- Email: [email protected]
Predicting coma and other low responsive patients outcome using event-related brain potentials: A meta-analysis
Clinical Neurophysiology 118 (2007) 606–614 www.elsevier.com/locate/clinph
Predicting coma and other low responsive patients outcome using event-related brain potentials: A meta-analysis J. Daltrozzo
a,b,c,*
, N. Wioland
a,b
, V. Mutschler
a,b
, B. Kotchoubey
c
a
c
Department of Neurology, University Hospital of Strasbourg, France b CNRS UPS 858, Louis Pasteur University of Strasbourg, France Department of Medical Psychology and Behavioral Neurobiology, University of Tu¨bingen, Germany See Editorial, pages 477–479
Abstract Objective: A meta-analysis was performed to estimate the predictive power (odd ratio, OR) for awakening of auditory event-related potential (ERP) components in low responsive patients with stroke or hemorrhage, trauma, anoxic, post-operative, and metabolic encephalopathy etiologies. Methods: We reviewed MEDLINE and analyzed citations for retrieved articles. Logistic regressions were applied on patient samples (Glasgow Coma Scale <12) across and for separate etiologies. Results: For stroke and hemorrhage the ORs with 95% confidence intervals were: 2.05 [1.12–3.75] (N100), 4.47 [1.92–10.44] (MMN), 10.29 [2.00–52.79] (P300), for trauma: 1.63 [0.70–3.80] (N100), 4.72 [1.35–16.44] (MMN), 12.89 [4.82–34.43] (P300), anoxic: 8.03 [2.83–22.75] (N100), 15.50 [4.27–56.26] (MMN), 5.93 [2.38–14.77] (P300), post-operative: 10.66 [1.98–57.50] (N100), metabolic encephalopathy: 2.12 [0.34–13.13] (N100), 3.60 [0.28–46.36] (MMN), 7.71 [0.75–79.77] (P300), and all etiologies: 2.85 [1.91–4.27] (N100), 6.53 [3.55–12.01] (MMN), and 8.79 [4.88–15.83] (P300). Based on six N100 studies (N = 548 patients), five MMN studies (N = 470), and six P300 studies (N = 313), the N100, MMN, or P300, when present, significantly predicted awakening, P300 and MMN being significantly better predictors than N100. Conclusions: The MMN and P300 appear to be reliable predictors of awakening. Significance: The prognostic assessment of low responsive patients with auditory ERP should take into account both MMN and P300. Ó 2006 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved. Keywords: Coma; Prognosis; ERP; N100; MMN; P300
1. Introduction Event-related potentials (ERPs) can be used to predict awakening from coma. The prognostic power of ERPs has been investigated in several prospective cohort studies. Reuter and Linke (1989) were the first to report the presence of an auditory ERP component, the P300 in four coma patients and they indicated that all these patients survived. Their findings were subsequently confirmed by Gott et al. (1991) and other authors. Most often, the visual detection of the P300 was performed only if an earlier ERP *
Corresponding author. Tel.: +33 388 127 414; fax: +33 388 116 568. E-mail address: [email protected] (J. Daltrozzo).
component, which is called the N100 and intended to indicate mainly a sensory cortical processing, was present (De Giorgio et al., 1993). Another ERP component, the mismatch negativity (MMN), can also be recorded in comatose patients and is also an indicator of prognosis (Kane et al., 1993). Both MMN and P300 reflect discrimination of sounds. The MMN is modality-specific (i.e., generated only by auditory stimuli), due to an automatic mechanism and depends on the physical distance between the two discriminated sounds (e.g., a difference of frequency, sound duration, or intensity). Conversely, the P300 is independent of sensory modality, can easily be modulated by arousal and attention, and does not depend on the physical difference between stimuli. In an attempt to better evaluate the prognostic
1388-2457/$32.00 Ó 2006 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.clinph.2006.11.019
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
power of these ERP components, later studies used larger samples of patients (e.g., Fischer et al., 2004; Gue´rit et al., 1999; Mutschler et al., 2001). They have confirmed that these ERP components, and particularly the MMN, are reliable predictors of awakening from coma. Fischer et al. (2004) were the first to demonstrate the predictive value of the MMN statistically with the odd ratio (OR) measure. The goal of our study is to perform meta-analyses in order to compare the prognostic power of several ERP components on a large pool of patients and with a statistical measure already used for such purpose, the OR. 2. Methods 2.1. Systematic search Two of us independently reviewed MEDLINE from January 1, 1989 to June 13, 2006, with key words: coma and (ERP or N1 or MMN or mismatch negativity or P300 or prognostic and auditory event-related potentials) and analyzed citations for retrieved articles. 2.2. Eligibility criteria The sources were independently selected by two of us on the basis of the following criteria: (1) the last report of each research team to avoid duplicate data or reports of the same team if populations were different according to the etiology of coma; (2) reports in English or in French; (3) at least one of ERP components N1, MMN and P300 was recorded; (4) patient’s Glasgow Coma Scale (GCS) score lower than 12 at the time of the ERP recording; (5) outcome of coma defined on the basis of the Glasgow Outcome Score (GOS, 1: death, 2: persistent vegetative state, 3: severe disability, 4: moderate disability, and 5: good recovery, Jennett and Bond, 1975; Sinson et al., 2001) and compared with the presence of the ERP component for each group of patients classified by etiology. We contacted authors to obtain further data if required for inclusion. 2.3. Extracted data From the selected studies the following data were extracted for each etiology: the number of True Positives (TP, number of patients with the ERP component who awakened, i.e., with GOS larger than 2), of False Negatives (FN, number of patients without the ERP component and with GOS >2), True Negatives (TN, number of patients without the ERP component and with GOS 62), and False Positives (FP, number of patients with the ERP component and with GOS 62). 2.4. Statistical analysis The sensitivity (TP/(TP + FN)), specificity (TN/(TN + FP)), positive predictive value (TP/(TP + FP)), and nega-
607
tive predictive value (TN/(TN + FN)) for awakening were estimated for each selected study. Exact 95% confidence intervals (CIs) were estimated using the binomial distribution (Fischer et al., 2006) with Microsoft Excel 97 SR-2. The prognostic power of ERPs was estimated with a logistic regression analysis. The logistic regression analysis is a statistical model used to study the relationships between explanatory qualitative (or quantitative) variables and a dependent qualitative variable. These relationships are estimated by the model in terms of OR. In the present study, all variables except the etiology are qualitative with two modalities, i.e., binary variables. The dependent variable is the outcome of the patient. It is equal to 1 if the patient subsequently awakens, i.e., if the GOS is larger than 2, and equal to 0 otherwise. The explanatory variables are the presence (variable equal to 1) or absence (variable equal to 0) of the N100, the MMN, or the P300, and the patient etiology. The etiology was considered as a nominal variable with five modalities for the N100 and the MMN and four modalities for the P300. The etiology with the largest number of patients was chosen as the reference modality for this explanatory variable (e.g., Fischer et al., 2004) in order to have the most stable regression model and thus the best OR estimates for the etiology variable. In addition to the logistic regression on the data of all patients (i.e., across all etiologies), the predictive power for awakening of each ERP component was estimated with patient samples grouped by etiology. For the relationship between the ERP variable and the outcome, an OR significantly larger than 1, i.e., the lower limit of the estimated 95% Confidence Interval (CI) of the OR is larger than 1, indicates that the presence of the ERP component is a significant predictor of awakening. Since ERP components are sensitive to vigilance and levels of sedation (Gue´rit, 2004), an OR significantly larger than 1 does not allow to conclude that the absence of the ERP component predicts significantly nonawakening (i.e., death or vegetative state). An OR smaller or not significantly different from 1 (i.e., the CI contains the value 1) indicates that the ERP is not predictive of outcome. For the relationship between an etiology (i.e. one modality of the etiology variable different from the reference modality) and the outcome, an OR significantly larger than 1 indicates that the etiology is significantly more related to awakening than the reference etiology. An OR significantly smaller than 1 indicates that the etiology is significantly less related to awakening than the reference etiology. An OR not significantly different from 1 (i.e., the CI contains the value 1) indicates that the relationship between the etiology and the outcome does not differ significantly from the relationship between the reference etiology and the outcome. The analysis of heterogeneous studies (with a randomeffects model) remains a controversial debate (Fleiss and Gross, 1991, p.129). Thus, we conducted logistic regressions with a fixed-effects model. This model assumes that the aggregated studies are homogeneous (i.e., the effect size, here the log OR, does not vary very much between studies).
