Brain Topography, Volume 17, Number 2, Winter 2004 (© 2004)
99
Localization of Brainstem Auditory Evoked Potentials in Primates: A Comparison of Localization Techniques Applied to Deep Brain Sources Joel B. Fontanarosa*, Robert E. Lasky+, Hyong C. Lee*, and Wim van Drongelen*
Summary: The objective of this study was to evaluate the performance of source localization techniques through localization of deep brain sources. To accomplish this, two replications of a brainstem auditory evoked potential (BAEP, left ear 60 dB nHL clicks) were recorded from five normal rhesus monkeys. We analyzed waves III and IV, as this portion of the BAEP corresponds to the deepest signal. Data were analyzed using five different source localization techniques: Moving Dipoles, Fixed Dipoles, MUSIC (Multiple Signal Classification) dipole scan, LORETA (Low Resolution Tomography), and LCMV (Linearly Constrained Minimum Variance) spatial filtering. The moving dipole, fixed dipole and MUSIC solutions were found to be, on average, 25.1 mm from the brainstem generators. LORETA detected sources within the brainstem 65% of the time. However, 90% of these localization results also included false detections defined as regions of the brain that were more than 2 cm away from the auditory pathway. LCMV included the brainstem in 90% of the trials and false detections in 40% of the cases. These findings indicate that evoked electrical activity from deep brain sources can be localized with cm accuracy. The dipole methods performed better than LORETA and LCMV. Given the depth and amplitude of the sources analyzed in this study, these results can be interpreted as an upper bound on the accuracy of each technique. Key words: Source localization; Dipole analysis; MUSIC; LCMV spatial filter; LORETA.
Introduction Recordings of electrical brain activity have excellent temporal resolution but poor spatial resolution. Over the past decades, source localization algorithms were developed to improve spatial localization by determining the position and strength of electrical sources associated with surface recorded electrical potentials (e.g., Scherg 1984; Scherg and von Cramon 1985; Mosher et al. 1992; Pascual-Marqui et al. 1994; Van Veen et al. 1997). Source localization from surface recorded electrical brain activity is useful and important in both research and clinical
* Department of Pediatrics, The University of Chicago, Chicago, IL, USA. + Center for Clinical Research and Evidence-Based Medicine, University of Texas–Houston Medical School, Houston, TX, USA. Accepted for publication: September 23, 2004. We thank Drs. V.L. Towle, J.D. Hunter, and K.E. Hecox for useful comments and discussion and G. Rao for his help with the LCMV implementation. This work was supported by a Falk Grant, a Peterman Grant, and a Howard Hughes Grant. Correspondence and reprint requests should be addressed to Wim van Drongelen, PhD, The University of Chicago Biological Science Division, Department of Pediatrics, MC 3055, 5841 S Maryland, Chicago, IL, USA, 60637. Fax: (773) 702-4786 Copyright © 2004 Springer Science+Business Media, Inc.
settings, and is easy to apply because the techniques are non-invasive and inexpensive. As methods to localize electrical brain activity improve, surface measurements of electric activity such as the electroencephalogram (EEG) or evoked potentials (EP) will become increasingly powerful research and clinical tools to explore spatio-temporal patterns of neural activity. Estimation of the location of the source of electric activity representing the underlying generators of an observed scalp potential is approached with combined forward and inverse solutions. The forward model describes the scalp potentials for a given (set of) source(s); the inverse solution relates a measured surface potential to the generator position(s). The instantaneous EEG or EP signals that are measured can be considered as the weighted linear summation of the currents occurring at each instant (Nunez 1981). Even with the availability of significant computational power, modeling the head consisting of different regions with different electrical properties is not feasible. Simplification of the anatomy is required, and therefore, the forward model varies with selected geometry and detail. Simple forward models describe the brain, skull and scalp as a set of three concentric spheres (e.g. Scherg and von Cramon 1985). More realistic models use boundary elements or finite elements to derive a head model from a series of MRI slices (e.g., Le and Gevins 1993; Fuchs et al. 2002). Given the source loca-
100
