Application of Machine Learning To Epileptic Seizure Detection Ali Shoeb ashoeb@mit.edu John Guttag guttag@mit.edu Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, Massachusetts, 02139 Abstract We present and evaluate a machine learning approach to constructing patient-specific classifiers that detect the onset of an epileptic seizure through analysis of the scalp EEG, a non-invasive measure of the brain's electrical activity. This problem is challenging because the brain's electrical activity is composed of numerous classes with overlapping characteristics. The key steps involved in realizing a high performance algorithm included shaping the problem into an appropriate machine learning framework, and identifying the features critical to separating seizure from other types of brain activity. When trained on 2 or more seizures per patient and tested on 916 hours of continuous EEG from 24 patients, our algorithm detected 96% of 173 test seizures with a median detection delay of 3 seconds and a median false detection rate of 2 false detections per 24 hour period. We also provide information about how to download the CHB-MIT database, which contains the data used in this study. seizures. The most common way to infer the onset of a seizure before it becomes clinically manifest is through analysis of the scalp electroencephalogram (EEG), a noninvasive, multi-channel recording of the brain's electrical activity. The characteristics of EEG vary significantly across patients. In fact, EEG associated with seizure onset in one patient may closely resemble a benign pattern within the EEG of another patient. This cross-patient variability in seizure and non-seizure activity causes patient non-specific classifiers to exhibit poor accuracy or long delays in declaring the onset of a seizure (Wilson et al., 2004). In some cases, however, patient non-specific classifiers can exhibit impressive performance when restricted to analyzing seizure types that vary little across patients (Meier et al., 2008). This paper is about using machine learning to construct patient-specific detectors capable of detecting seizure onsets quickly and with high accuracy. Unlike previous efforts, which focused on adult EEG, we evaluate our patient-specific detectors on pediatric scalp EEG because it exhibits greater variability in seizure and non-seizure activity. Patient-specific seizure onset detection remains a challenge because · Patients with epilepsy have considerable overlap in the EEG associated with seizure and nonseizure states. This forces algorithm designers to confront a steep tradeoff between detector sensitivity and specificity. · The EEG of epilepsy patients constantly transitions between regimes both within the seizure and non-seizure states, and is therefore a nonstationary process. Characterizing the short-term evolution of EEG activity can be critical to inferring the brain's underlying state. · Numerous medical applications require seizure onsets to be detected quickly, i.e., a detector needs to ascertain that the brain has transitioned into 1. Introduction Seizures are transient aberrations in the brain's electrical activity. People with epilepsy, a central nervous system disorder, suffer from recurrent seizures that occur at unpredictable times and usually without warning. Seizures can result in a lapse of attention or a whole-body convulsion. Frequent seizures increase an individual's risk of sustaining physical injuries and may even result in death. A device capable of quickly detecting and reacting to a seizure by delivering therapy or notifying a caregiver could ease the burden of Appearing in Proceedings of the 27 th International Conference on Machine Learning, Haifa, Israel, 2010. Copyright 2010 by the author(s)/owner(s). Application of Machine Learning To Epileptic Seizure Detection a seizure state using few samples from that state. This forces algorithm designers to confront another steep tradeoff, between detector latency and specificity. · Since seizures are rare events, algorithm designers must craft methods that work with a paucity of seizure training data. Since the goal of seizure detection is to segment the brain's electrical activity in real-time into seizure and non-seizure periods, one could consider using an online, unsupervised time-series segmentation algorithm. Unfortunately, the many regimes of seizure and nonseizure EEG (Agarwal et al., 1998) cause such algorithms to return numerous