# Seizure detection

Post-publication activity

Curator: Mark G. Frei

The term “seizure detection” generally refers to the use of an automated algorithm (a “seizure detection algorithm” or “SDA”) to recognize that a seizure is occurring (or has occurred) through analysis of biologic signals recorded from a patient with epilepsy. Essentially, the goal is to receive and analyze a set of signals and transform the information they contain into an output signal or indicator of whether or not the patient is in a state of seizure (Figure 1-Figure 2). Important objectives are to perform this transformation as quickly, efficiently and accurately as possible. For most applications, “real-time” seizure detection is needed, requiring “online” analysis of signal data points as soon as they are available (e.g., received into a processor), with detection decisions being made with negligible delay and without use of “future information” (i.e., only using information that is available at the time the detection is made).

 Figure 1: (A) Electrocorticogram (“ECoG”) recorded from two 8-contact needle electrodes implanted bilaterally in temporal lobes, showing a seizure. (B) The dimensionless foreground to background ratio of seizure power output by an SDA (Osorio et al. 1998) applied to the data in (A). Inset shows the output ratio (on a log scale) from 5 min before to 5 min after seizure onset. (C) The corresponding indicator functions of being in a state of seizure ($$I_{Sz}(t)$$) and of the SDA output being in a state of detection ($$I_{SDA}(t)$$). Figure 2: The output of the SDA shown in Figure 1 over a 25 hour period (five panels of 5 hr each). SDA detections detections are marked with red “+” below. All detections were "true positives."

The term “Seizure Detection” differs from “Seizure Prediction” (see, e.g., Mormann 2008). In the case of detection, the SDA is not expected to identify the presence of a seizure until after it has begun “electrographically” (i.e., characteristics of the seizure have appeared in the biologic signal(s) being monitored). In the case of prediction, the algorithm is expected to forecast impending seizures before they start. Of course, this distinction can become muddled when there is uncertainty about when the seizure starts and if the prediction algorithm is based on detecting a signal pattern that is consistently followed by a seizure. The latter may indicate that the seizure has actually already begun by the time the detected pattern appears. In this case the “prediction” method may be actually just “detecting” the seizure before its (subjectively scored) seizure onset.

## Motivation and Applications

Approximately 1-2% of the population suffers from seizures. The unpredictability of when seizures will suddenly occur is a primary reason for the disability associated with epilepsy, and this uncertainty dramatically impacts the quality of life for patients and their caregivers (Fisher et al. 2000). Despite most persons with epilepsy showing no effects of the disorder other than during and immediately following the occurrence of seizures, not knowing when these events may happen can prevent them from driving, swimming, cooking, etc. Removing this uncertainty, through sufficiently early and accurate seizure detection and immediate warning, may thus result in a significant quality of life improvement. Accurate detection and logging of seizures can be used to improve the diagnostic yield from patient monitoring during epilepsy surgery evaluation and to improve understanding of epilepsy as a dynamical disease. Additionally, accurate automatic seizure detection offers the potential for automated “closed-loop” therapy, in which a therapy such as electrical stimulation, drug infusion, cooling, or biofeedback may be delivered in response to a seizure detection (Osorio et al. 2001, Theodore and Fisher 2004, Osorio et al. 2005, Morrell 2006, Anderson et al, 2008, Stacey and Litt 2008, Rothman 2008, Osorio and Frei 2009). Closed-loop control of seizures has major advantages over therapy that does not utilize immediate “feedback” of the patient state. In particular, closed-loop therapy can be precisely timed (e.g., delivered immediately upon detection of a seizure) and even adaptively dosed (e.g., delivering different dose levels and using different therapy modalities and delivery sites depending upon measured seizure characteristics and patient states). In other words, the therapy can be tailored to be administered only when and where needed. Coupling the use of quantitative monitoring algorithms with therapy also enables the objective assessment of therapeutic efficacy, by enabling characterization of therapy dose-response and correlating administration with changes in seizure severity and frequency of occurrence over time.

