Fast evaluation of interictal spikes in longterm ... - Wiley Online Library

11 downloads 10020 Views 709KB Size Report
Mar 15, 2012 - In view of such findings, the advent of digital EEG and recordings ..... scheme provided only an approximate illustration of cluster centers and their ... Scherg M. (2011) Quick Guide on 3D Maps. http://www.besa.de/tutorials/.
Epilepsia, 53(7):1196–1204, 2012 doi: 10.1111/j.1528-1167.2012.03503.x

FULL-LENGTH ORIGINAL RESEARCH

Fast evaluation of interictal spikes in long-term EEG by hyper-clustering *Michael Scherg, *Nicole Ille, *Dieter Weckesser, *Arndt Ebert, *yAndrea Ostendorf, zxTobias Boppel, zSusanne Schubert, {#Pa˚l G. Larsson, #Oliver Henning, and z**Thomas Bast *BESA GmbH, Gra¨felfing, Germany; yDepartment of Natural Sciences, Ruhr West University, Mu¨lheim, Germany; Departments of zGeneral Pediatrics and xNeuroradiology, University Hospital, Heidelberg, Germany; Departments of {Neurosurgery and #Refractory Epilepsy, Oslo University Hospital, Oslo, Norway; and **Kork Epilepsy Center, Kehl-Kork, Germany

SUMMARY Purpose: The burden of reviewing long-term scalp electroencephalography (EEG) is not much alleviated by automated spike detection if thousands of events need to be inspected and mentally classified by the reviewer. This study investigated a novel technique of clustering and 24-h hyper-clustering on top of automated detection to assess whether fast review of focal interictal spike types was feasible and comparable to the spikes types observed during routine EEG review in epilepsy monitoring. Methods: Spike detection used a transformation of scalp EEG into 29 regional source activities and adaptive thresholds to increase sensitivity. Our rule-based algorithm estimated 18 parameters around each detected peak and combined multichannel detections into one event. Similarity measures were derived from equivalent location, scalp topography, and source waveform of each event to form clusters over 2-h epochs using a densitybased algorithm. Similar measures were applied to all 2-h clusters to form 24-h hyper-clusters. Independent raters evaluated electroencephalography data of 50 patients with epilepsy (25 children) using traditional visual spike

Although sharp transients in the scalp electroencephalography (EEG), usually called spikes or sharp waves, have been central in the diagnosis of epilepsies, the main interest during presurgical video-EEG monitoring of patients with intractable epilepsy is on defining the seizure-onset zone (Rosenow & Lders, 2001). A recent study of 152 patients after epilepsy surgery, however, documented a high predictive value (84%) of averaged interictal spikes to localize the relevant brain region by using high-density scalp recordings

Accepted March 15, 2012; Early View publication May 11, 2012. Address correspondence to Michael Scherg, BESA GmbH, Freihamer Str. 18, 82166 Grfelfing, Germany. E-mail: mscherg@besa.de Wiley Periodicals, Inc. ª 2012 International League Against Epilepsy

review and optimized hyper-cluster inspection. Congruence between visual spike types and epileptiform hyper-clusters was assessed on a sublobar level using three-dimensional (3D) peak topographies. Key Findings: Visual rating found 126 different epileptiform spike types (2.5 per patient). Independently, 129 hyper-clusters were classified as epileptiform and originating in separate sublobar regions (2.6 per patient). Ninety-one percent of visual spike types matched with hyper-clusters (temporal lobe spikes 94%, extratemporal 89%). Conversely, 11% of hyper-clusters rated epileptiform had no corresponding visual spike type. Numbers were comparable in adults and children. On average, 15 hyper-clusters had to be inspected and rated per patient with an evaluation time of around 5 min. Significance: Hyper-clustering over 24 h provides an independent tool for rapid daily evaluation of interictal spikes in long-term video-EEG monitoring. If used in addition to routine review of 2–5 min EEG per hour, sensitivity and reliability in noninvasive diagnosis of focal epilepsy increases. KEY WORDS: Spike detection, Spike clustering, Source localization, Source montages, 3D mapping.