segmentations beyond those demarcating seizure and non-seizure periods. A seizure detector needs to be taught which signal regimes and transitions are relevant. For this reason, we elect to solve the seizure detection problem within a supervised, discriminative framework. Within the discriminative framework we choose to solve a binary classification problem, despite the fact that the underlying physiological activity is multiclass. We do so because it is neither easy nor practical for an expert to identify and label the subclasses of the seizure and non-seizure states. In contrast, asking an expert to divide a record of the brain's electrical activity into two encompassing classes, seizure and nonseizure, is consistent with standard clinical practice. The key to our classifier's high accuracy is a completely automated process for constructing a feature vector that unifies in a single feature space the time-evolution of spectral and spatial properties of the brain's electrical activity. Previous patient-specific algorithms (Qu & Gotman, 1997; Shoeb et al., 2004; Haas et al., 2007) classified temporal, spectral, and spatial features separately, and required an individual skilled in interpreting the brain's electrical activity to specify how such features should be combined. Furthermore, our feature vector can be extended with information extracted from other physiologic signals. This is useful for detecting seizures associated with subtle changes in the EEG, but less subtle influence on other observable physiologic processes. It is important to distinguish our work from previous investigations concerned with using machine-learning to detect (Grewal & Gotman, 2005; Gardner et al., 2006) seizures using intracranial EEG. Algorithms that process intracranial EEG rely on features that cannot be observed within the scalp EEG because of the spatial averaging effect of the dura and skull. Furthermore, intracranial EEG is immune to corruption by artifacts (e.g. scalp muscle contractions) that can mask the onset of seizure activity within the scalp EEG. Finally, in evaluating our approach to seizure detection, we avoided testing methodology that might result in overly optimistic results. In (Gardner et al., 2006), seizures of a single type were used (temporal lobe seizures); data was hand-selected by an expert to be free of artifacts; and only 29 seizure and 41 nonseizure epochs each of 15-minute duration were tested (for a total of 17 hours of test data that included 29 seizures). In contrast, our dataset contains numerous seizure types as well as 916 hours and 173 test seizures. In (Mirowski et al., 2009), high specificity was achieved because test non-seizure feature vectors (33% of data) were not separated in time from training non-seizure feature vectors (66% of data). This creates testing and training sets that are highly correlated, since EEG exhibits significant temporal correlation. To evaluate our detector's specificity, we create test sets by removing hour-long records from a corpus of EEG data as opposed to removing second-long epochs. This paper is organized as follows. In Section 2 we review properties of the scalp Electroencephalogram (EEG). In Section 3 we discuss both the feature extraction and classification stages of our binary classifier. Our detector's performance is analyzed in Section 5. Finally, in Section 7, we illustrate how seizure detection can be enhanced through the addition of features extracted from another physiologic process. 2. The Scalp Electroencephalogram EEG measures the electrical activity of the brain using electrodes that are uniformly arrayed on the scalp. An EEG channel is formed by taking the difference between potentials measured at two electrodes, and captures the summed potential of millions of neurons. Following the onset of most seizures, a set of EEG channels develops rhythmic activity that is typically composed of multiple frequency components. The identity of the EEG channels involved and the structure of the rhythmic activity differs across individuals. For example, Figure 1 and Figure 2 illustrate seizures from different patients. Patient A's seizure in Figure 1 begins following 2994 seconds and is characterized by the appearance of rhythmic activity most prominently on the channels FP2-F4 and T8-P8. Patient B's seizure in Figure 2 begins at 1723 seconds with a spike followed by a period of low amplitude Application of Machine Learning To Epileptic Seizure Detection FP1-F7 F7-T7 T7-P7 P7-O1 FP1-F3 F3-C3 C3-P3 P3-O1 FP2-F4 F4-C4 C4-P4 P4-O2 FP2-F8 F8-T8 T8-P8 P8-O2 FZ-CZ CZ-PZ 2989 2990 2991 2992 2993 2994 2995 2996 2997 types of rhythmic activity are normal while others are abnormal but not associated with seizures. For example, the rhythmic activity observed between 2989-2992 seconds in Figure 1 is a normal feature of sleep EEG known as a spindle, and should not be confused with the seizure activity seen later in the same figure. 