## Algorithms for Seizure Detection

There are numerous seizure detection algorithms described in the literature. The most prominent early attempt at automated seizure detection was made by Gotman (Gotman 1982) which built upon earlier work that he and others had done that attempted to quantify EEG transients/spikes (see, e.g., Gotman and Gloor 1976, Ives et al. 1976) and nonstationarities (Lopes da Silva 1975). Today most algorithms for seizure detection are based upon moving-window analysis of electrical signals recorded from the scalp (the electroencephalogram or “EEG”) or directly from the brain (electrocorticogram or “ECoG” if from the cortex, otherwise intracranial EEG or “iEEG”)(see, e.g., Osorio et al. 1998 and references therein), though many other signals may also be used, including cardiac-based (Marshall et al. 1983, O’Donovan et al. 1996, Frei et al. 1996), chemical (Crick et al. 2007) and motion-related (Nijsen et al. 2005) signals. In each window, one or more quantifying measures are computed from the data and changes in their values are analyzed as a function of time. Commonly used measures for EEG quantification include (i) amplitude and/or signal power, often restricted to a particular frequency band (or weighted as a function of frequency) via application of a filter to the signals, (ii) frequency changes in the signal, (iii) phase variable changes, (iv) rhythmicity changes, and (v) a measure of distance between the signal segment and a template signal with known morphology. These quantities may also be combined to derive other measures, such as signal arc-length or line-length measures (Esteller et al. 2001), or measures of similarity between the power spectral densities obtained from different signal epochs (Murro et al. 1991, Alarcon et al. 1995, Gabor and Seyal 1996). Depending upon whether a measure utilizes a one, two, or several input signals to produce its output, the measure is referred to as a univariate, bivariate or multivariate measure, respectively. Differences between measures in the most recent moving window(s) are typically compared to reference or background values to identify statistically significant changes associated with the seizure activity. Often the ratio between the recent moving window (“foreground”) and past non-seizure values (“background”) is compared to a threshold in order to detect significant changes (Osorio et al. 1998).

## Some of the Challenges in Seizure Detection

There are several difficult challenges associated with the development of a seizure detection algorithm (“SDA”), including:

### What is a seizure and when to detect it?

Figure 3: (A) 60 s of single channel ECoG signal containing a seizure and annotations of 3 different patterns that could be used to train or adapt an SDA. The first pattern (“seg 1”) is a quasi-periodic train of spike and slow waves (SSW) with approximately one wave every 2 s; the second pattern (“seg 2”) consists of more frequent SSWs occurring approximately 2/s; and the third pattern (“seg 3”) is a higher frequency rhythmic discharge detected by the generic SDA in (Osorio et al. 1998). (B) The “detection ratio” outputs of three SDAs, each adapted to detect one of the three annotated patterns identified in (A).
Figure 4: When the detector adapted to detect the “seg 1” pattern is applied to a longer recording, the first 30 detections are shown (i.e., 30 single-channel ECoG traces, from successive detected events, with detection occurring for each trace at the right margin of the figure). The reader is challenged to consider each trace and see if they can determine differences that might foreshadow whether or not the signal will further evolve into the high frequency rhythmic discharge pattern (as in “seg 3” of Figure 3) often associated with unequivocal electrographic seizures. (See Figure 5 for the answer.) A further important challenge, typically requested of a SDA, is to determine by this point in time whether or not the signal change is going to evolve into a seizure with clinical/behavioral manifestations.
Figure 5: The same data as shown in Figure 4 except the 60 s of ECoG following each detection is also shown. Annotations in the right margin indicate the 14 events which evolved to produce generic SDA detections of high frequency rhythmic activity. For this subject, during 85.6 hr of continuous monitoring, 36% of such SSW train detections evolved to generic detections of high frequency rhythmic activity and 18% of those generic SDA detections evolved to become seizures with known clinical/behavioral manifestations. On the other hand, every known clinical seizure was preceded by a generic SDA detection and every such detection was preceded even earlier by a SSW train. This observation calls into question what exactly constitutes “a seizure” and when exactly one begins.