and individual realistic head models (Brodbeck et al., 2011). In view of such findings, the advent of digital EEG and recordings extending to more electrodes, detection, and quantification of interictal spikes is getting more relevant. For automated detection, various methods have been proposed (reviews in Wilson & Emerson, 2002; Halford, 2009) ranging from rule-based methods deriving mimetic parameters from spike waveshape and EEG background signals (Gotman & Gloor, 1976; Gotman & Wang, 1991) to trained neural networks (Wilson et al., 1999) and wavelet analysis (Goelz et al., 2000). These studies aimed at comparing the performance of human reviewers with computerized spike detection, typically using a small number of EEG epochs and spikes. Quantification of different types of spikes, their frequency, and circadian distribution was not feasible

1196

1197 Fast Evaluation of Interictal Spikes because marking and classifying all visible spikes in longterm EEG constitutes an overwhelming workload. Clinical routine evaluation of interictal spikes in longterm EEG is time-consuming and requires experienced reviewers for visual identification and quantification of epileptic discharges. Therefore, after careful review of the first hours, visual inspection is typically limited to epochs of 2– 5 min/h to document seemingly relevant spikes. Furthermore, the perception of a spike depends on the selected montage (Wilson et al., 1996), with the danger of overlooking relevant spike types. Automatic spike detection should help the reviewer to save time in finding relevant interictal patterns and providing comprehensive overview and decision making. However, many spike detectors mark all detections in any channel and/or montage as distinct events in the EEG and yield thousands of events to be classified—an additional workload not accepted in clinical routine. Therefore, Wilson et al. (1999) proposed to combine near-to-synchronous detections in different channels into one event and to apply hierarchical clustering to group spikes with similar topology and morphology. However, clustering was shown only for 101 spikes in one short EEG example. Van Hese et al. (2008) used a five-step method including dipole localization to find the most prominent focal spike cluster in eight pediatric routine EEGs. For long-term EEGs in larger patient groups, quantitative results on the number of spike and nonspike types using clustering have not yet been published. This study investigated a time-saving novel technique of clustering and 24-h hyper-clustering on top of automated detection in long-term EEGs of 50 patients undergoing presurgical epilepsy monitoring. The purpose was to assess whether a fast review of spike types defined automatically by hyper-clustering over 24 h was feasible and comparable to the focal spikes types defined by clinical review. Care was taken to optimize the review tools for a fast informed decision on the hyper-clusters, summarizing either similar epileptiform events or other EEG patterns, for example, artifacts.

Methods Patients In this retrospective study, we analyzed anonymized EEG data of 50 patients who had undergone presurgical epilepsy workup. Long-term monitoring data were collected from 25 patients each at the Pediatric Epilepsy Unit, University of Heidelberg, Germany, and at the Epilepsy Center in Sandvika, Norway. Ethics committee approval was obtained for post hoc analyses of EEG data at both institutions. Data were from consecutive patients having at least 22 h of EEG recorded within a 24 h epoch. Patients with generalized epilepsy (symptomatic or cryptogenic) or epileptic encephalopathies including West and Lennox-Gastaut