3. Seizure Detection as Binary Classification Our goal is to construct a function f (X ) that maps a feature vector X derived from the EEG onto the labels Y = ±1 depending on whether X is representative of seizure or non-seizure EEG. In the following subsections we discuss how we construct the feature vector X , the discriminant function f (X ), and the training sets. 3.1. Feature Vector Design In section 2, we noted that features important for characterizing EEG include its spectral structure, the channels on which it manifests, and its short-term temporal evolution. In the following subsections we illustrate how these features are extracted and encoded. 3.1.1. Spectral Features Figure 1. A seizure within the scalp EEG of Patient A. EEG. Next, rhythmic activity develops most prominently on the channel F3-C3, and, over the period of a few seconds, increases in amplitude and decreases in frequency. This seizure illustrates the non-stationarity of EEG within the seizure state. FP1-F7 F7-T7 T7-P7 P7-O1 FP1-F3 F3-C3 C3-P3 P3-O1 FP2-F4 F4-C4 C4-P4 P4-O2 FP2-F8 F8-T8 T8-P8 P8-O2 FZ-CZ CZ-PZ 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 Figure 2. A seizure within the scalp EEG of Patient B. Though seizures vary across individuals, the seizures of any given individual exhibit considerable consistency, provided that they emerge from the same brain region. Figure 3 illustrates a second seizure from patient B. Note the similarity in the spatial, spectral, and temporal character of this seizure and the seizure shown in Figure 2. FP1-F7 F7-T7 T7-P7 P7-O1 FP1-F3 F3-C3 C3-P3 P3-O1 FP2-F4 F4-C4 C4-P4 P4-O2 FP2-F8 F8-T8 T8-P8 P8-O2 FZ-CZ CZ-PZ 6206 6208 6210 6212 6214 6216 6218 6220 6222 The rhythmic activity associated with the onset of a seizure is often composed of multiple frequency components. For instance, the blue curve in Figure 4 represents the spectrum of the rhythmic activity observed on the channel FP2-F4 following the onset of the seizure in Figure 1. The rhythmic activity is composed of strong frequency components at 2, 5, and 11 Hz. Considering multiple spectral components is necessary for detecting seizures with high accuracy. The spectral content of a seizure epoch may overlap the dominant frequency of an epoch of non-seizure activity, but what distinguishes the two is the presence or absence of other spectral components. For example, the red curve in Figure 4 represents the spectrum of a sleep spindle. The two spectra overlap in the 10-12 Hz range, but the seizure spectrum contains stronger low-frequency spectral components. Because of EEG non-stationarity, it is important to extract spectral features from a reasonably small time epoch. Since EEG cannot be segmented into short physiologically relevant units, two second long epochs are commonly used. We extract the spectral structure of a sliding window of length L = 2 seconds by passing it through a filterbank composed of M = 8 filters, and Figure 3. Another seizure in the scalp EEG of Patient B. Not all rhythmic activity observed within the scalp EEG is a reflection of an underlying seizure. Certain Application of Machine Learning To Epileptic Seizure Detection 200 Seizure Onset Sleep Spindle 150 P8) that are not involved in the sleep spindle event. To automatically capture the spectral and spatial information contained within each two second EEG epoch at time t = T , we concatenate the M = 8 spectral energies extracted from each of N = 18 EEG channels. This process forms a feature vector XT with M × N = 144 elements as shown in the middle portion of Figure 5. 