Seizures are a symptom associated with abnormal electrical activity in the brain, sometimes described as an electrical storm in the brain or earthquake in the brain (Osorio et al. 2010). Unfortunately, answers to the "what is a seizure?" and "when to detect it?" questions remain elusive since there is currently no consensus or objective definition of what constitutes a seizure. This significantly hinders the perfection of seizure detection, as it is obviously difficult to develop an algorithm to detect events with perfect precision when the events themselves aren’t clearly defined.

Even when the EEG signal contains an unequivocal seizure, expert reviewers sometimes differ significantly (by tens of seconds or more) in their respective markings of the seizure’s electrographic onset (EO) and electrographic end (EE) times. This inter-rater variability in EO and EE markings complicates SDA development and performance assessment, since most applications desire rapid detection of the signal changes that immediately follow EO. Depending upon when the EO is marked, the signal change immediately following might, for example, comprise a train of spikes, a rhythmic high frequency oscillation, or some other type of signal change (see, e.g., Figure 3). Thus one can appreciate the importance of an adaptive SDA that can be tailored to detect virtually any type of relevant signal change, depending on what the electroencephalographer/epileptologist/user desires to detect or what is needed for the particular application.

A significant limitation in comparing expert visual analysis (EVA) scores with the results of an SDA is that EVA is typically retrospective – the electroencephalographer identifies that a seizure has occurred and then pages forward and backward through the signal display/printout to set the times at which they believe the event starts and ends. On the other hand, any SDA that operates in real-time is required to issue a decision at a particular time using only signal information available up to that point in time. While some SDAs have been developed that perform retrospectively to first determine a seizure’s presence and then back up to determine when it started (Chan et al. 2008), and these approaches have some utility for offline processing, the associated detection delay limits their usefulness for warning or closed-loop control applications.

Another difficulty in determining “what to detect and when” arises because in certain applications the user may not want to detect ‘’all’’ seizures. In seizure warning applications, for example, it may be desirable to only detect, or at least only warn of, the seizures that have clinical/behavioral manifestations or some loss of normal function. Classifying seizures as either “clinical” or “subclinical,” depending upon whether behavioral manifestations are present or absent, is an important objective but can be difficult to accomplish with a good degree of accuracy. Patients are often unaware that they have had a seizure, there isn’t always an expert observer present and cognitive/functional testing during seizures to assess potential impairment is rarely administered (Osorio and Frei 2010). Given these factors, a more proper terminology for seizure classification divides events into: (i) known clinical seizures (“KCSz”) and (ii) seizures not known to be clinical (“NKCSz”).

Practical strategies for seizure detection algorithm development, in the face of the aforementioned uncertainties, generally fall into two groups: (1) develop an SDA that detects certain specific signal characteristics that a group of seizures have in common (e.g., the earliest signal pattern common to all KCSz), or (2) develop multi-faceted SDAs which attempt to detect a variety of abnormal epileptiform activity present in brain signals, including not just unequivocal seizures but other relevant activity such as brief seizures, bursts, spike trains, and even single spikes, then correlate the occurrence of these detections with clinical events of interest in hopes of providing the most complete information possible about the brain system dynamics underlying seizures for that patient. The first approach has limitations in that seizures lacking the target signal characteristic set may go undetected, while the second approach tends to detect both artifacts and other paroxysmal events (both epileptic and non-epileptic) that may not generally be considered to be seizures.

A challenge in SDA development is to detect the earliest signal change that will evolve with high likelihood in a clinical seizure. However, since onset patterns of “subclinical seizures” (i.e., the electrographic seizures that aren’t known to be clinical) are often very similar to those of clinical seizures, the task involves a speed vs. accuracy trade-off: How far should the SDA let a seizure evolve before making a detection decision? The relative frequency of clinical vs. subclinical events, the negative consequences of improper classification (which may depend, e.g., upon the severity of clinical symptoms or side-effects of a detection-triggered therapy) and the patient-specific signal patterns are all important factors in deciding when to detect. Figure 3-Figure 5 provides an illustrative example. Figure 9-Figure 11 at the end of this article gives another such example for the interested reader.