tion(s), model geometry and electrical properties, the forward solution generates a unique solution for the scalp potential distribution. The inverse solution is ill-posed and requires additional constraints to associate a unique generator to a given scalp potential distribution. Inverse algorithms differ by the constraints they impose. The moving dipole and fixed dipole inverse approaches require a priori knowledge of the number of dipoles involved. The fixed dipole solution also implies absence of propagation of the dipole within the analysis window. The Multiple Signal Classification (MUSIC) (Mosher et al. 1992) we applied scans potential locations using a one-dipole model in order to find the optimal solution. As with the moving and fixed dipole algorithms, the number of dipoles that are present must be specified. For a known and small number of focal signals with a strong signal-to-noise ratio (SNR), the moving, fixed, and MUSIC dipole approaches can be expected to perform well. Algorithms that generate three dimensional (3D) activity distributions may be well-suited for localizing multi-focal and/or wide spread cortical activity patterns. This type of algorithm also requires constraints to obtain a unique inverse solution. The Low Resolution Tomography (LORETA) algorithm assumes a smooth 3D potential field (Pascual-Marqui et al. 1994). If the assumption of smoothness cannot be made or the number of active sources is not known a priori, then the Linearly Constrained Minimum Variance (LCMV) spatial filter (Van Veen et al. 1997) is the more appropriate algorithm. The purpose of this study was to estimate the accuracy of several source localization methods under the challenging conditions associated with localizing a deep source. To accomplish this, we evaluated the performance of the moving and fixed dipole algorithms, MUSIC dipole scan, LORETA, and LCMV to localize the sources of the brainstem auditory evoked potential (BAEP) in the rhesus monkey. Waves III and IV of the primate auditory brainstem response were selected as representatives of deep brain sources because their localization result can be validated by comparison to the generators that were determined electrophysiologically (Møller and Burgess 1986; Legatt et al. 1986; Urasaki E et al. 1995; Alegre et al. 2001). Previous studies localizing the BAEP have used recordings with a small number of electrodes in non-standard positions and spherical head models (Grandori 1984; Scherg 1984; Scherg and von Cramon 1985; Witt 1991). In this study, we use a montage of 32 electrodes positioned according to the 10-10 system (American electroencephalographic Society 1994), we used boundary element head models for each individual subject, and we evaluated the precision of the localizations by computing the Euclidian distance between the localized and theorized generator positions. We introduce a novel method to evaluate 3D distributions ob-
Fontanarosa et al.
tained from LORETA and the LCMV spatial filter in a semi-quantitative fashion. Because we recorded replicate BAEPs, we were also able to estimate the reliability of source localization of the applied algorithms.
Methods Data Acquisition and Preprocessing Five rhesus monkeys were included in the study. The sample consisted of 1 male and 4 female 17-year-old monkeys (17 years, 3 months to 17 years, 5 months). These monkeys had participated as controls in a study evaluating the effects of early lead exposure (Lasky et al. 1995). Thirty-two electrodes were applied to the scalp of the monkeys (impedances were less than 5 kΩs before and after data collection). The electrodes were positioned according to the extended 10-10 system (American Electroencephalographic Society 1994). The positions included were Fp1, Fp2, F7, F8, T3, T4, T5, T6, O1, O2, F1C1, F2C2, Cz, Pz, C1P1, C2P2, C3, C4, F3, F4, F7T3, F8T4, T3T5, T4T6, F5C5, F6C6, C5P5, C6P6, P3, P4, Fpz, and Fz, referenced to the average of the signals recorded from the mastoids. Recorded electrode coordinates were identified by a 3D localization system (Polhemus, Inc., Colchester, VT). Brainstem auditory evoked potentials (BAEP) were recorded with a 32-channel Pathfinder system (Nicolet Biomedical Inc., Madison, WI). Clicks (rarefaction, 100 µs duration, 60 dB nHL) were presented to the left ear at a rate of 27.3 Hz. Data were recorded at a sensitivity of 50µV, filtered between 0.1 – 1.5 kHz and sampled at a rate of 17 kHz per channel. 1000 trials were averaged for each data set, and two data sets were collected for each monkey. For data processing, the BAEP waveforms were imported into the CURRY software package (Neuroscan, El Paso, TX). A 2 ms pre-stimulus period was used to correct the baseline and to estimate the noise level of the recorded waveforms. The amplitudes from all channels were then transformed to signal-to-noise ratio (SNR) values. MRI Data, the Forward Model, and the Coordinate System For each monkey, a T1 weighted full head MRI (3D SPGR Coronal scan, TE=2.2, TR=11.4, acquisition matrix = 256 × 256, 2 excitations) was acquired on a GE Signa 1.5 T scanner. One hundred and twenty-four slices were recorded with slice thicknesses of 1.2 mm and a 0.0 mm skip between slices. The MRI data were imported into the CURRY software package for the construction of the boundary element model (BEM). The BEM contained surfaces for the scalp, skull, and brain compartments. The absolute conductivities assigned to the scalp, skull, and brain parenchyma were 0.33, 0.04125, and 0.33 S/m,
Source Localization of BAEP
101
Figure 1. Recorded brainstem auditory evoked potentials (BAEP) from subject 4. The 32 channels of baseline corrected data are superimposed in this figure and peaks I-IV are identified. The intervals used for the dipole algorithms and LCMV analysis are indicated. The time ranges used in the LORETA studies were the same as those used for the dipoles.