3.1.3. Time Evolution Decibels (dB) 100 50 0 -50 0 2 4 6 8 10 12 14 16 18 20 22 24 Frequency (Hz) Figure 4. Superposition of the frequency spectra of a sleep spindle and seizure activity. then measuring the energy falling within the passband of each filter. For channel k, the energy measured by filter i is denoted by the feature xi,k as shown on the left side of Figure 5. The filterbank spans the frequency range 0.5-25 Hz since most seizure and nonseizure EEG activity falls within this range. We characterized the performance of our detector as a function of using a filterbank composed of M = 2, 4, and 8 equal bandwidth filters. Increasing the number of filters did not greatly impact how quickly our detector recognizes the onset of a seizure, but helped to better discriminate between seizure and non-seizure states. The feature vector XT does not capture how an epoch relates to those in the recent past. Consequently, XT cannot represent how a seizure emerges from background EEG nor how it evolves. To capture such information, we form a stacked feature vector XT that is the result of concatenating the feature vectors from W contiguous, but non-overlapping 2 second epochs as shown on the right side of Figure 5. Note that encoding the temporal evolution of EEG through the formation of XT is not equivalent to forming a single feature vector XT using a longer epoch length. In the former approach, discrete events are preserved and, in the latter, the spectral signatures of discrete events are smeared. Electroencephalographers require an EEG abnormality to persist and evolve for a minimum of 6-10 seconds before considering the abnormality a seizure or a component of a seizure. To incorporate this domain knowledge, we set W = 3 so that our classifier considers the evolution of feature vectors over a period of 6 seconds. We also characterized the performance of our detector as W was swept from 2 to 4. As expected, we noted that increasing W decreases the detector's false detection rate and increases the latency with which it recognizes the onset of a seizure. 3.1.4. Non-EEG Features The earliest measurable sign of some seizures may not be rhythmic EEG activity, but a reflection of the physical sequelae of the seizure such as scalp muscle contractions or eye flutter. However, these activities are not reliable indicators of a seizure, since they are routinely observed outside the seizure state. To ascertain whether such routine activity is associated with a seizure, information beyond that within the EEG needs to be incorporated into the detection process. The additional information can be derived using a second physiologic signal whose dynamics are influenced by the presence or absence of a seizure. The second signal and the EEG complement each other if the changes in each signal that suggest the presence Figure 5. Feature extraction stages. 3.1.2. Spatial Features The identity of the EEG channels involved in seizure and non-seizure activity can further differentiate between these two classes. As an example, consider the sleep spindles between 2989-2992 seconds, and the seizure activity following 2994 seconds in Figure 1. The seizure involves EEG channels (C4-P4 and T8- Application of Machine Learning To Epileptic Seizure Detection 1 0.8 Energy (25 +/- 11Hz) 5 4 3 0.6 0.4 1 2 Figure 7. Training and missed test seizure within the EEG of Patient 15. 0.2 0 0 0.2 0.4 0.6 Energy (0-16Hz) 0.8 1 Figure 6. Nonlinear decision boundary separating seizure (red) and non-seizure (blue) feature vectors. later stages. In this example, the earliest seizure vector that is correctly classified is the third, which means a delay in declaring seizure onset. It is possible to choose SVM parameters that will force the construction of a more eccentric boundary that encloses the early seizure points, but at the expense of enclosing many more nonseizure points. We train the SVM on seizure vectors computed from the first S seconds of K seizures, and on non-seizure vectors computed from at least 24 hours of nonseizure EEG. A long period of non-seizure EEG is used to ensure that awake, sleep, abnormal, and artifactcontaminated EEG are included. Selecting the value of S involves a tradeoff. A small S focuses the SVM on seizure onset, but also causes the SVM to fail to detect a seizure whenever the onset changes. Increasing S expands the decision boundary and enables the detection of later seizure stages, but increases the detector's false-alarms because the extended boundary encloses more non-seizure vectors. In our experiments we set S = 20 seconds. The impact of the number of training seizures K on performance is discussed in Section 6. The SVM uses a number of hyperparameters, and it is not reasonable to expect clinicians to set them on a patient-specific basis. In our experiments we used a single fixed set of SVM hyperparameters