### The trade-off between SDA accuracy vs. detection speed

In seizure warning and automated delivery of therapy applications, it is desirable for any SDA to detect seizures as early as possible. However, this goal obviously conflicts with the desire for highly accurate detections, since the more signal the algorithm can “see” to analyze, the more information it has at its disposal for use in improving detection accuracy but the later the resulting detection will be. The two examples in Figure 3-Figure 11 illustrate this challenge. Speed and accuracy may both be improved by improving signal acquisition, such as by placing electrodes inside the brain and nearer to the location where the seizures are starting, or by utilizing multiple sensors to help separate signal changes of interest from artifacts. However, such improvements may come at the expense of increased invasiveness of monitoring.

### Dealing with non-stationarity and noise

Figure 6: Samples of EEG/ECoG containing various common artifacts: (A) clipping due to amplifier saturation, (B) 60 Hz “Mains” frequency electrical noise, (C) inductance/movement artifacts, (D) muscle artifact, and (E) ocular artifact due to eye movements.

The EEG, especially when recorded from the scalp, is prone to many different artifacts which obstruct the view of underlying brain activity. Figure 6 shows some of these common artifacts. Most SDAs are based, at least in part, on detecting changes in the distribution of signal energy as a function of time and frequency (time-frequency-energy or “TFE” analysis). One of the difficulties posed by artifacts/noise in the signal is that it can often overlap in the frequency domain with signals of interest (seizures). Also, the artifacts often come and go with time, as do seizures, making the degree of overlap at any given moment somewhat unpredictable. The underlying brain signals and their characteristics also change with time and state of the patient (i.e., they are non-stationary), which can pose a problem for certain popular conventional analysis techniques. Fourier analysis is perhaps the most common TFE analysis technique, but it has an underlying assumption that the signals being analyzed are stationary, which limits its utility for this application. Over the past two decades, wavelet-based TFE analysis and other related methods have become popular tools for use with EEG and other non-stationary signals, in part because their multi-resolution approach to TFE analysis deals with inherent non-stationarity in the signal more successfully (Schiff 1994, Osorio et al. 1998, Jouny et al. 2010). New methods for TFE analysis of non-stationary signals, such as Intrinsic Timescale Decomposition (ITD) (Frei and Osorio 2007) have recently provided additional tools to decompose these and other complex signals into components of interest, enabling separation of seizure/epileptiform signal components from background/normal components.

### Minimizing computational cost

The processors that are approved for use in today's implantable medical devices are much more limited in their processing capability than desktop computers or those used in other "non-implanted applications." Capabilities for implantable devices, including those currently under development, are 4-5 orders of magnitude lower than present-day computers, functioning on a level that is approximately equivalent to the IBM 286 PCs of the late 1970s to early 1980s (Werder 2007). Reasons for this include the cost of gaining regulatory approval for state of the art processors and the manufacturing constraints imposed by regulatory agencies. This imposes significant limitations upon the computational complexity of algorithms that may be used in seizure warning and closed-loop therapy devices. Despite the fact that processing power required for an SDA typically increases at least linearly with the sampling rate and with the number of signals being monitored, current trends toward use of very high sampling rates (e.g., 10 kHz or more) and micro-electrodes, which can require well over a hundred in number to cover even small regions of the brain, continue in hopes that these approaches will provide useful information that cannot otherwise be obtained. Other research has focused on developing SDAs which retain desirable characteristics of the most promising existing algorithms but require less computational power (Bhavaraju 2006, Raghunathan et al. 2009, Raghunathan et al. 2010).