608
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
Table 1 Selected sources Source *
Daltrozzo et al
ERP
Etiology
MMN
Anoxic Encephalopathy Stroke and hemorrhage Anoxic Encephalopathy Stroke and hemorrhage Anoxic Encephalopathy Stroke and hemorrhage Encephalopathy Post-operative Stroke and hemorrhage Trauma Encephalopathy Post-operative Stroke and hemorrhage Trauma
N100
P300
Fischer et al. (2004)
MMN
N100
N
TP
FN
TN
FP
28 4 12 29 4 11 28 4 12
0 0 1 3 0 0 1 0 0
4 0 0 1 0 2 3 0 1
20 3 11 22 3 6 23 4 11
4 1 0 3 1 3 1 0 0
7 54 124 96 7 54 125 96
1 14 33 20 4 32 63 50
4 30 53 63 1 12 23 33
2 10 31 10 0 8 20 7
0 0 7 3 2 2 19 6
Fischer et al. (2006)
MMN N100
Anoxic Anoxic
62 62
12 15
8 5
42 36
0 6
Grapperon et al. (2004)
N100
Anoxic Encephalopathy Stroke and hemorrhage Trauma Anoxic Encephalopathy Stroke and hemorrhage Trauma
2 1 21 9 2 1 21 9
1 0 12 3 0 0 4 0
0 1 8 3 1 1 16 6
0 0 1 0 0 0 1 2
1 0 0 3 1 0 0 1
P300
Gue´rit et al. (1999)
N100 P300
Anoxic Anoxic
21 22
10 5
1 6
7 11
3 0
Kane et al. (1996)
MMN P300
Trauma trauma
53 54
22 33
17 4
14 15
0 2
Lew et al. (2003)
N100 P300
Trauma Trauma
22 22
12 7
4 9
6 6
0 0
Mutschler et al. (2001)
N100
Anoxic Encephalopathy Stroke and hemorrhage Trauma Anoxic Encephalopathy Stroke and hemorrhage Trauma
77 19 40 10 77 19 40 10
11 7 13 5 9 6 12 5
3 0 0 0 5 1 1 0
32 2 6 0 49 5 13 2
31 10 21 5 14 7 14 3
P300
Naccache et al. (2005)
MMN MMN MMN
Encephalopathy Stroke and hemorrhage Trauma
4 20 6
2 6 1
1 4 2
1 9 3
0 1 0
Signorino et al. (1997)
P300
Trauma
25
13
5
5
2
ERP, event-related potential; N, number of patients; TP, true positives (i.e., number of patients with the ERP component and with GOS >2); FN, false negatives (i.e., number of patients without the ERP component and with GOS >2); TN, true negatives (i.e., number of patients without the ERP component and with GOS 62); FP, false positives (i.e., number of patients with the ERP component and with GOS 62). MMN, mismatch negativity. Note 1. The GOS was estimated: 1 month after coma onset in Naccache et al. (2005); 1.5 months in Mutschler et al. (2001); 3 months in Daltrozzo et al* and in Kane et al. (1996); 6 months in Lew et al. (2003) and in Signorino et al. (1997); 12 months in Fischer et al. (2006) and Fischer et al. (2004); and just before the patient leaves the ICU in Grapperon et al. (2004). The time of GOS assessment was not reported in Gue´rit et al. (1999). Note 2. The mean age of the patients was 49 years [1–93] (age range was not reported in Signorino et al. (1997)). In most sources the age range was very large (e.g., [8–93] in Fischer et al. (2004); [1–80] in Kane et al. (1996); [18–89] in Gue´rit et al. (1999)). Since, the patients’ age was not reported separately for TP, FN, TN, and FP by etiology, it was not possible to include this variable in the regression. * Je´roˆme Daltrozzo, Norma Wioland, Ve´ronique Mutschler, Philippe Lutun, Albert Jaeger, Bartholomeus Calon, Alain Meyer, Thierry Pottecher, Simone Lang, and Boris Kotchoubey, Cortical information processing in coma and other low responsive states, in preparation.
Table 2 Sensitivity (Se), specificity (Spe), positive predictive value (Ppv), and negative predictive value (Npv) for awakening of selected sources Sources
MMN Daltrozzo et al Fischer et al. (2004) Fischer et al. (2006) Kane et al. (1996) Naccache et al. (2005) All P300 Daltrozzo et al Grapperon et al. (2004) Gue´rit et al. (1999) Kane et al. (1996) Lew et al. (2003) Mutschler et al. (2001) Signorino et al. (1997) All
N
TP
FN
TN
FP
Se
%
95% CI
Spe
44 282 62 33
3 149 15 16
3 69 5 12
31 35 36 1
7 29 6 4
3/6 149/218 15/20 16/28
50 68 75 57
12–88 62–75 51–91 37–76
31/38 35/64 36/42 1/5
21 22 146 610
10 12 36 241
1 4 3 97
7 6 40 156
3 0 67 116
10/11 12/16 36/39 241/338
91 75 92 71
59–100 48–93 79–98 66–76
44 281 62 53 30 470
1 68 12 22 9 112
4 150 8 17 7 186
34 53 42 14 13 156
5 10 0 0 1 16
1/5 68/218 12/16 22/39 9/16 112/298
20 31 60 56 56 38
1–72 25–38 48–93 40–72 30–80 32–43
44 33
1 4
4 24
38 3
1 2
1/5 4/28
20 14
1–72 4–33
22 54 22 146 25 346
95% CI
Ppv
82 55 86 20
66–92 42–67 71–95 1–72
3/10 149/178 15/21 16/20
7/10 6/6 40/107 156/272
70 100 37 57
35–93 54–100 28–47 51–63
34/39 53/63 42/42 14/14 13/14 156/172
87 84 100 100 93 91
5 33 7 32 13 95
6 4 9 7 5 59
11 15 6 69 5 147
0 2 0 38 2 45
5/11 33/37 7/16 32/39 13/18 95/154
45 89 44 82 72 62
17–77 75–97 20–70 66–93 47–90 53–69
38/39 3/5 11/11 15/17 6/6 69/107 5/7 147/192
%
%
95% CI
Npv
%
95% CI
30 84 71 80
7–65 77–89 48–89 56–94
31/34 35/104 36/41 1/13
91 34 88 8
76–98 25–44 74–96 0–36
10/13 12/12 36/103 241/357
77 100 35 68
46–95 74–100 26–45 62–72
7/8 6/10 40/43 156/253
88 60 93 62
48–100 26–88 81–99 55–68
73–96 73–92 92–100 77–100 66–100 85–95
1/6 68/78 12/12 22/22 9/10 112/128
17 87 100 100 90 88
0–64 78–94 74–100 85–100 56–100 81–93
34/38 53/203 42/50 14/31 13/20 156/342
89 26 84 45 65 46
75–97 20–33 71–93 27–64 41–85 40–51
97 60
87–100 15–95
1/2 4/6
50 67
1–99 22–96
38/42 3/27
90 11
77–97 2–29
100 88 100 64 71 77
72–100 64–99 54–100 55–74 29–96 70–82
5/5 33/35 7/7 32/70 13/15 95/140
100 94 100 46 87 68
48–100 81–99 59–100 34–58 60–98 59–75
11/17 15/19 6/15 69/76 5/10 147/206
65 79 40 91 50 71
38–86 54–94 16–68 82–96 19–81 65–77
ERP, event-related potential; N, number of patients; TP, true positives (i.e., number of patients with the ERP component and with GOS >2); FN, false negatives (i.e., number of patients without the ERP component and with GOS >2); TN, true negatives (i.e., number of patients without the ERP component and with GOS 62); FP, false positives (i.e., number of patients with the ERP component and with GOS 62). Se, Sensitivity for awakening = TP/(TP + FN); Spe, specificity for awakening = TN/(TN + FP); Ppv, positive predictive value for awakening = TP/(TP + FP); Npv, negative predictive value for awakening = TN/(TN + FN). For nonawakening specificity exchange with sensitivity and Ppv with Npv. MMN, mismatch negativity. Note. The GOS was estimated: 1 month after coma onset in Naccache et al. (2005); 1.5 months in Mutschler et al. (2001); 3 months in Daltrozzo et ala and in Kane et al. (1996); 6 months in Lew et al. (2003) and in Signorino et al. (1997); 12 months in Fischer et al. (2006) and Fischer et al. (2004); and just before the patient leaves the ICU in Grapperon et al. (2004). The time of GOS assessment was not reported in Gue´rit et al. (1999). a Je´roˆme Daltrozzo, Norma Wioland, Ve´ronique Mutschler, Philippe Lutun, Albert Jaeger, Bartholomeus Calon, Alain Meyer, Thierry Pottecher, Simone Lang, and Boris Kotchoubey, Cortical information processing in coma and other low responsive states, in preparation.