within the SVMLight software package (Joachims, 1999). These parameters, which were determined through analysis of a different pediatric EEG dataset (Shoeb et al., 2004), include RBF kernel parameter = 0.1, and trade-off between classification margin and error C = 1. of a seizure rarely coincide during non-seizure states, and often coincide at the time of an actual seizure. Patient-specificity remains essential since the manner in which the EEG and the second signal jointly change during the seizure state is patient-specific. The electrocardiogram (ECG), a non-invasive measure of the electrical activity of the heart, is a good candidate for a complementary signal. It is easy to acquire, and there is significant evidence that seizures for many patients are associated with changes in heart rate (Zijlman et al., 2002). Unfortunately, ECG data was not available for most of the patients we were able to study. However, we did have one patient (patient 24) for whom the EEG and ECG were simultaneously recorded. In Section 7 we describe how features derived from that patient's ECG were concatenated with our EEG features in order to improve detector performance. 3.2. Feature Vector Classification We classify a feature vector as representative of seizure or non-seizure activity using a support-vector machine (SVM). Since the seizure and non-seizure classes are often not linearly separable, we generate non-linear decision boundaries using an RBF kernel. Still the SVM does not always enclose the most important data points. As an example, Figure 6 depicts a nonlinear boundary separating seizure (red) and non-seizure (blue) feature vectors extracted from a single EEG channel. The seizure feature vectors form two clusters: one cluster with fewer feature vectors that lies in close proximity to the non-seizure vectors, and a second cluster with a much greater number of vectors far away from the non-seizure vectors. The seizure vectors closest to the non-seizure vectors are associated with the onset of a seizure, and those further away are associated with 4. Evaluation Methodology 4.1. Scalp EEG Data Set The data set used to evaluate our detector consists of 916 hours of continuous scalp EEG sampled at 256 Hz. The data set was recorded from 23 pediatric patients at Children's Hospital Boston and one adult patient at Beth Israel Deaconess Medical Center. While the recordings were being made, the patients experienced 173 events that were judged to be clinical seizures by Application of Machine Learning To Epileptic Seizure Detection tector on NS seizure records and on NN S - 1 nonseizure records. The detector is then run on the withheld non-seizure record. This process is repeated NN S times. For each round, we note the number of false detections. Figure 8. Rhythmic burst and seizure within the EEG of Patient 13. 5. Performance 5.1. Sensitivity and Latency Overall, 96% of the 173 test seizures were detected. The mean latency with which the detector declared seizure onset was 4.6 seconds. More specifically, 50% of the 173 test seizures were detected within 3 seconds, 71% within 5 seconds, and 91% within 10 seconds. Our detector misses or has a large detection latency when a test seizure differs significantly in spatial or spectral character from all of the seizures in the training set. For instance, the rhythmic activity within the red box in the top panel of Figure 7 is associated with almost all of the seizures from Patient 15. The bottom panel of the figure highlights one seizure from the same patient that begins with a train of spikes instead of the rhythmic activity. When the detector is trained on all seizures except the one in the bottom panel and tested on the one in the bottom panel, the latency is 55 seconds. 5.2. Specificity Figure 9 shows the number of false detections declared by the detector per 24 hours for each of the 24 subjects. The median false detection rate is 2 false detections per 24 hour period. experts. For each clinical seizure, an expert indicated the earliest EEG change associated with the seizure. The data was segmented into one hour long records. Records that do not contain a seizure are called nonseizure records and those that contain one or more seizures are called seizure records. The pediatric EEG data used in this paper is contained within the CHB-MIT database, which can be downloaded from the PhysioNet website: http://physionet.org/physiobank/database/chbmit/ 4.2. Performance Metrics We characterized the performance of our seizure detector in terms of sensitivity, specificity, and latency. Sensitivity refers to the percentage of test seizures identified. Specificity refers to the number of times, over the course of 24 hours, that the detector declared the onset of seizure activity in the absence of an actual clinical seizure. Latency refers to the delay between expert-marked seizure onset within the EEG and detector declaration of seizure activity. 