### Use of a generic vs. an individualized SDA

There is a great deal of inter-subject variability in seizure characteristics, which makes it difficult for any prospective algorithm to detect everything. Thus far, there is no “one size fits all” SDA that will detect every seizure for every subject with any reasonable degree of accuracy. However, the frequent lack of a priori information about a subject’s seizures and need for a reasonably broadly applicable approach to seizure detection has led to the development of generic SDAs, which may be group-optimized in their parameter settings to detect as large a set of seizures as possible (Osorio et al. 1998). Yet, the stereotypic nature of seizures within a given subject affords significant improvement when generic SDAs are tailored or adapted to the make use of an individual’s or group's seizure fingerprint(s). Flexible SDAs, along with strategies for automating individualized adaptation of their parameters that allow efficient physician supervision (e.g., providing relevant patient-specific information, including samples of seizures one wishes to detect, less intense seizures one wishes to avoid detecting, and typical variations observed in non-seizure data), are likely to outperform a generic approach (Frei et al. 2002, Qu and Gotman 1993, Shoeb et al. 2004, Qu and Gotman 1995, Wilson 2006, Haas et al. 2007).

### SDA Performance Assessment

SDA performance is conventionally assessed using sensitivity, specificity and speed of detection in comparison to the “gold standard” of expert visual analysis of EEG and patient video by a trained electroencephalographer/epileptologist. The inter-relationship between sensitivity and specificity is often analyzed using Receiver Operating Characteristic (ROC) curves. However, the results obtained/reported using each of these standard statistics must be critically and carefully interpreted to be meaningful and are far from complete in terms of describing SDA performance. This may be illustrated with a few simple examples (see Figure 7-Figure 8). Extensive prospective validation of any SDA is required to properly assess its performance (see, e.g., Osorio et al. 2001).

 Figure 7: Example illustrating some of the difficulties in interpreting sensitivity and specificity metrics. Suppose a patient had 100 seizures during a study and every one of them was immediately preceded by an SDA detection, but each detection began 5 s before EO and ended 1 s before EO. Then although the SDA would actually be accurately predicting 100% of seizures and doing so exactly 5 s prior to each EO, each detection could be considered as an FP because they occur outside the expert-scored EO-to-EE intervals. Each seizure could also be considered to have been missed (i.e., scored as a false negative or FN) for the same reason. As a comparison, suppose instead that the algorithm detected each seizure right near its end, (e.g., 1 s before its scored end time (EE)) and had no other detections during the study. In this case, the SDA would have perfect sensitivity and perfect specificity, despite having an average detection delay from EO almost as large as the average seizure duration. The absurd fact that the first SDA, for which every seizure is missed and every detection is an FP actually outperforms the second SDA that had perfect sensitivity and specificity, illustrates the care with which these statistics must be interpreted. Figure 8: Examples illustrating why direct comparisons of (#FP/hr) reports can be misleading. All three plots show the indicator function of whether or not an SDA output, denoted by $$I_{SDA}(t)\ ,$$ is in a state of detection ($$I_{SDA}(t)=1$$) as a function of time of day. (A) An SDA with 100 FP/24 hr, but all the detections occur within a single 10 s period. This scenario can happen, e.g., when a feature is rapidly alternating above and below a detection threshold. (B) The same SDA output as shown in (A), but zoomed in to a 12 s window during which the FPs occur. If the multiple individual detections were considered together as a single “grouped FP” (or “FP cluster”), the FP rate would be lowered to 1.0 FP/24 hr. (C) An example of an SDA output with "only" 1.0 FP/24 hr, yet 90% of the time monitored is spent in a state of false detection. This scenario can happen, e.g., in the case of a detection threshold that is set much too low, or a broken sensor producing nearly continuous artifact that is being erroneously detected.

## Summary

Accurate automated seizure detection remains an important challenge and a critical first step in (1) removing the uncertainty associated with when epileptic seizures will occur, (2) furthering our understanding of seizures and what may cause them and (3) enabling the development of systems for automated “closed-loop” therapy intended to terminate seizures as they begin or even prevent their occurrence altogether.