syndromes were excluded, since this study was on detecting focal spikes. The 25 patients from the adult group had a mean age of 35 years (range 17–68 years). Age of the 25 children ranged from 2–16 years (mean 9.4). All 50 patients had seizures and were diagnosed having focal epilepsy. Thirteen adults had known lesions: ischemic infarction, 2; hippocampal sclerosis, 1; posttraumatic, 4; tumor, 4; dysplasia, 1; and cavernous malformation, 1. Eighteen children had known lesions: cortical malformation, 10; tuberous sclerosis, 2; tumor, 2; and median artery infarction, 4. The other cases were cryptogenic. EEG data EEG was recorded from 25 electrodes in Sandvika using the 19 electrodes of the 10–20 system plus three inferior electrodes on either side (F9/T10, T9/T10, P9/P10). In Heidelberg, 32 EEG channels were recorded with additional electrodes at FC1/FC2, FC5/FC6, CP1/CP2, CP5/CP6 using Pz-reference. EEG was average referenced to utilize all 33 electrodes (Scherg et al., 2002). In all cases, electrocardiography was recorded. Data were anonymized for digital review and analysis. For visual inspection, the EEG data were filtered digitally using a time-constant of 0.3 s and a 50 Hz notch if needed. For automated event detection, data were filtered into two separate buffers using a so-called ‘‘EEG-filter’’ and a ‘‘spike-filter.’’ The EEG filter consisted of a forward low filter of 1 Hz (6 dB/octave, equivalent to a 0.16-s time constant) and a high cutoff of 35 Hz (zero-phase, 24 dB/oct.). The spike filter was from 4 Hz (zero-phase, 12 dB/oct.) and to 35 Hz (zero-phase, 24 dB/oct.). Event detection Detection of events in the 24-h EEG studies and clustering were done offline to create spike and cluster files for each subject using the BESA Epilepsy 1.0 Detection software (http://www.besa.de). Transient events were detected by a novel, rule-based algorithm utilizing source montages to increase the sensitivity for spikes arising in fissural and obliquely oriented cortex. First, the EEG-filtered data blocks were checked for artifacts. Channels were excluded if exceeding a threshold of 2,000 lV peak to peak. Artifactfree channels were transformed into a reference-free source montage covering 29 brain regions (12 right, 12 left, 5 midline, cf. Fig. 1). The current vector of each brain region was estimated using three orthogonal dipoles oriented radially (toward the center of the head model sphere) and tangentially (one in horizontal and one in vertical plane). Because a spiking cortical surface can have any net orientation in 3D space, we used linear combinations of these three local dipole vectors to calculate the currents along their three axes (vectors: 1, 0, 0; 0, 1, 0; 0, 0, 1), along the six diagonals between the axes (orientations, not normalized: 1, €1, 0; €1, 0, 1; 0, €1 ,1), and along the four 3D diagonals of the unit cube around the local origin (orientations: €1, 1, €1). Epilepsia, 53(7):1196–1204, 2012 doi: 10.1111/j.1528-1167.2012.03503.x

1198 M. Scherg et al.

Figure 1. Equivalent locations of 29 regional sources within a spherical head model used to create the source montage for spike detection. Sources were distributed evenly over the whole brain and symmetric in the left (blue) and right (red) hemispheres. Left: Top view showing upper sources. Right: lateral view showing left and midline (black) sources. Epilepsia ILAE

Therefore, we ensured that a local spike of any orientation would project with at least 92% magnitude on one of these 13 vectors and be detected against EEG background activities of other orientations. Using the spike-filter, each of these 13 transformed signals was checked continuously for emerging peaks and further parameterized if exceeding an adaptive threshold. For each regional source, the mean vector amplitude was calculated in half-second intervals during the previous 10 s. Each source threshold was continuously adapted to the 1.6-fold maximum value of the lowest 40% of these 20 values. Thresholds below a minimum ranging linearly from 42 nAm at the most inferior temporal source to 32 nAm at the most superior source were set to this value. Peaks exceeding threshold had to be salient relative to the surrounding EEG as defined by four criteria: (1) ratio of detected peak relative to preceding peak >2.0; (2) ratio of detected peak or any other peak within €20 msec relative to next peak >0.25 (to detect small spikes followed by larger wave); (3) difference of peak amplitude to the mean of all other surrounding positive peaks within )300 msec to +100 msec divided by peak amplitude >70%; and (4) ratio of peak amplitude to the root-mean-square amplitude over the previous 10 s (crest factor) >1.4. Detected peaks were ignored if related to electrooculography (EOG) /ECG artifacts. EOG was calculated using three virtual channels (Scherg et al., 2002) to render blinks and horizontal and vertical eye movements. Peaks in EOG >80 lV and R-peaks in the QRS-complex of the ECG (detected using the algorithm of Hamilton & Tompkins, 1986) resulted in rejection of source peaks coinciding within €20 msec. For each detected peak, optimum orientation of the related regional source was found by principal components analysis (PCA) of the EEG filtered underlying three dipole Epilepsia, 53(7):1196–1204, 2012 doi: 10.1111/j.1528-1167.2012.03503.x