Figure 2. Source localization results for wave IV from subject 4. (a) An example of the dipole solutions: moving dipole (yellow), fixed dipole (red), and MUSIC (blue). The upper part of Figure 3a shows sagittal and coronal MRI slices with parts of the auditory pathway indicated: the superior olive (SO), the lateral lemniscus (LL), the inferior colliculus (IC), and the medial geniculate body (MG). The lower part of Figure 3a is a three dimensional reconstruction of the brain viewing the posterior, left and top of the brainstem (upper row) and whole brain (bottom row). The clipped result for the LCMV spatial filter result (b) and the LORETA result (c) are superimposed on a three dimensional reconstruction of posterior, left, and top views of the brainstem(upper rows) and the whole brain (bottom rows).
102
respectively (Hoekema et al. 2003). In this study, the x-axis was the inter-aural axis (positive left), the y-axis was the axis connecting the inion to the nasion (positive inion), and the z-axis was perpendicular to the x, y plane (positive up). Measured electrode positions were imported into the CURRY software package. Electrodes were fitted to the BEM model by using coordinates generated from the Polhemus system with the nasion, inion, and the left and right pre-auricular notches as landmarks. Source Localization Techniques and Data Analysis The localization techniques applied in this study can be subdivided into dipole localization algorithms (Moving Dipoles, Fixed Dipoles, and the Multiple Signal Classification (MUSIC) dipole scan) and current density methods (Low Resolution Tomography (LORETA) and Linearly Constrained Minimum Variance spatial filtering (LCMV)). The dipole algorithms produce an inverse solution corresponding to activity at a single point. In contrast, the current density methods produce a result that is an overall map of the current or activity levels associated with the electrical activity recorded at each electrode position. These two different types of localization algorithms were analyzed separately. All inverse algorithms are available in the CURRY software package with the exception of the LCMV spatial filter technique, which was added as a plug-in module. Nomenclature used to identify Waves I-IV is consistent with the literature (Møller and Burgess 1986; Alegre et al. 2001). The generators against which the results were compared for waves III and IV (figure 1) were the contralateral superior olivary nucleus (SO) and the lateral lemniscus (LL), respectively (figure 2a). As the LCMV spatial filter requires a relatively long epoch (section LORETA and LCMV), the LCMV "generator" was identified as the pathway from the superior olivary complex (SO) along the lateral lemniscus (LL) up to the medial geniculate body (MG). Generator locations and waves III-IV in the recordings were identified according to those previously described in studies about primate BAEP generators (Jewett and Williston 1971; Allen and Starr 1978; Stockard et al. 1979; Buchwald 1983; Legatt et al. 1986; Møller and Burgess 1986; Urasaki et al. 1995; Alegre et al. 2001). There is some controversy in the literature in regard to the precise location of these generators in different primate species; in this study we used the set of generators specifically for the rhesus monkey as described by Møller and Burgess (1986), and we evaluated effects of alternative source specification in the discussion. Dipole Analysis and MUSIC We applied a moving dipole algorithm to compute a single dipole for each time sample. The fixed dipoles al-
Fontanarosa et al.
gorithm and MUSIC each compute a single dipole that is fixed in space for an epoch of the waveform. The epoch’s time range was selected as the peak corresponding to wave III or IV. The number of samples in each analysis varied depending on the wave. The time window analyzed was chosen as a symmetrical window around the peak. The dipole results were constrained to a single dipole whose position coincided with the maximum value in the Global Field Power calculated in CURRY. The distance between the coordinates for the result of a wave and the coordinates for its putative generator in the MRI image was calculated for each of the axes (x, y, and z) and represented as dx, dy, and dz. The Euclidian distance dr (=[dx2+dy2+dz2]1/2) was used for overall evaluation. The data were analyzed with SPSS for Windows (SPSS, Chicago, IL). The generator coordinates were subtracted from the calculated dipole coordinates, and the resulting deviation scores (dx, dy, dz, and dr) were the variables used in all subsequent analyses. Descriptive statistics were calculated to provide information about the errors in localization for each coordinate, algorithm, and wave. Intra-class correlation coefficients (ICC) were computed for each pair of reproduced results (trial 1 and trial 2 for each coordinate, algorithm, and wave combination) to assess reliability of source localization. The ICC values were calculated using the Two-way Random Effects Model with Absolute Agreement Definition using the SPSS software package. ICC reliability results were evaluated according to guidelines established for kappa statistics by Landis and Koch (1977). A three-way repeated measures analysis of variance (ANOVA) was used to evaluate differences in localization as a function of three within subject factors: coordinate (dx, dy, dz, and dr), algorithm (fixed dipole, moving dipole, and MUSIC), and wave (III and IV). LORETA and LCMV The LORETA analysis was conducted over the same time range as the fixed and moving dipole algorithms for each wave. The points analyzed corresponded to the same points analyzed for the dipole examinations: i.e. the "peak" point as determined by the maximum of the Global Field Power. For the LCMV spatial filter, we required a minimum of 64 sample points (corresponding to an epoch of 3.75 ms) following the recommendation by Van Veen et al (1997) to include at least 2N samples for the analysis of an N channel data set. The duration of the BAEP is 5-6 ms from stimulus onset to the end of the brainstem part of the waveform. Because of the epoch constraints imposed by the LCMV spatial filter, the time range for LCMV began immediately before wave III, and ended 65-75 sample points (3.8-4.4 ms) later (figure 1). For this technical reason, the LCMV output may also con-