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
N100 Daltrozzo et ala Fischer et al. (2004) Fischer et al. (2006) Grapperon et al. (2004) Gue´rit et al. (1999) Lew et al. (2003) Mutschler et al. (2001) All
609
610
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
Table 3 Sources included in the meta-analysis ERP
Etiology
N
TP
FN
TN
FP
N100
Stroke and hemorrhage Trauma Anoxic Post-operative Encephalopathy
197 137 129 54 31
88 70 25 32 11
33 40 5 12 2
33 13 61 8 5
43 14 38 2 13
MMN
Stroke and hemorrhage Trauma Anoxic Post-operative Encephalopathy
156 155 90 54 15
40 43 12 14 3
57 82 12 30 5
51 27 62 10 6
8 3 4 0 1
P300
Anoxic Trauma Stroke and hemorrhage Encephalopathy
127 111 52 23
15 58 12 6
14 18 2 1
83 28 24 9
15 7 14 7
ERP, Event-Related Potential; N, Number of patients; TP, True Positives (i.e., number of patients with the ERP component and with GOS >2); FN, False Negatives (i.e., number of patients without the ERP component and with GOS >2); TN, True Negatives (i.e., number of patients without the ERP component and with GOS 62); FP, False Positives (i.e., number of patients with the ERP component and with GOS 62). MMN, mismatch negativity. Note: The GOS was assessed between 1 and 12 months after the incident for most of the selected sources (Table 1).
We retained only homogeneous studies using the following method. Studies with the highest weighted squared difference between the log OR of the individual study and the log OR of the aggregated studies were considered as outliers and thus removed if the Q test for homogeneity of studies was significant (Fleiss and Gross, 1991). Q is a measure of the heterogeneity between a set of studies. If N studies are aggregated, Q is the sum of the above-mentioned
weighted squared difference and is assumed to follow a chi squared distribution with N 1 degrees of freedom. The weighting factor of these regressions for the assessment of study quality was the number of patients of each study. In order to test if the ORs of different ERP components differed significantly from each other, the log OR was assumed to be normally distributed (the natural logarithm transformation is frequently used to normalize data, e.g., Thompson and Sharp, 1999). The log OR were ordered by ascending value. The variance of these log OR was derived from the CIs as follows: var (log OR) = {log [max(CI)/min(CI)] /(2*1.96)}2 (i.e., the relation between the variance and the 95% CI under the assumption of normality). Then, the differences (delta) between each two to-be-compared log OR were computed with their respective variances, the sum of the variances of each term of these differences. Z scores were computed as follows: Z = delta/ square root[var (delta)]. And the significance of these Z was assessed using one-tailed Bonferroni corrected p-values. We assessed potential publication bias using funnel plots (i.e., a symmetrical plot of log OR estimates against the inverse of their variance indicates that there is no publication bias, Egger et al., 1997). In order to further assess the homogeneity of the included studies and the robustness of the OR estimation for the N100, MMN, and P300, a metaanalysis was conducted for each ERP component after removal of each individual study. The regressions were conducted with SPSS version 11.0.1 for Windows. 3. Results Our systematic search (see Section 2.1) found 154 publications. The application of the above-mentioned eligibility
Table 4 Estimated ORs Prognostic variable
Reference
EOR
Min (95% CI)
Max (95% CI)
p Value
N100 Stroke and hemorrhage Trauma Anoxic Post-operative Encephalopathy
Absent Stroke and hemorrhage
2.85 – 2.86 0.21 3.04 0.39
1.91 – 1.69 0.13 1.42 0.18
4.27 – 4.84 0.35 6.53 0.85
<0.001 <0.001 <0.001 <0.001 0.004 0.018
MMN Stroke and hemorrhage Trauma Anoxic Post-operative Encephalopathy
Absent Stroke and hemorrhage
6.53 – 2.81 0.23 3.13 0.71
3.55 – 1.64 0.12 1.43 0.23
12.01 – 4.81 0.42 6.86 2.22
<0.001 <0.001 <0.001 <0.001 0.004 0.557
P300 Anoxic Trauma Stroke and hemorrhage Encephalopathy
Absent Anoxic
8.79 – 5.16 0.65 0.70
4.88 – 2.70 0.28 0.23
15.83 – 9.84 1.51 2.09
<0.001 <0.001 <0.001 0.319 0.517
Goodness-of-fit test (Hosmer–Lemeshow): p = 0.217 for the estimation of the odd ratio of the N100, p = 0.842 for the estimation of the odd ratio of the MMN, and p = 0.875 for the estimation of the odd ratio of P300. (–), values corresponding to the reference modality not estimated by the model. EOR, estimated odd ratio; MMN, mismatch negativity.
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
Daltrozzo et al
611
Daltrozzo et al
Daltrozzo et al
Fischer et al, 2004
Kane et al, 1996
Fischer et al, 2004
Grapperon et al, 2004
Mutschler et al, 2001 Guérit et al, 1999
Naccache et al, 2005
Signorino et al, 1997
Mutschler et al, 2001
Summary
Summary
Summary
-10
10
-10
0
-10
10
N100 log OR (95 % CI)
MMN log OR (95 % CI)
Summary log OR 1.05 (95% CI, 0.65-1.45)
Summary log OR 1.88 (95% CI, 1.27-2.49)
0
10
P300 log OR (95 % CI) Summary log OR 2.17 (95% CI, 1.59-2.76)
Fig. 1. Forest plots of the natural logarithm of the estimated odd ratios (ORs) for the N100, MMN, and P300. This figure shows the natural logarithm of the estimated ORs and 95% confidence intervals (CI) (limit lines) for each individual study included in the meta-analysis and each component of evenrelated potential. The summary ORs and 95% CIs are given at the bottom of each graph. The log OR were biased due to null frequencies for some studies (see false positives, Table 1), which are not plotted. However, these biases did not contribute to the studies heterogeneity according to the Qdf statistic (see Section 2).
criteria (see Section 2.2) retained 10 studies (Table 1). Considering these data collapsed across sources, the MMN showed a larger specificity (Spe, 91%, CI: 85–95) and posi-
tive predictive value (Ppv, 88%, CI: 81–93) than the P300 (Spe, 77%, CI: 70–82, Ppv, 68%, CI: 59–75) and the N100 (Spe, 57%, CI: 51–63, Ppv, 68%, CI: 62–72) (Table 2). The
Table 5 Estimated ORs Etiology
Prognostic variable
Reference
Min (95% CI)
Max (95% CI)
p value
N
All
N100 MMN P300
Absent Absent Absent
2.85 6.53 8.79
1.91 3.55 4.88
4.27 12.01 15.83
<0.001 <0.001 <0.001
548 470 313
Stroke and Hemorrhage
N100 MMN P300
Absent Absent Absent
2.05 4.47 10.29
1.12 1.92 2.00
3.75 10.44 52.79
0.020 0.001 0.005
197 156 52
Trauma
N100 MMN P300
Absent Absent Absent
1.63 4.72 12.89
0.70 1.35 4.82
3.80 16.44 34.43
0.263 0.015 <0.001
137 155 111
Anoxic
N100 MMN P300
Absent Absent Absent
8.03 15.50 5.93
2.83 4.27 2.38
22.75 56.26 14.77
<0.001 <0.001 <0.001
129 90 127
Post-operative
N100 MMN
Absent Absent
10.66 –
1.98 –
57.50 –
0.006 –
54 54
N100 MMN P300
Absent Absent Absent
2.12 3.60 7.71
0.34 0.28 0.75
13.13 46.36 79.77
0.421 0.326 0.087
31 15 23
Encephalopathy
(–), OR not estimated by the model. EOR, estimated odd ratio; MMN: mismatch negativity. N, number of patients included in the regression.
