The electroencephalogram (EEG) provides a real-time measure of brain electrical activity and has proven to be useful in a wide range of clinical settings, for example, evaluation of seizures, mental status change, coma, and classification of epilepsy.^1,2 The clinical applications of continuous EEG monitoring can involve recordings from scalp or intracranial electrodes obtained in the hospital (e.g., video scalp and intracranial EEG [IEEG] monitoring), on an outpatient basis (routine scalp EEG), and even in nonmedical settings (ambulatory EEG). More recently, a cranially implanted device (the RNS system, NeuroPace Inc.) using continuous IEEG for real-time seizure detection and electrical stimulation to abort seizures is undergoing clinical trial.³ This broad range of clinical applications underscores the usefulness of EEG for studying brain dysfunction and highlights the potential applications of automated EEG analysis. In this chapter, we give an overview of automated detection of epileptiform activity. The primary goal is to discuss epileptiform event detection in the context of modern data mining and pattern recognition.^4–6 The chapter is not intended to be a comprehensive review of epileptiform spike and seizure detection.^7–10

The quantity of data generated by continuous EEG recording can be significant. Historically, the analysis and storage of these data has been challenging due to relatively limited computational resources. These challenges persist today, despite the ubiquity of cheap gigaflop servers and terabyte storage systems, for a different reason: the established clinical work flow for EEG analysis breaks down for modern data sets. Current state-of-the-art recording systems now acquire hundreds of channels of EEG data at high sampling rates. These large-scale data sets can be thousands of times larger than traditionally acquired clinical EEG, which is nominally collected at 200 to 500 Hz. Although the hardware aspect of collecting, analyzing, and storing these data is feasible, the “gold standard” of analysis by human expert visual review of entire records is not. The sheer increase in data for modern recordings places a significant strain on epilepsy monitoring unit (EMU) staff and resources for all but the simplest of analyses. Moreover, the hunt for multiscale events ranging in duration from milliseconds to minutes requires more detailed and repeated review of EEG. The need for reliable automated analysis tools in clinical and research electrophysiology is clearly recognized, and there is a growing community of researchers pursuing these goals. The field should certainly benefit from parallel efforts in other clinical fields, such as genomics, that are driving efforts in database management, data mining, and pattern recognition.⁴

Without automated analyses tools, clinical video-EEG evaluation continues to rely primarily on human expert visual review. Video-EEG monitoring (VEM) has long been a cornerstone in the evaluation of patients with seizure disorders.^1,2 Generally, the primary goal of long-term VEM is to record the patient’s habitual seizures. However, even when seizures are not captured, the long-term interictal recording can be useful by recording interictal epileptiform activity. In particular, epileptiform transients such as spikes, sharp waves, and temporal intermittent rhythmic delta activity recorded with scalp EEG support a diagnosis of epilepsy.¹ These interictal recordings also can be useful for quantifying the spatial and temporal distribution of interictal epileptiform transients. For example, in the evaluation of patients for temporal lobe epilepsy surgery, there is evidence that the frequency of spikes (number of spikes/min) ipsilateral to seizure onset is associated with surgical outcome.¹¹ The presence of spikes and seizures contralateral to any planned surgery are important.¹² Similarly, there is accumulating evidence that the spatial distribution of IEEG spikes¹³ and high-frequency oscillations¹⁴ has prognostic implications in neocortical epilepsy surgery.

Recent advances in neurophysiological data acquisition systems now enable recording over the physiological range of human brain activity.^15–18 Wide bandwidth EEG recordings (0.01–1000 Hz) show that interictal high-frequency oscillations are signatures of an epileptic brain and may have greater specificity than interictal spikes for the region of the brain that must resected for seizure freedom.^14,18 However, high-frequency oscillations are relatively short duration events (50–250 msec) that are difficult to visualize with standard clinical EEG viewing parameters. The development of automated high-frequency oscillation detectors is an area of active research.¹⁹ Recently, high-frequency epileptiform transients, fast ripples,¹⁵ and pathologic gamma and ripple oscillations¹⁴ have been described in EEG recordings as specific interictal biomarkers of an epileptic brain.

The development of EEG analyses now indicates that the entire EEG record, not only seizures, contains useful diagnostic information. However, identifying epileptiform transients in continuous multiday VEM recordings presents a practical challenge for manual visual review. In fact, even electrographic and subtle clinical seizures are easily missed in long-term recordings that rely on expert visual review. Of course, this fact is compounded in busy EMUs where multiple patients are monitored continuously. Therefore, mining large EEG data sets to obtain quantitative information for ictal and interictal electrographic events is clinically relevant. As discussed above, this includes data mining for interictal spikes, high-frequency oscillations, subclinical electrographic seizures, and clinical seizures. Although it is generally assumed that human experts, that is, electroencephalographers, are the gold standard for marking interictal epileptiform activity and seizures, it is not feasible to routinely label entire EEG records using manual review. Detailed labeling of the occurrence of epileptiform activity in multiday long-term EEG recordings necessitates automated detectors.

Over the past decade, more sophisticated analyses have been applied to the problem of automated EEG event detection and include techniques from many areas of signal processing (e.g., filtering, wavelets, and time–frequency analyses) and machine learning (feature selection, manifold learning, and modern classifiers, including neural networks and support machines). In addition, more recently, the literature makes clear how difficult it is to accurately detect interictal spikes and seizures.^8,20 Similarly, recent studies have shown the significant challenges for automated detection of high-frequency oscillations in wide bandwidth EEG recordings.¹⁹ The robust detection of these classic events remains a very active area of research.