Source Localization of BAEP
tain locations corresponding to activities occurring after waves III and IV. Unlike the dipole techniques, current density methods assign a strength to every point in the brain volume. In general, they identify a cluster of points with strengths considerably above the rest of the activity, not a single point. Given the high SNR of the activity localized and the spatially compact nature of the source generator, the current density volume solution, unlike the noise, should be spatially stable if the source is modeled well by the algorithm. We therefore defined the source region as those points above a threshold of activity; the threshold itself was chosen so that raising its value by 5% of the maximum value did not significantly change the source region. This 5% change in setting the threshold was established empirically as the minimum value that generated stable results across all cases. The volume determined with this threshold will be referred to as the clipped result (CR). To obtain a semi-quantitative evaluation for LORETA and LCMV, a series of questions were asked of each result. We examined whether the clipped result (CR) includes parts of the generator pathway (GP) and the brainstem (B): i.e. we determined if the statements 1) GP⊂CR, and 2) B∩CR were true. Sensitivity and selectivity of the localization were further addressed by the examination of two statements: 3) A source DETECT was scored when part of the clipped result was within 2 cm of the generator pathway. 4) A ¬FALSE_DETECT (¬ denotes the logical negation symbol "not") was scored when the entire clipped result was within 2 cm of the generator pathway. Specifically, those two questions addressed whether the minimum (MIN) and maximum (MAX) distance between the elements of the CR and GP ( CR - GP ) exceeded a 2 cm threshold: MIN CR - GP < 2 cm (DETECT), and MAX CR - GP < 2 cm (¬ FALSE_DETECT), respectively. The fifth question evaluated if statements 3 and 4 were both true, i.e., if detection without false detection occurred in a single trial. The questions were designed such that a positive response to each question would indicate a favorable aspect of the localization result. The answers to these questions were analyzed for reliability by calculating the proportion of agreement between replicates for each subject.
Results The 32 BAEP superimposed waveforms (figure 1) were inspected for peaks I-IV to identify the relevant segments of the waveform to be localized. A set of localization results is depicted in figure 2. The epochs indicated by the horizontal lines in figure 1 were used as the input for the dipole analysis (dipoles, figure 2a), spatial filter (LCMV, figure 2b), and LORETA (figure 2c). The 3D reconstructions of the brainstem and brain in posterior, lat-
103
eral and top views, with the superimposed dipole solutions are shown in the bottom part of figure 2a: the yellow and red sticks indicate the moving and stationary dipole solutions respectively; the blue symbol indicates the solution from the MUSIC algorithm. The results from the spatial filter and LORETA (figure 2b and 2c) are displayed as the clipped areas superimposed on the 3D brain reconstruction. In this example, all dipole methods detected the source in the right side of the superior part of the brainstem, close to the generators indicated by SO (Superior Olivary Complex), LL (Lateral Lemniscus), IC (Inferior Colliculus), and MG (Medial Geniculate body). The errors in localizing deep brain sources associated with the dipole source localization algorithms are shown for each subject in table I. All three dipole algorithms used produced similar results. The average error of the dipole algorithms (dr) was between 22.3 mm and 28.0 mm (average: 25.1 mm). The average errors in each direction for each wave are depicted as a bar graph in figure 3. Overall, the z-direction (superior-inferior) contributed most to the error (figure 3a,b). However, in individual cases the maximum contribution of error in localizations could be attributed to other directions. Furthermore, it is interesting to note that the average error is in the positive x-direction, the negative y-direction, and the positive z-direction for each algorithm and for both waves III and IV. These results indicate that the overall error vector points to the center of the head. In the dipole and MUSIC localizations, the only significant (p<0.05) effect from a 3(algorithm) × 3(coordinate) × 2(wave) repeated measures analysis of variance (ANOVA) was the coordinate main effect (F(2,8)=12.4; p=0.004). The coordinate main effect indicates that the difference in the errors associated with the x, y, and z coordinates was statistically significant. Specifically, the greatest errors were observed along the z-axis. The reliability of the dipole results as indicated by intra-class correlation coefficients (ICCs) for the Euclidian distances dx, dy, dz, and dr revealed that the dipole algorithms in this study showed reliability that was good (ICC > 0.4) to excellent (ICC > 0.75) (Landis and Koch 1977). The only exceptions were traced to subject 3 who had inconsistent wave IV localization results. Table II shows the results of the current density analyses, LORETA and LCMV. Part of the brainstem was contained in the current density localization volume solution in 60% of the wave III LORETA results, and in 70% of the wave IV LORETA results. However, 90% of the LORETA results overall showed some aspects of the localization result that were distant (> 2 cm) from the auditory pathway, and none of the LORETA results included a result entirely in the brainstem. In 80% of the cases LORETA produced a current density within 2 cm of the pathway, while the generator was only included in the localization result in 30%
104
Fontanarosa et al.