signals in an interval of €80 msec around the peak. The resulting source waveform of the first PCA component was used to estimate 18 parameters for each peak source (cf. Appendix). Based on these, spike probability was determined by a discriminant analysis based on training data derived from 285 spike segments (group 1) and 420 falsepositive transients (group 2). Segments were selected from previously published EEG data of 59 patients (Bast et al., 2004, 2005, 2006; Ramantani et al., 2006). The same data were used to define the limits of the other parameters and thresholds described in this section. Spike probability was determined as the posterior group-specific probability of the 18-dimensional vector discriminating the spike group from the false positives. Detected events with spike probability 0.65 with the trimmed means. Minimally, four events were retained for small clusters. Finally, mean locations, waveforms, and topographies of all cluster cores in every 2-h epoch were calculated for the ensuing hyper-clustering process. Hyper-clustering and decision workflow To combine similar clusters over the 24-h review period, the same cluster algorithm was applied to the mean core locations, topographies, and waveforms of all clusters found in the 12 2-h epochs. This second cluster process used a similarity threshold of 0.75 for both topographies and waveforms to form so-called hyper-clusters. There was no lower limit in the number of clusters per hyper-cluster. Therefore,

already two clusters could form a hyper-cluster. In order not to lose any cluster information, e.g. from rare events during one epoch, even single clusters without similarity to any others were counted as separate hyper-clusters. After selecting a patient from the database, hyper-clusters were calculated, sorted by frequency, and displayed for inspection in the BESA Epilepsy 1.0 Evaluation software (http://www.besa.de). To facilitate decision, each hypercluster was initially displayed along with the most-populated 2-h cluster and the EEG around the first event in this cluster (Fig. 2). The rater clicked on different events and cluster epochs to inspect the EEG, 3D peak maps, and spike waveforms to decide whether the displayed hyper-cluster comprised epileptiform events or not. After each decision, the next hyper-cluster was displayed automatically until all hyper-clusters were classified. Epilepsia, 53(7):1196–1204, 2012 doi: 10.1111/j.1528-1167.2012.03503.x

1200 M. Scherg et al. Hyper-clustering rating The rating of the hyper-clusters was done by one of the nonclinical authors sufficiently trained to recognize spikelike transients, normal EEG patterns, and artifacts. The display was standardized using a virtual montage (Scherg et al., 2002) with 33 systematically grouped channels (F9, T9, P9, F10, T10, P10, Fp1, F7, FC5, T7, CP5, P7, O1, Fp2…O2, F3, FC1, C3, CP1, P3, F4…P4, Fz, Cz, and Pz) against an average reference based on all 10-10-electrodes and a band-pass of 2–35 Hz (zero-phase-shift filters). During inspection of events in this standardized display, filters could be switched off briefly by releasing the filter button to confirm artifacts such as eye movements. The rating included a visual analysis of the spike topography. The 3D map at the event peak was rotated automatically to display the map center from the best among 19 different viewpoints. Therefore, the rater could compare rapidly several individual peak maps with the mean over all events of the displayed hyper-cluster (Fig. 2) to identify the negative peak and center of the event topography (Scherg, 2011). The center of the—typically dipolar—map was used to rate the side, lobe, and sublobar region of each hyper-cluster considered epileptiform (for definitions and abbreviations see Table 1). Visual rating of spike types Two senior and one junior epileptologist undertook the visual rating of spike types independently of the hyper-cluster rating using traditional EEG review montages in the BESA Research 5.2 software (http://www.besa.de). The protocol reflected our typical presurgical workup using EEG segments of 2–5 min duration per hour for spike review. First, 1–2 h periods of EEG both in sleep and