## Other Resources

Publicly Available Databases: Recently several databases of biological signal data (e.g., EEG, ECoG or EKG) containing seizures have been made publicly available. The use of common datasets is essential in comparing performance between different detection algorithms and to properly assess prospective performance on data that are not used in developing the algorithm parameters. Some such databases can be found at:

Some Existing Commercially-Available Seizure Detection Algorithms:

## Glossary of Terms

• Electrographic Onset (EO) – The time at which the seizure is marked to have begun through expert visual analysis of the EEG/ECoG.
• Electrographic End (EE) – The time at which the seizure is marked to have ended through expert visual analysis of the EEG/ECoG.
• Clinical Onset (CO) – The time at which an expert observer, or an observer familiar with the patient’s seizures first identifies a behavioral manifestation of a seizure.
• Known Clinical Seizure (KCSz) – A seizure which is known to have clinical manifestations (where clinical/behavioral manifestations are determined, e.g., by behavioral observation or responsiveness testing).
• Seizure Not Known to be Clinical (NKCSz) – Any electrographic seizure which is not known to have clinical/behavioral manifestations.
• False Positive (FP) – The occurrence of a seizure detection at a time when no seizure is occurring.
• False Negative (FN) – The failure to detect a seizure that has occurred.
• True Positive (TP) – The detection of a seizure that has occurred.
• True Negative (TN) – Correctly being in the state of non-detection during a time when a seizure is not occurring.

• Sensitivity – A measure of how likely an epoch during a seizure is to be detected. The classical definition, for a binary classifier, is given by$Sensitivity = \frac{\mbox{Number of True Positives}}{\mbox{Number of True Positives} + \mbox{Number of False Negatives}}$
• Specificity - A measure of how likely an epoch during a non-seizure epoch is to be detected. The classical definition, for a binary classifier, is given by$Specificity = \frac{\mbox{Number of True Negatives}}{\mbox{Number of True Negatives} + \mbox{Number of False Positives}}$