EOR
612
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
few false positives for the MMN component imply that the specificity and positive predictive value for awakening (i.e., the sensitivity and negative predictive value for nonawakening) are below 100%. For the logistic regression analysis, the studies were tested for homogeneity. The MMN studies were not significantly heterogeneous (Qdf=4 = 7.56, p = 0.11) but significant heterogeneity was found for N100 (Qdf=6 = 14.6, p = 0.02) and P300 (Qdf=6 = 16.6, p = 0.01) studies. Thus, the N100 data of Fischer et al. (2006), which resulted in a large log OR compared to the other studies, and the P300 data of Grapperon et al. (2004), with a small log OR, were rejected. After these rejections, the tests of homogeneity were not anymore significant for the N100 (Qdf=5 = 8.49, p = 0.13) and the P300 (Qdf=5 = 8.35, p = 0.14) studies. The remaining data, which entered in the meta-analysis, are summarized in Table 3. The estimated ORs are presented in Table 4 and Fig. 1. The OR for the N100 (ORN100) was estimated with the data of 548 patients, the OR for the MMN (ORMMN) with 470 patients, and the OR for the P300 (ORP300) with 313 patients. ORN100 was significantly smaller than ORP300 (p = 0.001) and ORMMN (p = 0.013). ORP300 was larger than ORMMN but this difference was not significant (p = 0.754). The prognosis was the worst for the patient with anoxia or metabolic encephalopathy and the best for trauma or brain surgery (Table 4). According to the meta-analysis performed on the P300 sources, the rate of awakening of patients with stroke or hemorrhage was not significantly different from the OR of patients with anoxia. The OR of anoxic patients was significantly smaller than the ORs for trauma and brain surgery (p < 0.001, with the N100 and MMN sources). The patient suffering from metabolic encephalopathy had an OR significantly smaller than those with trauma (p < 0.001 with the N100 sources and p = 0.001 with the P300 sources) and brain surgery (p < 0.001, with the N100 sources). The ORs for stroke and hemorrhage were significantly smaller than the OR for trauma (p < 0.001, with the P300 sources). The general tendency (i.e., ORP300 > ORMMN > ORN100) observed across all etiologies is also found when ORs are estimated with patient samples grouped by stroke and hemorrhage, trauma, and metabolic encephalopathy etiology (Table 5) but not with anoxic patients (i.e., ORMMN > ORN100 > ORP300). In stroke and hemorrhage patients, the P300 predicts awakening significantly better than the N100 (p < 0.001). The apparent superiority of the P300 compared to the MMN in patients with stroke and hemorrhage, trauma, and metabolic encephalopathy is not significant. In anoxic patients, the MMN seems to be a better predictor than the P300, but this difference does not reach significance. The funnel plots (Fig. 2) did not indicate an asymmetry and thus the possibility of publication bias. Fig. 3 indicates that the OR estimation remains stable against the removal of an individual study.
1 / variance
15
0 0
N100 log OR
4
1 / variance
10
0 0
MMN log OR
3
P300 log OR
5
1 / variance
10
0 0
Fig. 2. Funnel plots. This figure shows the logarithm of the estimated ORs for each individual study included in the meta-analysis and each component of even-related potential on the horizontal axis, and an estimate of their precision (inverse of the variance) is plotted on the vertical axis. The vertical axis cuts the horizontal axis at the summary log ORs. In a few studies, the false positive frequency was null (Table 1). Thus, the estimated log OR were biased and are not plotted. Noteworthy, these biases did not contribute to the studies heterogeneity according to the Qdf statistic (see Section 2).
4. Discussion In patients suffering from stroke, hemorrhage, trauma, or metabolic encephalopathy, the odd ratios (ORs) of the N100 (ORN100), MMN (ORMMN), and P300 (ORP300) were found as follows: ORP300 larger than ORMMN larger than ORN100, with a nonsignificant difference between ORP300 and ORMMN.
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
613
none
none
none
Signorino et al, 1997
Mutschler et al, 2001
Naccache et al, 2005
Lew et al, 2003
Kane et al, 1996
Mutschler et al, 2001 Lew et al, 2003 Guérit el al, 1999
Kane et al, 1996
Fischer et al, 2004
Guerit et al, 1999
Fischer et al, 2006
Fischer et al, 2004
Daltrozzo et al
Daltrozzo et al
Daltrozzo et al
0
5
P300 log OR (95 % CI)
Grapperon et al, 2004
0
5
MMN log OR (95 % CI)
0
2
N100 log OR (95 % CI)
Fig. 3. Assessment of Robustness of the odds ratio (OR) estimation for the N100, MMN, and P300. This figure shows the logarithm of the estimated ORs and 95% confidence intervals (CI) (limit lines) for each component of even-related potential and each meta-analysis performed after removal of each individual study. The largest deviation from the meta-analysis of all studies is due to the removal of a high percentage of patients (i.e., 59.8% if Fischer et al., 2004 is removed from the MMN studies and 46.6% if Mutschler et al., 2001 is removed from the P300 studies) or to a low goodness of fit (Hosmer– Lemeshow p < .05 if Fischer et al., 2006 is removed from the MMN studies). Note that the goodness of fit is correct (Hosmer–Lemeshow p > .05) for all other OR estimations.
This trend (ORP300 > ORMMN > ORN100) with an equivalence between ORP300 and ORMMN was confirmed when the ORs of ERPs were estimated on larger patient samples with a regression analysis and the etiology as explanatory variable. Thus, our data do not confirm that the MMN is a better predictor of awakening than the P300 (Kane et al., 2000). The meta-analysis also indicated that the presence of each of the N100, MMN, and P300 components is a highly significant predictor of awakening, the P300 and MMN being significantly better predictors than the N100. However, since the ERP components are not mandatory evoked potentials, even in healthy participants (Gue´rit, 2004), their absence does not predict nonawakening (see Section 2). It could be argued that the comparison of the ORs between six N100 studies, five MMN studies, and six P300 studies is not meaningful because they were estimated on different patient samples. However, these samples largely overlapped. The fact that the estimated ORs for the etiology variable were almost identical between samples (Table 4) indicates that the differences between them were moderate. Further, each OR was estimated on a set of several studies; thus can be considered independent of methodologies. The homogeneity of the included studies was not only warranted by the Q test (Fleiss and Gross, 1991) but also by an assessment of the robustness of the OR estimation (Fig. 3) indicating stability against removal of an individual study.