4.3. Performance Metric Measurement To estimate our detector's performance on data from a patient, we use a leave-one-record-out cross-validation scheme. We considered an evaluation method based on leaving out epochs rather than hour-long records, but that approach leads to misleadingly good results by including in the training set feature vectors in close temporal proximity to those in the test data. Let NN S denote the number of non-seizure records and NS denote the number of seizure records. To estimate the detector's latency and sensitivity, we train the detector on NN S non-seizure records from a patient (median NN S = 33) and on NS - 1 seizure records (median NS = 5). The detector is then tasked with detecting the seizure in the withheld seizure record. This process is repeated NS times so that each seizure record is tested. For each round we record whether the test seizure was detected, and if so, with what latency. To estimate the detector's specificity, we train the de- Figure 9. Specificity of our patient-specific detector. For Patient 13 the algorithm declared 20 false detections per 24 hour period. This is attributable to the presence of bursts of activity that resemble the seizure's onset. The top panel of Figure 8 illustrates a burst of activity that elicits a false detection. The burst closely resembles the seizure activity highlighted within the bottom panel of Figure 8. The main difference between the two is the seizure's temporal extent. Application of Machine Learning To Epileptic Seizure Detection To improve the specificity of our detector for Patient 13, one could modify the detector so that it declares a seizure only when activity suspected of being a seizure persists for a duration of time greater than the average length of the rhythmic bursts. The cost of such a modification would be an increase in detection latency. 5.3. Rate of Learning Figure 10 illustrates how the average detection latency and miss rate decrease with an increasing number of training seizures. The curve was derived using data from five randomly selected patients, but is representative of the behavior of the detector. With a single training seizure, the detector has an average latency greater than 7 seconds and misses more than 45% of the test seizures. With three training seizures, the detector has an average latency close to 4 seconds and misses less than 5% of test seizures. Fp1-F3 F3-C3 C3-P3 P3-O1 Fp2-F4 F4-C4 C4-P4 P4-O2 Fp1-F7 F7-T3 T3-T5 T5-O1 Fp2-F8 F8-T4 T4-T6 T6-O2 1484 1486 1488 1490 1492 1494 1496 1498 1500 1502 1504 1506 1508 1510 140 130 120 Beats Per Minute 110 100 Seizure Onset 90 80 70 1460 1465 1470 1475 1480 1485 Time (Seconds) 1490 1495 1500 Figure 11. A seizure with an onset lacking rhythmic activity, but which is accompanied by a heart-rate accleration. 7. Combining Physiologic Signals For Better Seizure Detection In this section, we illustrate how combining EEG and ECG information can improve seizure detection performance. Figure 10. Latency and sensitivity as a function of number of training seizures K 7.1. EEG-ECG Based Seizure Detection To include information within the ECG into the detection process, we augment the feature vector shown in Figure 5 with two ECG features. The features, which are extracted from 6 second epochs, are mean heart rate (X1,ECG ), and the difference between the terminal and initial heart rates within an epoch (X2,ECG ). Concatenating EEG and ECG features enables the SVM to automatically learn the relationship between spectral, spatial, and temporal information extracted from the EEG with heart rate and heart rate change information extracted from the ECG. 7.2. Evaluation 7.2.1. Data Set We recorded 10 seizures and 66 hours of synchronized EEG-ECG from Patient 24. Figure 11 illustrates a typical seizure. The onset of the seizure, at 1486 seconds, involves rapid eye-blinking, which manifests in 6. Comparison to a Patient Non-Specific Classifier We compared our performance to the