were lower [[126]
]. sEMG alone detected tonic and hypermotor seizures, while ACM detected hypermotor, tonic, and clonic seizures [[126]
].3.2.3 Magnetometer and ACM
The magnetometer is used as a seizure detection sensor because it characterizes movement in the horizontal plane, as seen in a tonic seizure [
[130]
]. Magnetometers measure body inclination using a 3D Earth's magnetic field sensor and track changes in orientation in 3D space [[131]
]. Three sensor modules, each containing a tri-axial ACM and a tri-axial magnetometer, were placed in the head and both wrists of patients [[130]
]. In a first phase, sensors discriminated tonic activity, clonic activity, hypermotor movements, and no movements [[130]
]. Afterwards the system was tried in 86 patients in an epilepsy monitoring unit [[132]
]. There was accurate classification of tonic seizures in 62% of cases and tonic-clonic in 90% of the cases [[132]
]. Hypermotor seizures were classified as tonic-clonic seizures in 100%. The detectors had a 80% sensitivity with 95% specificity [[132]
].3.2.4 Video, ACM and radar-induced activity recording (VARIA)
Radars use wave transmission to identify motion caused, for example, by a change in the patient's position in bed. The combination of radar, video and ACM may increase the ability to detect seizures. This multimodal system considers the top 10% of events per night with the most abnormal movements, regardless of the number of seizures that occurred [
[23]
]. VARIA was tested in an 8-year-old patient with focal epilepsy with secondarily generalization. This system achieved a sensitivity of 56% and 20 false alarms per night [[23]
]. The system detected some seizures that were missed by the staff [[23]
]. Limitations of this system include numerous, easily removable bracelets [[23]
].3.2.5 Multi-modal intelligent seizure acquisition (MISA) system
The MISA system includes sEMG, magnetometers, ACM, and gyroscopes allowing full body movement description. Gyroscopes provide information on the rotation of each joint [
[125]
]. The system consists of a suit with 16 sensors, each one containing 3D accelerometer, 3D magnetometer, and 3D gyroscope, as well as 28 sEMG electrodes placed on 14 muscles [[127]
]. The MISA system was first tried on three subjects that simulated tonic-clonic, versive asymmetric tonic, and myoclonic seizures [[127]
]. The system achieved the best results when all modalities were used [[127]
]. The non-specific system detected 98% of simulated seizures with one false positive per hour, and the subject specific system detected all seizures, with only one false positive in 4 h [[127]
]. On another study the MISA system had the best results followed by the combination ACM and sEMG [[125]
]. Some patients found the suit uncomfortable so the prototype is under modification, probably involving smaller and fewer electrodes [125
, 128
].