A few low responsive patients could have a GOS larger than 2 at the time of the ERP recording. When these patients are more likely to have ERPs present, the prognostic value of ERPs becomes probably less for them. The GOS was assessed between 1 and 12 months after the coma onset for most of the selected sources (Table 1). Although this difference may have slightly affected the results, it should be taken into account that the acute coma state rarely lasts longer than two or three weeks, after which the outcome can be defined. The difference between 1-month and 1-year assessments could rather be critical for the outcome of the subsequent vegetative state (i.e., GOS = 2), which was not the issue of the present study. The fact that the N100 tends to be a poorer predictor of awakening than the two later, more ‘‘cognitive’’ ERP components (except perhaps in anoxic patients where ORMMN > ORN100 > ORP300, Table 5) should probably not be discussed in the present article, before this tendency is confirmed by another, larger study. However, the issue has an important methodological aspect, since some investigators look for the late components only in the presence of N100. In a study of the persistent vegetative state, we showed that some patients can demonstrate very clear, and statistically highly significant, late components (including P300) without any visible sign of N100 (Kotchoubey et al., 2005; Fig. 4). Therefore, defining the presence of one component (which may be a better predictor) as dependent on the presence of another component (which
614
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
is possibly a worse predictor) would decrease the predictive value of the ERP method as a whole. Our data are in line with most literature indicating that traumatic and post-operative etiologies are related to the best chance of awakening whereas the lowest rate of awakening is obtained for anoxia and metabolic encephalopathy. Thus, Fischer et al. (2004) reported a significantly smaller OR for anoxia compared to traumatic and post-operative etiologies. This consistency could be expected since a big portion of the data used for the estimation of the ORMMN and ORN100 (60% and 51%, respectively) stems from Fischer et al. (2004). Obviously, traumatic and postoperative etiologies often relate to focal brain injuries while anoxia and metabolic encephalopathy are related to very diffuse injuries. Thus, our results confirm that the spatial extension of the brain injury in coma is markedly related to the probability of awakening. For certain etiologies like metabolic encephalopathy or post-operative, the number of patients is small (Table 3). Thus, the reliability of the estimated OR for these etiologies is questionable. 5. Conclusion This study is the first to compare the prognostic power of three ERP components: the N100, the MMN, and the P300, using a statistical measure (the odd ratio), on the basis of a large number of patients (313–548). The analysis reveals that the presence of each of these ERP components significantly predicts awakening. The P300 and MMN had significantly larger predictive powers than the N100. This trend was confirmed in patient samples of stroke and hemorrhage, trauma, and metabolic encephalopathy etiology but not with anoxic. Our data confirm that the etiology is a strong indicator of outcome, the likelihood of subsequent awakening being higher in trauma and post-operative etiologies than in anoxia and metabolic encephalopathy. Acknowledgements This study was supported by the Deutsche Forschungsgemeinshaft (SFB 550 to B.K.) and the French Ministry of Health (PHRC 2004 R-03-03). References De Giorgio CM, Rabinowicz AL, Gott PS. Predictive value of P300 eventrelated potentials compared with EEG and somatosensory evoked potentials in non-traumatic coma. Acta Neurol Scand 1993;87:423–7. Egger M, Davey Smith G, Schneider M, Minder C. Biais in meta-analysis detected by a simple, graphical test. BMJ 1997;315:629–34. Fischer C, Luaute´ J, Ne´moz C, Morlet D, Kirkorian G, Mauguie`re F. Improved prediction of awakening or nonawakening from severe
anoxic coma using tree-based classification analysis. Crit Care Med 2006;34(5):1520–4. Fischer C, Luaute´ J, Adeleine P, Morlet D. Predictive value of sensory and cognitive evoked potentials for awakening from coma. Neurology 2004;63:669–73. Fleiss JL, Gross AJ. Meta-analysis in epidemiology, with special reference to studies of the association between exposure to environmental tobacco smoke and lung cancer: A critique. J Clin Epidemiol 1991;44(2):127–39. Gott PS, Rabinowicz AL, De Giorgio CM. P300 auditory event-related potentials in nontraumatic coma. Association with Glasgow Coma Score and awakening. Arch Neurol 1991;48:1267–70. Grapperon J, Vidal F, Bruschera D, Cantais E, Salinier L, Costes O, et al. Auditive and somato-sensory event-related potentials in comas: Prognostic value for awakening, social and professional reinstatement. Ann Franc¸aises Anesth Re´anim 2004;23:102–8. Gue´rit JM. Prognostic contribution for potentials evoked in unit of intensive care. Ann Franc¸aises Anesth Re´anim 2004;23:99–101. Gue´rit JM, Verougstraete D, De Tourtchaninoff M, Debatisse D, Witdoeckt C. ERPs obtained with the auditory oddball paradigm in coma and alterated states of consciousness: Clinical relationships, pronostic value, and origin of components. Clin Neurophysiol 1999;110:1260–9. Jennett B, Bond M. Assessment of outcome after severe brain damage. Lancet 1975;1(7905):480–4. Kane NM, Butler SR, Simpson T. Coma outcome prediction using eventrelated potentials: P3 and mismatch negativity. Audiol Neurol 2000;5:188–91. Kane NM, Curry SH, Rowlands CA, Manara AR, Lewis T, Moss T, et al. Event related potentials – neurophysiological tools for predicting emergence and early outcome from traumatic coma. Intensive Care Med 1996;22:39–46. Kane NM, Curry SH, Butler SR, Cummins BH. Electrophysiological indicator of awakening from coma. Lancet 1993;341:688. Kotchoubey B, Lang S, Mezger G, Schmalohr D, Schneck M, Semmler A, et al. Information processing in severe disorders of consciousness: Vegetative state and minimally conscious state. Clin Neurophysiol 2005;116:2441–53. Lew HL, Dikmen S, Slimp J, Temkin N, Lee EH, Newell D, et al. Use of somatosensory-evoked potentials and cognitive event-related potentials in predicting outcomes of patients with severe traumatic brain injury. Am J Phys Med Rehabil 2003;82:53–61. Mutschler V, Wioland N, Delabranche X, Le Gourrier L, Calon B. Valeur pronostique des potentiels evoques exogenes et endogenes dans le coma chez l’adulte. In: Gue´rit JM, editor. L‘e´valuation neurophysiologique des comas, de la mort ence´phalique et des e´tats ve´ge´tatifs chroniques. Marseille, France: Solal; 2001. p. 183–94. Naccache L, Puybasset L, Gaillard R, Serve E, Willera JC. Auditory mismatch negativity is a good predictor of awakening in comatose patients: a fast and reliable procedure. Clin Neurophysiol 2005;116:988–9. Reuter PS, Linke DB. P300 and coma. In: Maurer K, editor. Topographic Brain Mapping of EEG and Evoked Potentials. Berlin, Heidelberg, New York, Tokyo: Springer; 1989. p. 192–6. Signorino M, D’acunto S, Cercaci S, Pietropaoli P, Angeleri F. The P300 in traumatic coma: Conditioning of the odd-ball paradigm. J Psychophysiol 1997;11:59–70. Sinson G, D’acunto S, Cercaci S, Pietropaoli P, Angeleri F. Magnetization transfer imaging and proton MR spectroscopy in the evaluation of axonal injury: Correlation with clinical outcome after traumatic brain injury. Am J Neuroradiol 2001;22:143–51. Thompson SG, Sharp SJ. Explaining heterogeneity in meta-analysis: A comparison of methods. Statist Med 1999;18:2693–708.
Predicting coma and other low responsive patients outcome using event-related brain potentials: A meta-analysis J. Daltrozzo
a,b,c,*
, N. Wioland
a,b
, V. Mutschler
a,b
, B. Kotchoubey
c
a
c
Department of Neurology, University Hospital of Strasbourg, France b CNRS UPS 858, Louis Pasteur University of Strasbourg, France Department of Medical Psychology and Behavioral Neurobiology, University of Tu¨bingen, Germany See Editorial, pages 477–479
Abstract Objective: A meta-analysis was performed to estimate the predictive power (odd ratio, OR) for awakening of auditory event-related potential (ERP) components in low responsive patients with stroke or hemorrhage, trauma, anoxic, post-operative, and metabolic encephalopathy etiologies. Methods: We reviewed MEDLINE and analyzed citations for retrieved articles. Logistic regressions were applied on patient samples (Glasgow Coma Scale <12) across and for separate etiologies. Results: For stroke and hemorrhage the ORs with 95% confidence intervals were: 2.05 [1.12–3.75] (N100), 4.47 [1.92–10.44] (MMN), 10.29 [2.00–52.79] (P300), for trauma: 1.63 [0.70–3.80] (N100), 4.72 [1.35–16.44] (MMN), 12.89 [4.82–34.43] (P300), anoxic: 8.03 [2.83–22.75] (N100), 15.50 [4.27–56.26] (MMN), 5.93 [2.38–14.77] (P300), post-operative: 10.66 [1.98–57.50] (N100), metabolic encephalopathy: 2.12 [0.34–13.13] (N100), 3.60 [0.28–46.36] (MMN), 7.71 [0.75–79.77] (P300), and all etiologies: 2.85 [1.91–4.27] (N100), 6.53 [3.55–12.01] (MMN), and 8.79 [4.88–15.83] (P300). Based on six N100 studies (N = 548 patients), five MMN studies (N = 470), and six P300 studies (N = 313), the N100, MMN, or P300, when present, significantly predicted awakening, P300 and MMN being significantly better predictors than N100. Conclusions: The MMN and P300 appear to be reliable predictors of awakening. Significance: The prognostic assessment of low responsive patients with auditory ERP should take into account both MMN and P300. Ó 2006 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved. Keywords: Coma; Prognosis; ERP; N100; MMN; P300
1. Introduction Event-related potentials (ERPs) can be used to predict awakening from coma. The prognostic power of ERPs has been investigated in several prospective cohort studies. Reuter and Linke (1989) were the first to report the presence of an auditory ERP component, the P300 in four coma patients and they indicated that all these patients survived. Their findings were subsequently confirmed by Gott et al. (1991) and other authors. Most often, the visual detection of the P300 was performed only if an earlier ERP *
Corresponding author. Tel.: +33 388 127 414; fax: +33 388 116 568. E-mail address: [email protected] (J. Daltrozzo).