Table I. Error in the dipole coordinate in mm, for the inter-aural axis (left-positive, dx), inion-nasion axis (inion-positive, dy), up-down axis (up-positive, dz), and Euclidian distance (dr). The SNR values shown correspond to the maximum signal to noise ratio (SNR) of the sample window analyzed for each wave. The error for the moving dipole, fixed dipole, and MUSIC analysis are presented for each monkey (1-5), each wave (III and IV), and each replicate recording. Moving Dipoles
Fixed Dipoles
MUSIC
SNR
Subject
Wave
dx
dy
dz
dr
dx
dy
dz
dr
dx
dy
dz
dr
1
III
7.7
-7.2
24.7
26.9
8.8
-6.0
25.0
27.2
10.3
-6.7
26.7
29.4
5.9
III
11.5
-5.2
25.5
28.5
13.1
-6.3
18.4
23.4
14.3
-4.7
22.7
27.2
7.4
III
11.9
-7.6
13.6
19.6
9.9
-8.2
14.3
19.2
9.1
-6.2
16.6
19.9
9.1
III
8.5
-3.4
21.2
23.1
7.4
-5.1
22.0
23.8
9.0
-10.2
18.5
23.0
10.7
III
0.8
14.6
38.4
41.1
5.4
14.7
35.7
39.0
9.7
16.0
28.3
33.9
8.9
III
1.5
24.5
31.5
39.9
1.4
22.5
34.2
41.0
9.7
18.0
28.3
34.9
11.5
III
8.6
-19.5
7.2
22.5
2.1
-16.0
11.9
20.1
8.3
-18.1
4.7
20.5
11.9
III
3.5
-18.6
4.2
19.4
-1.5
-18.3
-5.0
19.0
4.3
-20.1
2.7
20.7
12.9
III
1.4
-22.9
14.9
27.4
1.0
-23.0
14.8
27.4
5.7
-23.6
14.6
28.3
6.0
III
3.8
-24.5
19.2
31.4
1.4
-23.1
15.0
27.6
3.7
-27.6
24.6
37.2
11.4
Average
5.9
-7.0
20.0
28.0
4.9
-6.9
18.6
26.8
8.4
-8.3
18.8
27.5
9.6
95% CI
2.6
10.0
6.6
4.7
2.9
9.4
7.3
4.8
1.9
9.6
5.7
4.0
1.6
2 3 4 5
1 2 3 4 5
SEM
1.3
5.1
3.3
2.4
1.5
4.8
3.7
2.4
1.0
4.9
2.9
2.0
0.8
Median
5.8
-7.4
20.2
27.1
3.8
-7.3
16.7
25.5
9.1
-8.5
20.6
27.8
9.9
Range
11.1
49.0
34.2
21.7
14.6
45.6
40.7
21.9
10.6
45.6
25.6
17.2
7.0
IV
11.8
0.0
17.3
20.9
11.5
-0.8
15.6
19.4
11.3
-0.5
16.7
20.2
7.5
IV
14.7
-1.8
11.4
18.7
16.7
-5.0
8.7
19.5
13.3
-0.5
10.7
17.1
8.1
IV
5.8
-6.2
10.3
13.3
7.7
-10.0
7.9
14.9
8.1
-8.4
9.8
15.2
8.8
IV
7.7
-10.6
19.6
23.6
6.3
-10.3
18.2
21.8
8.0
-14.4
-16.3
23.2
10.2
IV
-2.5
-4.0
26.4
26.8
24.2
-3.0
1.8
24.5
5.7
8.8
20.8
23.3
8.5
IV
14.4
26.7
22.6
37.8
27.6
20.4
19.0
39.2
17.7
16.8
20.8
32.1
9.4
IV
4.3
-12.9
7.0
15.3
2.5
-12.5
3.8
13.3
6.3
-14.1
0.5
15.5
12.2
IV
0.4
-12.1
2.4
12.3
-1.7
-12.3
-0.7
12.4
2.3
-14.1
-3.5
14.7
13.3
IV
4.4
-27.5
16.3
32.3
5.4
-26.5
17.7
32.3
9.7
-25.3
11.6
29.5
6.2
IV
3.9
-23.8
19.9
31.3
5.4
-26.7
17.4
32.3
3.7
-27.3
17.6
32.7
10.8
Average
6.5
-7.2
15.3
23.2
10.6
-8.7
10.9
23.0
8.6
-7.9
8.9
22.3
9.5
95% CI
3.5
9.2
4.6
5.4
5.9
8.3
4.7
5.6
2.9
8.8
7.4
4.3
1.3
SEM
1.8
4.7
2.3
2.7
3.0
4.3
2.4
2.9
1.5
4.5
3.8
2.2
0.7
Median
5.1
-8.4
16.8
22.3
7.0
-10.2
12.2
20.7
8.1
-11.3
11.2
21.7
9.1
Range
17.2
54.2
24.0
25.5
29.3
47.1
19.7
26.8
15.4
44.1
37.1
18.0
7.1
Source Localization of BAEP
105
Figure 3. Bar graphs depicting the average localization errors for each algorithm in the x (inter-aural axis left-positive), y (inion-nasion axis, inion-positive), and z (up-down axis, up-positive) directions for (a). wave III and (b). wave IV. Error bars indicate the SEM.