The detection and labeling of interictal and ictal epileptiform activity in EEG records (Figure 3-1) can be broadly categorized into three different approaches:

Manual Expert EEG Review

• 
  

  Considered the gold standard

  
  
• 
  

  Not feasible for large-scale data sets

  
  
• 
  

  More variability than is widely acknowledged

High-Sensitivity Automated Detection Combined with Expert Review

• 
  

  Primary approach for automated analysis

  
  
• 
  

  Can significantly reduce data to review

  
  
• 
  

  Does not provide detector specificity

Fully Automated Detection and Labeling of Epileptiform Events

• 
  

  Requires high specificity and sensitive detectors

  
  
• 
  

  Efficient approach if detectors can be realized

  
  
• 
  

  Required for real-time device applications

Manual Expert EEG Review

This is generally considered to be the gold standard for detecting and labeling epileptiform activity in EEG records. It is labor intensive and except for relatively small “toy” data sets, not routinely feasible. In addition, the reliability and performance of expert manual review have been shown to be inconsistent. A commonly overlooked fact is that manual expert EEG review itself suffers from a number of problems. The repeatability of EEGers reading large records is inexact, and agreement between reviewers can vary drastically. The challenge for human reviewers, and machines, is greater for brief interictal transient events¹⁹ compared to seizure detection.¹⁰ There have been few attempts in the literature to address the reliability of the expert review. Wilson and Emerson described a perception threshold by which not all spikes are graded equally, and indeed found that there are spikes agreed on by reviewers, and others that are not.⁸ The perception score is akin to Likert scales²¹ to capture the degree of epileptiform “eventness.”

Although these tools from statistical data mining are useful, they do not obviate the fact that what is used as the gold standard is not foolproof. Of course, the above difficulties are compounded by large data sets generated by long-term recording at high sample rates from a large number of channels.

Nonetheless, every effort must be made to obtain high-fidelity gold standard data sets, which are critically important for the development of automated detectors. As described below, the rigorous development of spike and seizure detectors ultimately requires independent gold standard or “ground truth” data for training and evaluating the performance of detectors.

Highly Sensitive Automated Detection

This common approach involves using automated detection followed by expert visual review (Figure 3-2).^22,23 One advantage is that a detector is not required to exhibit high specificity, but rather relies on high sensitivity to identify all events of interest (along with a large number of false-positive detections to be pruned by subsequent manual visual review). The compromise between sensitivity and specificity is selected to provide a highly sensitive detector at the expense of specificity, that is, great accuracy accompanied by a high number of false positives. This can be interpreted within the context of receiver operating characteristic (ROC) curves (Figure 3-3). By selecting a threshold with high sensitivity, the automated detector identifies epileptiform activity at the expense of specificity and false-positive detections. The approach relies on subsequent manual expert review to eliminate false-positive events. In certain contexts, this approach can be viewed as a semisupervised algorithm. Although the approach still relies on expert review, the volume of data that have to be manually reviewed is often dramatically collapsed (by 100 times or more). The EEG is reduced to a series of candidate event intervals, for example, a 5-second window for spikes, and the manual review consists only of binary labeling of these intervals (as “event” or “nonevent”). One notable drawback to this approach is that it does not provide a means to estimate the number of false-negative events. Often in the case of EEG analysis, however, a reasonable assumption is that the high sensitivity of the detector ensures that the number of false negatives is negligible.

Fully Automated Detection

Fully automated detection does not rely on expert visual review and therefore requires both highly sensitive and specific detection of epileptiform activity. In many applications, fully automated detection is not required. For example, in a retrospective review of large continuous video-EEG data sets from the EMU, the second approach (highly sensitive automated detection combined with expert review) is more appropriate.

An example of a clinical application that does require fully automated seizure detection is the concept of seizure detection and responsive stimulation.³ The advances that have occurred make it possible for an implantable device to perform real-time automated seizure detection from IEEG and to trigger electrical stimulation designed to abort seizures. Automated seizure detection using IEEG has not received a great deal of attention, and the true specificity and sensitivity of these detectors are not yet well established. The first-generation device offering responsive stimulation is the NeuroPace RNS system, which uses a seizure detector built around three IEEG features: line length, half-wave, and amplitude.²⁴

The detection of epileptiform spikes and seizures from scalp and IEEG recordings is a complicated task with a long history.^{7–9,20,24,25} During the past few decades, the problem of event detection in noisy signals has benefited from advances in pattern recognition, machine learning, and data mining, to name a few areas of research. Each of these areas is likely to continue to affect the long-standing clinical problem of spike and seizure detection.

The detection of a particular class of EEG events (e.g., seizures and spikes) can be cast into the general framework of modern pattern recognition, an interdisciplinary domain spanning computer science, engineering, neuroscience, and medicine.⁴ In the following section, we review EEG event detection within the context of pattern recognition. Most of the literature describing epileptiform spike or seizure detection can also be viewed within the following framework.

Signal Acquisition

The performance of an EEG pattern recognition system (i.e., epileptic spike and seizure detection) fundamentally depends on characteristics of the sensing system, for example, signal and recording bandwidth, resolution, signal-to-noise ratio, and electrode geometry and placement. In the field of EEG, there is now considerable interest in brain activity that is outside the historically clinical “standard” recording bandwidth (∼1–100 Hz). In many applications, as the fidelity of data improves, detector performance improves.⁴

Signal Preprocessing

Most detection algorithms initially process the data prior to analysis. Preprocessing operations commonly include filtering (e.g., bandpass filtering), denoising (e.g. removing 60 Hz line noise), sample rate conversion, signal detrending, and artifact detection. Additionally, more sophisticated and application-specific preprocessing my be required, for example, data whitening or equalization,²⁶ to normalize the signal amplitude, spectrum, or energy across the time interval of interest. Independent component analysis (ICA)^27,28 and principal component analysis (PCA)^29,30 have been shown to be useful for removal of noncerebral components such as muscle, eye blinks, and reference electrode contamination.