component, which is called the N100 and intended to indicate mainly a sensory cortical processing, was present (De Giorgio et al., 1993). Another ERP component, the mismatch negativity (MMN), can also be recorded in comatose patients and is also an indicator of prognosis (Kane et al., 1993). Both MMN and P300 reflect discrimination of sounds. The MMN is modality-specific (i.e., generated only by auditory stimuli), due to an automatic mechanism and depends on the physical distance between the two discriminated sounds (e.g., a difference of frequency, sound duration, or intensity). Conversely, the P300 is independent of sensory modality, can easily be modulated by arousal and attention, and does not depend on the physical difference between stimuli. In an attempt to better evaluate the prognostic
1388-2457/$32.00 Ó 2006 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.clinph.2006.11.019
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
power of these ERP components, later studies used larger samples of patients (e.g., Fischer et al., 2004; Gue´rit et al., 1999; Mutschler et al., 2001). They have confirmed that these ERP components, and particularly the MMN, are reliable predictors of awakening from coma. Fischer et al. (2004) were the first to demonstrate the predictive value of the MMN statistically with the odd ratio (OR) measure. The goal of our study is to perform meta-analyses in order to compare the prognostic power of several ERP components on a large pool of patients and with a statistical measure already used for such purpose, the OR. 2. Methods 2.1. Systematic search Two of us independently reviewed MEDLINE from January 1, 1989 to June 13, 2006, with key words: coma and (ERP or N1 or MMN or mismatch negativity or P300 or prognostic and auditory event-related potentials) and analyzed citations for retrieved articles. 2.2. Eligibility criteria The sources were independently selected by two of us on the basis of the following criteria: (1) the last report of each research team to avoid duplicate data or reports of the same team if populations were different according to the etiology of coma; (2) reports in English or in French; (3) at least one of ERP components N1, MMN and P300 was recorded; (4) patient’s Glasgow Coma Scale (GCS) score lower than 12 at the time of the ERP recording; (5) outcome of coma defined on the basis of the Glasgow Outcome Score (GOS, 1: death, 2: persistent vegetative state, 3: severe disability, 4: moderate disability, and 5: good recovery, Jennett and Bond, 1975; Sinson et al., 2001) and compared with the presence of the ERP component for each group of patients classified by etiology. We contacted authors to obtain further data if required for inclusion. 2.3. Extracted data From the selected studies the following data were extracted for each etiology: the number of True Positives (TP, number of patients with the ERP component who awakened, i.e., with GOS larger than 2), of False Negatives (FN, number of patients without the ERP component and with GOS >2), True Negatives (TN, number of patients without the ERP component and with GOS 62), and False Positives (FP, number of patients with the ERP component and with GOS 62). 2.4. Statistical analysis The sensitivity (TP/(TP + FN)), specificity (TN/(TN + FP)), positive predictive value (TP/(TP + FP)), and nega-
607
tive predictive value (TN/(TN + FN)) for awakening were estimated for each selected study. Exact 95% confidence intervals (CIs) were estimated using the binomial distribution (Fischer et al., 2006) with Microsoft Excel 97 SR-2. The prognostic power of ERPs was estimated with a logistic regression analysis. The logistic regression analysis is a statistical model used to study the relationships between explanatory qualitative (or quantitative) variables and a dependent qualitative variable. These relationships are estimated by the model in terms of OR. In the present study, all variables except the etiology are qualitative with two modalities, i.e., binary variables. The dependent variable is the outcome of the patient. It is equal to 1 if the patient subsequently awakens, i.e., if the GOS is larger than 2, and equal to 0 otherwise. The explanatory variables are the presence (variable equal to 1) or absence (variable equal to 0) of the N100, the MMN, or the P300, and the patient etiology. The etiology was considered as a nominal variable with five modalities for the N100 and the MMN and four modalities for the P300. The etiology with the largest number of patients was chosen as the reference modality for this explanatory variable (e.g., Fischer et al., 2004) in order to have the most stable regression model and thus the best OR estimates for the etiology variable. In addition to the logistic regression on the data of all patients (i.e., across all etiologies), the predictive power for awakening of each ERP component was estimated with patient samples grouped by etiology. For the relationship between the ERP variable and the outcome, an OR significantly larger than 1, i.e., the lower limit of the estimated 95% Confidence Interval (CI) of the OR is larger than 1, indicates that the presence of the ERP component is a significant predictor of awakening. Since ERP components are sensitive to vigilance and levels of sedation (Gue´rit, 2004), an OR significantly larger than 1 does not allow to conclude that the absence of the ERP component predicts significantly nonawakening (i.e., death or vegetative state). An OR smaller or not significantly different from 1 (i.e., the CI contains the value 1) indicates that the ERP is not predictive of outcome. For the relationship between an etiology (i.e. one modality of the etiology variable different from the reference modality) and the outcome, an OR significantly larger than 1 indicates that the etiology is significantly more related to awakening than the reference etiology. An OR significantly smaller than 1 indicates that the etiology is significantly less related to awakening than the reference etiology. An OR not significantly different from 1 (i.e., the CI contains the value 1) indicates that the relationship between the etiology and the outcome does not differ significantly from the relationship between the reference etiology and the outcome. The analysis of heterogeneous studies (with a randomeffects model) remains a controversial debate (Fleiss and Gross, 1991, p.129). Thus, we conducted logistic regressions with a fixed-effects model. This model assumes that the aggregated studies are homogeneous (i.e., the effect size, here the log OR, does not vary very much between studies).
608
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
Table 1 Selected sources Source *
Daltrozzo et al
ERP
Etiology
MMN
Anoxic Encephalopathy Stroke and hemorrhage Anoxic Encephalopathy Stroke and hemorrhage Anoxic Encephalopathy Stroke and hemorrhage Encephalopathy Post-operative Stroke and hemorrhage Trauma Encephalopathy Post-operative Stroke and hemorrhage Trauma
N100
P300
Fischer et al. (2004)
MMN
N100
N
TP
FN
TN
FP
28 4 12 29 4 11 28 4 12
0 0 1 3 0 0 1 0 0
4 0 0 1 0 2 3 0 1
20 3 11 22 3 6 23 4 11
4 1 0 3 1 3 1 0 0
7 54 124 96 7 54 125 96
1 14 33 20 4 32 63 50
4 30 53 63 1 12 23 33
2 10 31 10 0 8 20 7
0 0 7 3 2 2 19 6
Fischer et al. (2006)
MMN N100
Anoxic Anoxic
62 62
12 15
8 5
42 36
0 6
Grapperon et al. (2004)
N100
Anoxic Encephalopathy Stroke and hemorrhage Trauma Anoxic Encephalopathy Stroke and hemorrhage Trauma
2 1 21 9 2 1 21 9
1 0 12 3 0 0 4 0
0 1 8 3 1 1 16 6
0 0 1 0 0 0 1 2
1 0 0 3 1 0 0 1
P300
Gue´rit et al. (1999)
N100 P300
Anoxic Anoxic
21 22
10 5
1 6
7 11
3 0
Kane et al. (1996)
MMN P300
Trauma trauma
53 54
22 33
17 4
14 15
0 2
Lew et al. (2003)
N100 P300
Trauma Trauma
22 22
12 7
4 9
6 6
0 0
Mutschler et al. (2001)
N100
Anoxic Encephalopathy Stroke and hemorrhage Trauma Anoxic Encephalopathy Stroke and hemorrhage Trauma
77 19 40 10 77 19 40 10
11 7 13 5 9 6 12 5
3 0 0 0 5 1 1 0
32 2 6 0 49 5 13 2
31 10 21 5 14 7 14 3
P300
Naccache et al. (2005)
MMN MMN MMN
Encephalopathy Stroke and hemorrhage Trauma
4 20 6
2 6 1
1 4 2
1 9 3
0 1 0
Signorino et al. (1997)
P300
Trauma
25
13
5
5
2
ERP, event-related potential; N, number of patients; TP, true positives (i.e., number of patients with the ERP component and with GOS >2); FN, false negatives (i.e., number of patients without the ERP component and with GOS >2); TN, true negatives (i.e., number of patients without the ERP component and with GOS 62); FP, false positives (i.e., number of patients with the ERP component and with GOS 62). MMN, mismatch negativity. Note 1. The GOS was estimated: 1 month after coma onset in Naccache et al. (2005); 1.5 months in Mutschler et al. (2001); 3 months in Daltrozzo et al* and in Kane et al. (1996); 6 months in Lew et al. (2003) and in Signorino et al. (1997); 12 months in Fischer et al. (2006) and Fischer et al. (2004); and just before the patient leaves the ICU in Grapperon et al. (2004). The time of GOS assessment was not reported in Gue´rit et al. (1999). Note 2. The mean age of the patients was 49 years [1–93] (age range was not reported in Signorino et al. (1997)). In most sources the age range was very large (e.g., [8–93] in Fischer et al. (2004); [1–80] in Kane et al. (1996); [18–89] in Gue´rit et al. (1999)). Since, the patients’ age was not reported separately for TP, FN, TN, and FP by etiology, it was not possible to include this variable in the regression. * Je´roˆme Daltrozzo, Norma Wioland, Ve´ronique Mutschler, Philippe Lutun, Albert Jaeger, Bartholomeus Calon, Alain Meyer, Thierry Pottecher, Simone Lang, and Boris Kotchoubey, Cortical information processing in coma and other low responsive states, in preparation.