and 20% of the results analyzed for wave III and wave IV, respectively. The results usually included several different regions of the brain. LORETA results were similar for wave III and wave IV. LCMV included the generator in the localization in 70% of the trials and in 60% of the trials, no parts of the localized regions was distant (> 2cm) from the auditory pathway. LCMV included the position of the brainstem in 9 out of 10 localization results. The source was detected to within 2 cm without a false detection (as indicated in statement 5 in table II) in 1 out of 10 cases for each LORETA analysis and in 6 out of
10 cases for LCMV. The reliability of the localization results is characterized by the Agreement Ratio reported in Table II. The localization results were in agreement for both replicates most consistently for questions 1 and 3, followed by questions 2 and 4.
Discussion In this study, we used waves III and IV of the brainstem auditory evoked potential (BAEP) to evaluate the performance of different source localization algo-
106
Fontanarosa et al.
Table II. Summary of the results obtained using LORETA and LCMV. Five true or false statements were used to evaluate localization performance: 1. The Generator Pathway (GP) is a subset of (⊂) the clipped localization result (CR). 2. The volume corresponding to the Brainstem (B) in the model intersects with (∩) the clipped localization result (CR). 3. The clipped current density localization volume detected a region within 2 cm of the known generator. 4. The clipped current density localization volume included only regions of the brain within 2 cm of the known generator (i.e. ¬ FALSE DETECT, where ¬ = logical "not"). 5. Questions 3 and 4 were both true, indicating that the generator was localized to 2 cm accuracy without any outliers for a given trial. To evaluate our findings, we looked at the True Score Ratio (number of true responses/ number of trials evaluated) and the Agreement Ratio (which cites number of agreements between replicate trials within a subject/ number of subjects) for each question.