performance of the Reveal algorithm (Wilson et al., 2004). Reveal is an offline, commercially available, patient non-specific detector that uses machine learning. Reveal relies on a neural network trained on hundreds of seizure and nonseizure epochs from a large number of pediatric and adult patients. When evaluated on the data set discussed in Section 4.1, Reveal detected 61% of seizures with a false detection rate of 33 false detections per 24 hour period. Although trained on a lot more data, Reveal's sensitivity and specificity are far worse than the 97% sensitivity and 2 false detections per 24 hour that we report. Application of Machine Learning To Epileptic Seizure Detection References Agarwal, R., Gotman, J., Flanagan, D., and Rosenblatt, B. Automatic EEG analysis during long-term monitoring in the ICU. Electroencephalography and Clinical Neurophysiology, 107(1):44­58, July 1998. Gardner, A., Krieger, A., Vachtsevanos, George, and Litt, Brian. One-class novelty detection for seizure analysis from intracranial eeg. Journal of Machine Learning, 7:1025­1044, Dec 2006. Grewal, S. and Gotman, J. An automatic warning system for epileptic seizures recorded on intracerebral eegs. Clinical Neurophysiology, 116(10):2460­2472, Oct 2005. Haas, S., Frei, M., and Osorio, I. Strategies for adapting automated seizure detection algorithms. Medical Engineering and Physics, 29(8):895­909, October 2007. Joachims, T. Making large-scale svm learning practical. Advances in Kernel Methods-Support Vector Learning, 1999. Meier, R, Dittrich, H, Schulze-Bonhage, A, and Aertsen, A. Detecting epileptic seizures in long-term human EEG: A new approach to automatic online and reat-time detection and classification of polymorphic seizure patterns. Journal of Clinical Neurophysiology, 25:119­131, Jun 2008. Mirowski, P., Madhavan, D., LeCun, Y., and Kuzniecky, R. Classification of patterns of eeg synchronization for seizure prediction. Electroencephalography and Clinical Neurophysiology, 120 (11):1927­1940, Nov 2009. Qu, H. and Gotman, J. A patient-specific algorithm for the detection of seizure onset in long-term EEGmonitoring: Possible use as a warning device. IEEE Transactions On Biomedical Engineering, 44:115­ 122, Feb 1997. Shoeb, A., Edwards, H., Connolly, J., Bourgeois, B., Treves, S., and Guttag, J. Patient-specific seizure onset detection. Epilepsy and Behavior, 5(4):483­ 98, Aug 2004. Wilson, S., Scheuer, M., Emerson, R., and Gabor, A. Seizure detection: Evaluation of the Reveal algorithm. Clinical Neurophysiology, 10:2280­2291, Oct 2004. Zijlman, M., Flanagan, D., and Gotman, J. Heart rate changes and ECG abnormalities during epileptic seizures: prevalence and definition of an objective clinical sign. Epilepsia, 43(8):847­854, 2002. Figure 12. Latency of detectors that use feature vectors containing EEG alone, or both EEG and ECG. the EEG as high-amplitude deflections on the channel FP1-F7. With the onset of eye-flutter, the patient's heart rate accelerates as can be seen in the heart rate profile in the bottom panel of Figure 11. Finally, at 1492 seconds, rhythmic activity appears on the channel T4-T6, and the patient's heart rate remains elevated. 7.2.2. Performance We compared the performance of the detector in Figure 5 when feature vectors are assembled using both EEG and ECG, and when they are assembled using EEG alone. Figure 12 illustrates the latency with which each detector declares the ten test seizures. The detector that only used EEG had a mean latency of 4.2 seconds, detected all but seizure 4 (note the missing bar), and declared 9 false detections per 24 hour period. The addition of ECG information improved performance. The detector that combined EEG and ECG had a mean latency of 2.7 seconds, detected all seizures, and declared 5 false detections per 24 hour period. 8. Conclusion We presented and evaluated a method of using SVM's to construct patient-specific classifiers that use scalp EEG signals to detect the onset of epileptic seizures. We believe that the method outlined in this paper is suitable for clinical use. Using the detector requires no expertise in machine learning, and the labeling of training data can be derived from information that is routinely supplied as part of clinical care. We are now in the process of working with physicians to conduct a prospective study in which detectors built in the way described here are used as part of a closed loop control system for a vagus nerve stimulator.