Table 2 Sensitivity (Se), specificity (Spe), positive predictive value (Ppv), and negative predictive value (Npv) for awakening of selected sources Sources
MMN Daltrozzo et al Fischer et al. (2004) Fischer et al. (2006) Kane et al. (1996) Naccache et al. (2005) All P300 Daltrozzo et al Grapperon et al. (2004) Gue´rit et al. (1999) Kane et al. (1996) Lew et al. (2003) Mutschler et al. (2001) Signorino et al. (1997) All
N
TP
FN
TN
FP
Se
%
95% CI
Spe
44 282 62 33
3 149 15 16
3 69 5 12
31 35 36 1
7 29 6 4
3/6 149/218 15/20 16/28
50 68 75 57
12–88 62–75 51–91 37–76
31/38 35/64 36/42 1/5
21 22 146 610
10 12 36 241
1 4 3 97
7 6 40 156
3 0 67 116
10/11 12/16 36/39 241/338
91 75 92 71
59–100 48–93 79–98 66–76
44 281 62 53 30 470
1 68 12 22 9 112
4 150 8 17 7 186
34 53 42 14 13 156
5 10 0 0 1 16
1/5 68/218 12/16 22/39 9/16 112/298
20 31 60 56 56 38
1–72 25–38 48–93 40–72 30–80 32–43
44 33
1 4
4 24
38 3
1 2
1/5 4/28
20 14
1–72 4–33
22 54 22 146 25 346
95% CI
Ppv
82 55 86 20
66–92 42–67 71–95 1–72
3/10 149/178 15/21 16/20
7/10 6/6 40/107 156/272
70 100 37 57
35–93 54–100 28–47 51–63
34/39 53/63 42/42 14/14 13/14 156/172
87 84 100 100 93 91
5 33 7 32 13 95
6 4 9 7 5 59
11 15 6 69 5 147
0 2 0 38 2 45
5/11 33/37 7/16 32/39 13/18 95/154
45 89 44 82 72 62
17–77 75–97 20–70 66–93 47–90 53–69
38/39 3/5 11/11 15/17 6/6 69/107 5/7 147/192
%
%
95% CI
Npv
%
95% CI
30 84 71 80
7–65 77–89 48–89 56–94
31/34 35/104 36/41 1/13
91 34 88 8
76–98 25–44 74–96 0–36
10/13 12/12 36/103 241/357
77 100 35 68
46–95 74–100 26–45 62–72
7/8 6/10 40/43 156/253
88 60 93 62
48–100 26–88 81–99 55–68
73–96 73–92 92–100 77–100 66–100 85–95
1/6 68/78 12/12 22/22 9/10 112/128
17 87 100 100 90 88
0–64 78–94 74–100 85–100 56–100 81–93
34/38 53/203 42/50 14/31 13/20 156/342
89 26 84 45 65 46
75–97 20–33 71–93 27–64 41–85 40–51
97 60
87–100 15–95
1/2 4/6
50 67
1–99 22–96
38/42 3/27
90 11
77–97 2–29
100 88 100 64 71 77
72–100 64–99 54–100 55–74 29–96 70–82
5/5 33/35 7/7 32/70 13/15 95/140
100 94 100 46 87 68
48–100 81–99 59–100 34–58 60–98 59–75
11/17 15/19 6/15 69/76 5/10 147/206
65 79 40 91 50 71
38–86 54–94 16–68 82–96 19–81 65–77
ERP, event-related potential; N, number of patients; TP, true positives (i.e., number of patients with the ERP component and with GOS >2); FN, false negatives (i.e., number of patients without the ERP component and with GOS >2); TN, true negatives (i.e., number of patients without the ERP component and with GOS 62); FP, false positives (i.e., number of patients with the ERP component and with GOS 62). Se, Sensitivity for awakening = TP/(TP + FN); Spe, specificity for awakening = TN/(TN + FP); Ppv, positive predictive value for awakening = TP/(TP + FP); Npv, negative predictive value for awakening = TN/(TN + FN). For nonawakening specificity exchange with sensitivity and Ppv with Npv. MMN, mismatch negativity. Note. The GOS was estimated: 1 month after coma onset in Naccache et al. (2005); 1.5 months in Mutschler et al. (2001); 3 months in Daltrozzo et ala and in Kane et al. (1996); 6 months in Lew et al. (2003) and in Signorino et al. (1997); 12 months in Fischer et al. (2006) and Fischer et al. (2004); and just before the patient leaves the ICU in Grapperon et al. (2004). The time of GOS assessment was not reported in Gue´rit et al. (1999). a Je´roˆme Daltrozzo, Norma Wioland, Ve´ronique Mutschler, Philippe Lutun, Albert Jaeger, Bartholomeus Calon, Alain Meyer, Thierry Pottecher, Simone Lang, and Boris Kotchoubey, Cortical information processing in coma and other low responsive states, in preparation.
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
N100 Daltrozzo et ala Fischer et al. (2004) Fischer et al. (2006) Grapperon et al. (2004) Gue´rit et al. (1999) Lew et al. (2003) Mutschler et al. (2001) All
609
610
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
Table 3 Sources included in the meta-analysis ERP
Etiology
N
TP
FN
TN
FP
N100
Stroke and hemorrhage Trauma Anoxic Post-operative Encephalopathy
197 137 129 54 31
88 70 25 32 11
33 40 5 12 2
33 13 61 8 5
43 14 38 2 13
MMN
Stroke and hemorrhage Trauma Anoxic Post-operative Encephalopathy
156 155 90 54 15
40 43 12 14 3
57 82 12 30 5
51 27 62 10 6
8 3 4 0 1
P300
Anoxic Trauma Stroke and hemorrhage Encephalopathy
127 111 52 23
15 58 12 6
14 18 2 1
83 28 24 9
15 7 14 7
ERP, Event-Related Potential; N, Number of patients; TP, True Positives (i.e., number of patients with the ERP component and with GOS >2); FN, False Negatives (i.e., number of patients without the ERP component and with GOS >2); TN, True Negatives (i.e., number of patients without the ERP component and with GOS 62); FP, False Positives (i.e., number of patients with the ERP component and with GOS 62). MMN, mismatch negativity. Note: The GOS was assessed between 1 and 12 months after the incident for most of the selected sources (Table 1).