LORETA wave III
LORETA wave IV
LCMV combined waves
True Score Ratio
Agreement Ratio
True Score Ratio
Agreement Ratio
True Score Ratio
Agreement Ratio
1. GP ⊂ CR
3/10
4/5
2/10
5/5
7/10
4/5
2. B ∩ CR
6/10
1/5
7/10
2/5
9/10
4/5
3. Detect
8/10
3/5
8/10
3/5
10/10
5/5
4. ¬ False Detect
1/10
4/5
1/10
4/5
6/10
1/5
5. Both 3 and 4 true.
1/10
0/5
1/10
0/5
6/10
1/5
rithms. We considered this part of the BAEP as a worst-case scenario for evaluation of localization algorithms because the signal is generated deep within the brain. Using 32 channel recordings and individualized boundary element models (BEM), the dipole localization techniques applied to these data had centimeter accuracy (~ 2.5 cm) (table I). The average localization errors were least along the x-axis (left-right) and greatest along the z-axis (superior-inferior). The intra-class coefficients (ICCs) that were computed suggest that the dipole solutions for these sources are consistent. Many ICCs indicate excellent consistency, and the measures that were unreliable were largely due to a single outlier (subject 3). These reliability estimates include not only the reliability of the algorithm, but also reflect reliability of the physiological response. The average errors in the dipole localization results were similar for waves III and IV (table I, figure 3). Average localizations were above (dz), contralateral (dx), and in anterior to (dy) the "gold standard", the putative generator. Interestingly, these errors for both waves point towards the center of the head model. Potentially, a localization error may be due, in part, to mis-specification of the real sources. In this study we used the sources described for the rhesus monkey by Møller and Burgess (1986). Other studies (Legatt et al. 1986; Alegre et al. 2001) discussed the source of wave IV of the BAEP as being more complex, with multiple generators contributing to each wave as seen on the surface. Legatt and co-workers (1986) described the presence of the following additional activities for wave IV (their wave 7): the ipsilateral lateral lemniscus, the contralateral inferior colliculus, and both
medial geniculate nuclei. Assuming these activities also play a role in the rhesus monkey, some of the error in wave IV in the x, y, and z-directions may be explained by the contribution of this subset of generators. However, there is little support from the literature for additional generators for wave III. Therefore, our finding that the average errors for both waves III and IV are similar in sign and magnitude, suggests that the localization errors recorded were most likely due primarily to other noise factors and modeling errors, rather than source mis-specifications. Recent studies of the accuracy of dipole inverse solution algorithms (Krings et al. 1999) using a 4-shell spherical model and a standard 10-20 array of 21 electrodes showed an average localization error of 17 mm, and an error of 13 mm when using 41 electrodes. Fuchs et al. (2002) compared BEM and spherical models for randomly distributed sources in the head using 71 electrodes and obtained an average localization error of 6.9 mm for a standard BEM model and 10.8 mm for a spherical shell model. Our localization errors averaged 25.1 mm for the dipole-based techniques. Our localization error is greater but of the same order of magnitude as previously reported localizations. Taking into account that we used 32 channels, and because the generators in this study are deep brain sources, our finding are not unexpected and agree well with the previously reported data on detection accuracy. Our study confirms that current algorithms and head models can be used to localize generators to within several centimeters and extends this finding to deep brain sources. On the other hand, it is
Source Localization of BAEP
clear there is room for improvement; an increase in accuracy of one order of magnitude is desirable to make localization procedures more useful both clinically and for many research applications. The output of the LORETA and LCMV algorithms is a current distribution profile. Therefore it is difficult to compare the performance of the dipole techniques directly with the results obtained with the LORETA or LCMV techniques. Current density algorithms calculate a value for a set of points distributed evenly throughout the brain volume. To obtain meaningful results for this study, we developed a rule to define the source volume by clipping the result. As can be seen in table II, the source volume (the clipped result, CR) frequently contained the correct location for the generator pathway (GP); LORETA generated a clipped result within 2 cm of the source in 80% and LCMV in 100% of the cases (table II). Although the LORETA and LCMV results often included the correct location, we obtained false detections in 90% and 40% of the cases respectively: i.e. the solution also included regions distant (>2 cm) from the auditory pathway. The LCMV result is for the longer epoch including waves III and IV; LORETA performed similarly for both waves. The relatively poor performance of LORETA on our data set is not unexpected because this method was designed to select the smoothest of all possible three-dimensional current distributions. The nature of the BAEP sources where local activity conducts through the brainstem is not very compatible with this constraint. However, the finding of the low reliability and limited accuracy of LORETA for the localization of deep and spatially compact sources is notable because it contrasts with existing descriptions of mm resolution of LORETA (Pascual-Marqui 1999; Pascual-Marqui 2002). This difference is perhaps due to methodological differences or to the different type of signals evaluated in this study. To estimate the spatial filter for the LCMV method, we had to use a relatively long epoch (figure 1) and in our approach, following the procedure described by Van Veen et al. (1997), we used this same epoch to perform the localization. Since this epoch includes multiple deep sources within a small interdistance, it is to be expected that a relatively wide spread volume for the CR was found (Van Drongelen et al. 1996, their figure 1). Because BAEP sources are briefly active and spatially compact, our findings confirm the expectation that the dipole techniques perform best for the sources localized in this study. The discussion of algorithm performance emphasizes an important point about source localization of surface recorded neural potentials. Solutions are ill-posed and require constraints to identify a "correct" solution. Consequently, no single algorithm is appropriate for all applications, and the better matched the algorithm constraints to the localization problem, the better the results. Therefore, researchers and clinicians armed with a tool
107
box of algorithms might benefit from an eclectic approach. An important task is to specify the sources to be localized in sufficient detail that an appropriate source localization algorithm can be selected. For situations in which there is insufficient information to adequately "constrain" potential algorithms, a multistage localization procedure may prove useful. For example, a decomposition algorithm or an algorithm like the LCMV spatial filter may be used initially to identify the number of dipoles responsible for the recorded activity. Informed by the number of sources likely involved, a second stage dipole algorithm may perform optimally.