## References

• Alarcon G, Binnie CD, Elwes RD, Polkey CE. Power spectrum and intracranial EEG patterns at seizure onset in partial epilepsy. Electroencephalogr Clin Neurophysiol 1995;94:326–37.
• Anderson WS, Kossoff EH, Bergey GK, Jallo GI. Implantation of a responsive neurostimulator device in patients with refractory epilepsy. Neurosurg Focus. 2008 Sep;25(3):E12.
• Bhavaraju NC, Frei MG, Osorio I. Analog seizure detection and performance evaluation. IEEE Trans Biomed Eng. 2006 Feb;53(2):238-45.
• Chan AM, Sun FT, Boto EH, Wingeier BM. Automated seizure onset detection for accurate onset time determination in intracranial EEG. Clin Neurophysiol. 2008 Dec;119(12):2687-96.
• Crick EW, Osorio I, Bhavaraju NC, Linz TH, Lunte CE. An investigation into the pharmacokinetics of 3-mercaptopropionic acid and development of a steady-state chemical seizure model using in vivo microdialysis and electrophysiological monitoring. Epilepsy Res. 2007 May;74(2-3):116-25.
• Esteller R, Echauz J, Tcheng T, Litt BA, Pless BA. Line length: an efficient feature for seizure onset detection. In: Echauz J, editor. Engineering in Medicine and Biology Society, 2001. Proceedings of the 23rd annual international conference of the IEEE, vol. 1702; 2001; 2001. p. 1707–10.
• Fisher RS, Vickrey BG, Gibson P, Hermann B, Penovich P, Scherer A. The impact of epilepsy from the patient's perspective I. Descriptions and subjective perceptions. Epilepsy Res 2000; 41: 39-51.
• Frei MG, Osorio I, Davidchack R. A reappraisal of the value of EKG in the detection of epileptic seizures. Abstract; Epilepsia. 1996:37(S5).
• Frei MG, Haas SM, Osorio I. Adaptation of a Real-Time Seizure Detection Algorithm. B. Pasik-Duncan: Stochastic Theory and Control, LNCIS 280, pp.131-136, 2002.
• Frei MG, Osorio I. Intrinsic time-scale decomposition: time-frequency-energy analysis and real-time filtering of non-stationary signals. Proc. R. Soc. A (2007) 463, 321-342. doi:10.1098/rspa.2006.1761.
• Frei MG, Zaveri HP, Arthurs S, Bergey GK, Jouny CC, Lehnertz K, Gotman J, Osorio I, Netoff TI, Freeman WJ, Jefferys J, Worrell G, Quyen Mle V, Schiff SJ, Mormann F. Controversies in epilepsy: debates held during the Fourth International Workshop on Seizure Prediction. Epilepsy Behav. 2010 Sep;19(1):4-16. Epub 2010 Aug 13.
• Gotman J and Gloor P. Automatic recognition and quantification of interictal epileptic activity in the human scalp EEG. Electroencephalography and Clinical Neurophysiology, 41(5):513–29, 1976.
• Gotman J. Automatic recognition of epileptic seizures in the EEG. Electroencephalography and Clinical Neurophysiology, Nov;54(5):530–40, 1982.
• Haas SM, Frei MG, Osorio I. Strategies for adapting automated seizure detection algorithms. Med Eng Phys. 29:895-909(2007).
• Ives JR, Thompson CJ, Gloor P. Seizure monitoring: a new tool in electroencephalography. Electroenceph. Clin. Neurophysiol., 1976, 41:422-427.
• Jouny CC, Bergey GK, Franaszczuk PJ. Partial seizures are associated with early increases in signal complexity. Clin Neurophysiol. 2010 Jan;121(1):7-13.
• Lopes da Silva FH, Dijk A, Smits H. Detection of nonstationarities in EEGs using the autoregressive model – an application to EEGs of epileptics. In H. Kunkel and G. Dolce (Eds.), CEAN – Computerized EEG analysis. Gustav Fischer Verlag, Stuttgart, 1975.
• Marshall DW, Westmoreland BF, Sharbrough FW. Ictal tachycardia during temporal lobe seizures. Mayo Clin Proc. 1983 Jul;58(7):443-6.
• Morrell M. Brain stimulation for epilepsy: can scheduled or responsive neurostimulation stop seizures? Curr. Opin. Neurol 2006; 19: 164-8.
• Murro AM, King DW, Smith JR, Gallagher BB, Flanigin HF, Meador K. Computerized seizure detection of complex partial seizures. Electroencephalography and Clinical Neurophysiology, Oct;79(4): 330–3, 1991.
• Nijsen TM, Arends JB, Griep PA, Cluitmans PJ. The potential value of three-dimensional accelerometry for detection of motor seizures in severe epilepsy. Epilepsy Behav. 2005 Aug;7(1):74-84.
• O'Donovan C, Geller EB, Burgess R, Lüders H. Ictal Tachycardia During Partial Seizures Independent of Mesial Temporal Involvement. Epilepsia 37[5], 82. 1996.
• Osorio I, Frei MG, Wilkinson SB. Real time automated detection and quantitative analysis of seizures and short term predictions of clinical onset. Epilepsia. 1998;39(S16):615-627.
• Osorio I, Frei MG, Manly BFJ, Sunderam S, Bhavaraju NC, Wilkinson SB. An introduction to contingent (closed-loop) brain electrical stimulation for seizure blockage, to ultra-short term clinical trials, and to multidimensional statistical analysis of therapeutic efficacy. J Clin Neurophysiol. 2001 Nov;18(6):533-44.
• Osorio I, Frei MG, Giftakis J, Peters T, Ingram J, Turnbull M, Herzog M, Rise M, Schaffner S, Wennberg R, Walczak T, Risinger M, Ajomone-Marsan C. Performance re-assessment of a real-time seizure detection algorithm on long ECoG series. Epilepsia. 2002 Dec;43(12):1522-35.
• Osorio I, Frei MG, Sunderam S, Giftakis J, Bhavaraju NC, Schaffner SF, Wilkinson SB. Automated seizure abatement in humans using electrical stimulation. Ann Neur. 2005 Jan 24;57(2):258-68.
• Osorio I, Frei MG. Comments on requirements for valid and meaningful assessment of therapies for epilepsy and a note of clarification about "open-" vs. "closed-loop" modalities. Epilepsia. 2007 Feb;48(2):406.
• Osorio I, Frei MG. Seizure abatement with single dc pulses: is phase resetting at play? Int J Neural Syst. 2009 Jun;19(3):149-56.
• Osorio I, Frei MG, Sornette D, Milton J, Lai YC. Epileptic seizures: Quakes of the brain? Phys Rev E Stat Nonlin Soft Matter Phys. 2010 Aug;82(2 Pt 1):021919.
• Osorio I, Frei M. Feasibility of automated warning in subjects with localization-related epilepsies. Epilepsy Behav. 2010 Oct 26 [Epub ahead of print] doi:10.1016/j.yebeh.2010.09.021
• Qu H, Gotman J. Improvement in seizure detection performance by automatic adaptation to the EEG of each patient. Electroencephalogr Clin Neurophysiol. 1993 Feb;86(2):79-87.
• Qu H, Gotman J. A patient-specific algorithm for the detection of seizure onset in long-term EEG monitoring: Possible use as a warning device. IEEE Transactions of Biomedical Engineering, 44(2):115–22, 1997.
• Raghunathan S, Gupta SK, Ward MP, Worth RM, Roy K, Irazoqui PP. The design and hardware implementation of a low-power real-time seizure detection algorithm. J Neural Eng. 2009 Oct;6(5):056005.
• Raghunathan S, Gupta SK, Markandeya HS, Roy K, Irazoqui PP. A hardware-algorithm co-design approach to optimize seizure detection algorithms for implantable applications. J Neurosci Methods. 2010 Oct 30;193(1):106-17. Epub 2010 Aug 14
• Rothman, SM. Beyond Prediction – Focal Cooling and Optical Activation to Terminate Focal Seizures pp. 269-282; in B. Schelter, J. Timmer, A. Schulze-Bonhage (Eds.), 2008 Seizure Prediction in Epilepsy: From Basic Mechanisms to Clinical Applications, Wiley-VCH, Berlin.
• Schiff SJ, Aldroubi A, Unser M, Sato S. Fast wavelet transformation of EEG. Electroencephalogr Clin Neurophysiol. 1994 Dec;91(6):442-55.
• Shoeb A, Edwards H, Connolly J, Bourgeois B, Treves ST, Guttag J. Patient-specific seizure onset detection. Epilepsy and Behavior, 5:483–98, 2004.
• Stacey WC, Litt B. Technology insight: neuroengineering and epilepsy-designing devices for seizure control. Nat Clin Pract Neurol 2008; 4: 190-201.
• Theodore WH, Fisher RS. Brain stimulation for epilepsy. Lancet Neurology 2004; 3: 111-8.
• Werder J. Computing Constraints of an Implanted Device Compared to a Traditional Computer. Presentation at the Third International Workshop on Seizure Prediction, Freiburg, Germany. September 30, 2007.
• Wilson SB. Algorithm architectures for patient dependent seizure detection. Clinical Neurophysiology, 117(6):1204–16, 2006.