Table 1. Definitions for comparison of spike types Side R L M B Lobe F

Right Left Midline Bilateral Frontal

C

Central

P

Parietal

T

Temporal

O

Occipital

wakefulness were reviewed by the junior rater under the guidance of one senior rater to identify different visual spike types in each patient. Second, the junior systematically reviewed the first 5 min of each hour (or later, if EEG was too artifactual) to find out if additional spike types existed. All events were marked as spike patterns 1–4 according to apparent frequency. In addition, apparently different spikes were marked by pattern 5. In a third step, every single marked event was reviewed by the senior who confirmed epileptic spikes and removed doubtful or false patterns. Spike patterns 1–4 were averaged and typical examples of each pattern and the mean patterns were mapped at the spike peak. As described for hyper-clusters, the center of the 3D map was used similarly to rate side, lobe, and sublobar region of each visual spike type. Comparison of visual and hyper-cluster spike types To compare spike types obtained by visual spike (VS) scanning and hyper-cluster (HC) rating, the first four visual spike types were used, or less if fewer were scored. Only in 4 adults and 10 children were more than four patterns observed and rated as pattern 5, but this mostly contained rare and inconsistent events. VS types with less than four events during visual EEG review were excluded. Up to six HC spike types were allowed in the comparison, since sometimes more than one HC was attributed to the same sublobar region. Correct detection was assumed for each VS type if a corresponding HC type existed based on the following criteria: (1) side and lobe were identical (midline and bilateral were considered the same since both exhibited similar midline peaks); (2) if sublobar region was different, correspondence was assumed only if regions were adjacent. For example, occipital spikes rated visually as OM r were considered corresponding to hyper-cluster OL r, whereas VS rated anterior temporal was not equivalent to HC rated posterior temporal, since regions were not adjacent. Missing detection was assumed if no HC corresponded.

Results Sublobar regions FM Medial near Fz FL Lateral FP Frontopolar CM Medial near Cz CL Lateral PM Medial near Pz PL Lateral TA Anterior/polar TB Basal TL Lateral TP Posterior TS Superior/Sylvian fissure OM Medial near Oz OL Lateral

The large frontal lobe was subdivided into two regions (F, C).

Epilepsia, 53(7):1196–1204, 2012 doi: 10.1111/j.1528-1167.2012.03503.x

Visual spike types Visual spike scanning of the 24 h recordings yielded 126 different epileptiform spike types in the 50 patients, that is, 2.5 spike types on average per patient. In five temporal and seven extratemporal cases two spike types were assigned to the same sublobar region but counted as one. Thirty-one patients had spikes only in one hemisphere. Of these, 17 were in one, 13 in two, and 1 in three lobes. Nine patients had independent spikes in one lobe of each hemisphere, six of these bitemporal. In these cases, the proportion of independent spikes in the left versus right hemispheres was similar for visually marked and hyper-cluster spike types. Six patients had independent multifocal spikes with both hemispheres and more than one lobe involved.

1201 Fast Evaluation of Interictal Spikes Origin as defined by rating the map centers was temporal in 50 spikes types and extratemporal in 65 (Table 2). The number of spike types was slightly larger for children (65, average 2.6) than adults (61, average 2.4). Adults had more temporal lobe spike types (34 vs. 22), whereas children had more extratemporal spike types (43 vs. 16; McNemar chisquare test: v2 = 5.78, p = 0.0162). The number of spikes marked during visual rating ranged from zero over two in two patients with temporal lobe epilepsy to >1,200 in three adults and nine children. Therefore, visual marking was very time consuming.

Table 2. Hyper-cluster detection of visually defined spike types

Children (25) Temporal Extratemp. Adults (25) Temporal Extratemp. All (50) Temporal Extratemp. Total