References Alegre, M., Gurtubay, I.G., Iriarte, J., Cirodia, E., Manrique, M. and Artieda, J. Brainstem auditory evoked potentials (BAEPs) in the cynomologous macaque monkey. Equivalence with human BAEPs and proposal of a new nomenclature. Hear Res., 2001, 151: 115-120. Allen, A.R. and Starr, A. Auditory brain stem potentials in monkey (M. mulatta) and man. Electroencephalogr. Clin. Neurophysiol., 1978, 45: 53-63. American Electroencephalographic Society Guidelines in Electroencephalography, Evoked Potentials, and Polysomnography. J. Clin. Neurophysiol., 1994, 11: 111-113. Buchwald, J.S. Auditory evoked responses in clinical populations and in the cat. Auris. Nasus. Larynx., 1983, 10: 87-95. Fuchs, M., Kastner, J., Wagner, M., Hawes, S. and Ebersole, J.S. A standardized boundary element method volume conductor model. Clin. Neurophysiol., 2002, 113: 702-712. Grandori, F. Dipole localization methods (DLM) and auditory evoked brainstem potentials. Rev. Laryngol. Otol. Rhinol., 1984, 105: 171-178. Hoekema, R., Wieneke, G.H., Leijten, F.S., van Veelen, C.W., van Rijen, P.C., Huiskamp, G.J., Ansems, J., and van Huffelen, A.C. Measurement of the conductivity of skull, temporarily removed during epilepsy surgery. Brain Topogr., 2003, 16: 29-38. Jewett, D.L., Romano, M.N., and Williston, J.S. Human auditory evoked potentials: possible brain stem components detected on the scalp. Science, 1970, 13: 1517-1518. Krings, T., Chiappa, K.H., Cuffin, B.N., Cochius, J.I., Connolly, S. and Cosgrove, G.R. Accuracy of EEG dipole source localization using implanted sources in the human brain. Clin. Neurophysiol., 1999, 110: 106-114. Landis, J.R. and Koch, G.G. The measurement of observer agreement for categorical data. Biometrics, 1977, 33: 159-174. Lasky, R.E., Maier, M., Snodgrass, E., Hecox, K. and Laughlin, N. The effects of lead on otoacoustic emissions and auditory evoked potentials in monkeys. Neurotoxicol. Teratol., 1995, 17: 633-644. Le, J. and Gevins, A. Method to reduce blur distortion from EEG’s using a realistic head model. IEEE Trans. Biomed. Eng., 1993, 40: 517-528. Legatt, A.D., Arezzo, J.C. and Vaughan, H.G. Short-latency au-
108
ditory evoked potentials in the monkey. II. Intracranial Generators. Electroencephalogr. Clin. Neurophysiol., 1986, 64: 53-73. Møller, A.R. and Burgess, J. Neural generators of the brain-stem auditory evoked potentials (BAEPs) in the rhesus monkey. Electroencephalogr. Clin. Neurophysiol., 1986, 65: 361-372. Mosher, J.C., Lewis, P.S. and Leahy, R.M. Multiple dipole modeling and localization from spatio-temporal MEG data. IEEE Trans. Biomed. Eng., 1992, 39: 541-557. Nunez, P.L. Electric fields of the brain: the neurophysics of EEG. Oxford University Press, Oxford, 1981. Pascual-Marqui, R.D., Michel, C.M., and Lehmann, D. Low resolution electromagnetic tomography: a new method for localizing electrical activity in the brain. Int. J. Psychophysiol., 1994, 18: 49-65. Pascual-Marqui, R.D. Review of methods for solving the EEG i n v e r s e p r ob l e m . I n t e r n at i o n al J o u r n a l o f Bioelectromagnetism, 1999, 1: 75-86. Pascual-Marqui, R.D., Esslen, M., Kochi, K. and Lehmann, D. Functional imaging with Low Resolution Brain Electromagnetic Tomography (LORETA): A Review. Methods and Findings in Experimental and Clinical Pharmacology, 2002, 24: 91-95.
Fontanarosa et al.
Scherg, M. Spatio-temporal modelling of early auditory evoked potentials. Rev. Laryngol. Otol. Rhinol., 1984, 105: 163-170. Scherg, M. and von Cramon, D. A new interpretation of the generators of BAEP waves I-V: results of a spatio-temporal dipole model. Electroencephalogr. Clin. Neurophysiol., 1985, 62: 290-299. Stockard, J.J., Stockard, J.E. and Sharbrough, F.W. Brainstem auditory-evoked responses. Arch. Neurol., 1979, 36: 597-598. Urasaki, E., Uematsu, S. and Lesser, R.P. Brainstem auditory evoked potentials recorded around the human upper brainstem area. Neurol. Med. Chir., 1995, 35: 353-359. Van Drongelen, W., Yuchtman, M., Van Veen, B.D. and Van Huffelen, A.C. A spatial filtering technique to detect and localize multiple sources in the brain. Brain Topogr., 1996, 9: 39-49. Van Veen, B.D., van Drongelen, W., Yuchtman, M. and Suzuki, A. Localization of brain electrical activity via linearly constrained minimum variance spatial filtering. IEEE Trans. Biomed. Eng., 1997, 44: 867-880. Witt, J.C., Towle, V.L., Bolanos, J. and Spire, J.P. Tetrahedral recording of 3-D BAEPs: evidence for the centered dipole model. Electroencephalogr. Clin. Neurophysiol., 1991, 80: 551-560.