Internal References

• Mormann, F. (2008) Seizure Prediction. Scholarpedia, 3(10):5770.

• Lesser RP, Webber WR. Seizure detection: reaching through the looking glass. Clin Neurophysiol. 2008 Dec;119(12):2667-8.
• Sunderam, S. How to Detect and Quantify Epileptic Seizures, in Epilepsy: The Intersection of Neurosciences, Biology, Mathematics, Engineering and Physics, Osorio I., Zaveri H, Frei MG,  Arthurs S (eds). CRC Press 2011.
 Figure 9: (A) 20 s of single channel ECoG signal containing a seizure. (B) Magnification of the seizure signal around the electrographic onset. This signal pattern is used to adapt/train an SDA to rapidly detect events with similar power spectral density sustained for at least 2 s. Figure 10: The adapted SDA, trained using the epoch in Figure 9(B), is applied to a multiday ECoG recording from the same subject as in Figure 9 and 15 s traces are shown for the first 50 resulting detections. All detections have similar signal characteristics in the two-second window immediately preceding the detections. The reader is challenged to (prospectively) guess which of these detections will evolve to electrographic seizures that last at least 10 s. (See Figure 11 for the answer.)
Figure 11: The same data as shown in Figure 10 except the 15 s of ECoG following each detection is also shown. Only 8 of the 50 detections correspond to electrographic seizure events that last at least 10 s.