Detected

Not detected

Total

Detected %

16 43

1 5

17 48

94 90

34 22

2 3

36 25

94 88

50 65 115

3 8 11

53 73 126

94 89 91

Hyper-cluster spike types Between 5 and 25 hyper-clusters had to be classified (mean 14.5 per patient). With some experience, this procedure took approximately 5 min on average, since hyperclusters containing artifacts were discarded rapidly (Fig. 3). Two hundred fourteen hyper-clusters were rated as epileptiform and redisplayed to define the originating side, lobe, and sublobar region using the 3D maps (Fig. 4). Eighty-five of these represented events similar to one of the other 129 hyper-clusters being assigned to the same sublobar region. Therefore, in the 50 patients, 129 different spike types were observed using hyper-clustering, that is, 2.6 on average. Comparison of hyper-cluster and visual spike types The number of different spike types defined by visual scanning (VS, 126) was similar to the number of spike types obtained by hyper-clustering (HC, 129). The region assigned by independent rating agreed in 115 of the 126 VS spike types (91%) with at least one HC (Table 2). Overall, correct detection rate was nonsignificantly higher in temporal lobe (94%) as compared to extratemporal spike types (89%). Neither was there a significant difference in the number of missed VS spike types between children (6) and adults (5). Conversely, 14 of the 129 hyper-clusters rated as epileptiform had no correspondent VS spike type (adults: 6, children: 8). In many cases with missed detection of visual spike types, the number of visually observed spikes was

Figure 3. Hyper-cluster 2, comprised of lateral eye-movement artifacts and rated as nonepileptiform (cf. decision button upper right). The filter button (2–35 Hz) was released briefly to confirm the artifact signals. Epilepsia ILAE Epilepsia, 53(7):1196–1204, 2012 doi: 10.1111/j.1528-1167.2012.03503.x

1202 M. Scherg et al.

Figure 4. Hyper-cluster 2 of another patient. The negative peak of the near-to-tangential map is below CP5, whereas the map center is between C3 and FC5. Rated as epileptiform with location CL l (cf. Table 1). Epilepsia ILAE

small or spikes were blurred by background activities, for example, in frontocentral regions.

Discussion The focus of this study was not on single spike detection but on the different types of focal spikes found by two independent methods in 50 patients. Congruence between spike types detected by visual scanning and hyper-clustering was high, with HC detection rates of 94% in temporal lobe and 89% in extratemporal lobe epilepsy. This illustrates the potential risk of missing 6–11% of focal spike types when using hyper-clustering alone. However, if used in addition to visual review based on 2–5 min/h, hyper-clusters can ascertain visual spike findings and increase overall sensitivity and reliability, since 11% had no corresponding VS. For example, in the 15-year-old female patient without visually rated spikes, five left temporal-basal spikes were clustered. Hyper-clustering provided quantitative numbers on frequency and time of occurrence for each spike type, for example, temporal lobe spikes at night (Fig. 2). Numbers depend on the perception of each spike (Wilson et al., 1996) and should be used with caution. The more a spike is emerging from surrounding EEG the more likely is its visual recognition and automated detection. More prominent spikes are more easily counted, whereas smaller spikes may go undetected. Cluster counts could also depend to some degree on artifact events with similar location by reducing Epilepsia, 53(7):1196–1204, 2012 doi: 10.1111/j.1528-1167.2012.03503.x

the likelihood for clustering of neighbors in our densitybased algorithm. Furthermore, artifacts, for example, from eye movements, can exhibit topographies similar to temporal and frontal spikes. Therefore, clusters may contain spikes and some other events. Because this influences relative frequencies, only cluster cores were used for quantitative evaluation. Although the presented detection rates were approximately 90% for spike types, interrater congruence for single spikes was 35 Hz. 18. Jump parameter: To identify electrode artifacts within €100 msec, the maximum absolute difference between adjacent values in the solely high-pass EEG-filtered first PCA component was determined and multiplied by 0.005. The hyperbolic tangent of this value yielded a jump parameter ranging from 0 to 1. A value close to 1 indicated a high-amplitude signal change.

Approximate Location Location r was approximated by weighting the normalized electrode coordinate vectors of the 81 standard 10–10 electrodes with squared voltage values vi (Scherg & von Cramon, 1984) given by correlating the first PCA source waveform with the 81 EEG-filtered average-referenced signals from )200 to +100 msec around the peak: r ¼ ðRv2i ei Þ=Rv2i

Suggest Documents