We retained only homogeneous studies using the following method. Studies with the highest weighted squared difference between the log OR of the individual study and the log OR of the aggregated studies were considered as outliers and thus removed if the Q test for homogeneity of studies was significant (Fleiss and Gross, 1991). Q is a measure of the heterogeneity between a set of studies. If N studies are aggregated, Q is the sum of the above-mentioned
weighted squared difference and is assumed to follow a chi squared distribution with N 1 degrees of freedom. The weighting factor of these regressions for the assessment of study quality was the number of patients of each study. In order to test if the ORs of different ERP components differed significantly from each other, the log OR was assumed to be normally distributed (the natural logarithm transformation is frequently used to normalize data, e.g., Thompson and Sharp, 1999). The log OR were ordered by ascending value. The variance of these log OR was derived from the CIs as follows: var (log OR) = {log [max(CI)/min(CI)] /(2*1.96)}2 (i.e., the relation between the variance and the 95% CI under the assumption of normality). Then, the differences (delta) between each two to-be-compared log OR were computed with their respective variances, the sum of the variances of each term of these differences. Z scores were computed as follows: Z = delta/ square root[var (delta)]. And the significance of these Z was assessed using one-tailed Bonferroni corrected p-values. We assessed potential publication bias using funnel plots (i.e., a symmetrical plot of log OR estimates against the inverse of their variance indicates that there is no publication bias, Egger et al., 1997). In order to further assess the homogeneity of the included studies and the robustness of the OR estimation for the N100, MMN, and P300, a metaanalysis was conducted for each ERP component after removal of each individual study. The regressions were conducted with SPSS version 11.0.1 for Windows. 3. Results Our systematic search (see Section 2.1) found 154 publications. The application of the above-mentioned eligibility
Table 4 Estimated ORs Prognostic variable
Reference
EOR
Min (95% CI)
Max (95% CI)
p Value
N100 Stroke and hemorrhage Trauma Anoxic Post-operative Encephalopathy
Absent Stroke and hemorrhage
2.85 – 2.86 0.21 3.04 0.39
1.91 – 1.69 0.13 1.42 0.18
4.27 – 4.84 0.35 6.53 0.85
<0.001 <0.001 <0.001 <0.001 0.004 0.018
MMN Stroke and hemorrhage Trauma Anoxic Post-operative Encephalopathy
Absent Stroke and hemorrhage
6.53 – 2.81 0.23 3.13 0.71
3.55 – 1.64 0.12 1.43 0.23
12.01 – 4.81 0.42 6.86 2.22
<0.001 <0.001 <0.001 <0.001 0.004 0.557
P300 Anoxic Trauma Stroke and hemorrhage Encephalopathy
Absent Anoxic
8.79 – 5.16 0.65 0.70
4.88 – 2.70 0.28 0.23
15.83 – 9.84 1.51 2.09
<0.001 <0.001 <0.001 0.319 0.517
Goodness-of-fit test (Hosmer–Lemeshow): p = 0.217 for the estimation of the odd ratio of the N100, p = 0.842 for the estimation of the odd ratio of the MMN, and p = 0.875 for the estimation of the odd ratio of P300. (–), values corresponding to the reference modality not estimated by the model. EOR, estimated odd ratio; MMN, mismatch negativity.
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
Daltrozzo et al
611
Daltrozzo et al
Daltrozzo et al
Fischer et al, 2004
Kane et al, 1996
Fischer et al, 2004
Grapperon et al, 2004
Mutschler et al, 2001 Guérit et al, 1999
Naccache et al, 2005
Signorino et al, 1997
Mutschler et al, 2001
Summary
Summary
Summary
-10
10
-10
0
-10
10
N100 log OR (95 % CI)
MMN log OR (95 % CI)
Summary log OR 1.05 (95% CI, 0.65-1.45)
Summary log OR 1.88 (95% CI, 1.27-2.49)
0
10
P300 log OR (95 % CI) Summary log OR 2.17 (95% CI, 1.59-2.76)
Fig. 1. Forest plots of the natural logarithm of the estimated odd ratios (ORs) for the N100, MMN, and P300. This figure shows the natural logarithm of the estimated ORs and 95% confidence intervals (CI) (limit lines) for each individual study included in the meta-analysis and each component of evenrelated potential. The summary ORs and 95% CIs are given at the bottom of each graph. The log OR were biased due to null frequencies for some studies (see false positives, Table 1), which are not plotted. However, these biases did not contribute to the studies heterogeneity according to the Qdf statistic (see Section 2).
criteria (see Section 2.2) retained 10 studies (Table 1). Considering these data collapsed across sources, the MMN showed a larger specificity (Spe, 91%, CI: 85–95) and posi-
tive predictive value (Ppv, 88%, CI: 81–93) than the P300 (Spe, 77%, CI: 70–82, Ppv, 68%, CI: 59–75) and the N100 (Spe, 57%, CI: 51–63, Ppv, 68%, CI: 62–72) (Table 2). The
Table 5 Estimated ORs Etiology
Prognostic variable
Reference
Min (95% CI)
Max (95% CI)
p value
N
All
N100 MMN P300
Absent Absent Absent
2.85 6.53 8.79
1.91 3.55 4.88
4.27 12.01 15.83
<0.001 <0.001 <0.001
548 470 313
Stroke and Hemorrhage
N100 MMN P300
Absent Absent Absent
2.05 4.47 10.29
1.12 1.92 2.00
3.75 10.44 52.79
0.020 0.001 0.005
197 156 52
Trauma
N100 MMN P300
Absent Absent Absent
1.63 4.72 12.89
0.70 1.35 4.82
3.80 16.44 34.43
0.263 0.015 <0.001
137 155 111
Anoxic
N100 MMN P300
Absent Absent Absent
8.03 15.50 5.93
2.83 4.27 2.38
22.75 56.26 14.77
<0.001 <0.001 <0.001
129 90 127
Post-operative
N100 MMN
Absent Absent
10.66 –
1.98 –
57.50 –
0.006 –
54 54
N100 MMN P300
Absent Absent Absent
2.12 3.60 7.71
0.34 0.28 0.75
13.13 46.36 79.77
0.421 0.326 0.087
31 15 23
Encephalopathy
(–), OR not estimated by the model. EOR, estimated odd ratio; MMN: mismatch negativity. N, number of patients included in the regression.
EOR
612
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
few false positives for the MMN component imply that the specificity and positive predictive value for awakening (i.e., the sensitivity and negative predictive value for nonawakening) are below 100%. For the logistic regression analysis, the studies were tested for homogeneity. The MMN studies were not significantly heterogeneous (Qdf=4 = 7.56, p = 0.11) but significant heterogeneity was found for N100 (Qdf=6 = 14.6, p = 0.02) and P300 (Qdf=6 = 16.6, p = 0.01) studies. Thus, the N100 data of Fischer et al. (2006), which resulted in a large log OR compared to the other studies, and the P300 data of Grapperon et al. (2004), with a small log OR, were rejected. After these rejections, the tests of homogeneity were not anymore significant for the N100 (Qdf=5 = 8.49, p = 0.13) and the P300 (Qdf=5 = 8.35, p = 0.14) studies. The remaining data, which entered in the meta-analysis, are summarized in Table 3. The estimated ORs are presented in Table 4 and Fig. 1. The OR for the N100 (ORN100) was estimated with the data of 548 patients, the OR for the MMN (ORMMN) with 470 patients, and the OR for the P300 (ORP300) with 313 patients. ORN100 was significantly smaller than ORP300 (p = 0.001) and ORMMN (p = 0.013). ORP300 was larger than ORMMN but this difference was not significant (p = 0.754). The prognosis was the worst for the patient with anoxia or metabolic encephalopathy and the best for trauma or brain surgery (Table 4). According to the meta-analysis performed on the P300 sources, the rate of awakening of patients with stroke or hemorrhage was not significantly different from the OR of patients with anoxia. The OR of anoxic patients was significantly smaller than the ORs for trauma and brain surgery (p < 0.001, with the N100 and MMN sources). The patient suffering from metabolic encephalopathy had an OR significantly smaller than those with trauma (p < 0.001 with the N100 sources and p = 0.001 with the P300 sources) and brain surgery (p < 0.001, with the N100 sources). The ORs for stroke and hemorrhage were significantly smaller than the OR for trauma (p < 0.001, with the P300 sources). The general tendency (i.e., ORP300 > ORMMN > ORN100) observed across all etiologies is also found when ORs are estimated with patient samples grouped by stroke and hemorrhage, trauma, and metabolic encephalopathy etiology (Table 5) but not with anoxic patients (i.e., ORMMN > ORN100 > ORP300). In stroke and hemorrhage patients, the P300 predicts awakening significantly better than the N100 (p < 0.001). The apparent superiority of the P300 compared to the MMN in patients with stroke and hemorrhage, trauma, and metabolic encephalopathy is not significant. In anoxic patients, the MMN seems to be a better predictor than the P300, but this difference does not reach significance. The funnel plots (Fig. 2) did not indicate an asymmetry and thus the possibility of publication bias. Fig. 3 indicates that the OR estimation remains stable against the removal of an individual study.
1 / variance
15
0 0
N100 log OR
4
1 / variance
10
0 0
MMN log OR
3
P300 log OR
5
1 / variance
10
0 0
Fig. 2. Funnel plots. This figure shows the logarithm of the estimated ORs for each individual study included in the meta-analysis and each component of even-related potential on the horizontal axis, and an estimate of their precision (inverse of the variance) is plotted on the vertical axis. The vertical axis cuts the horizontal axis at the summary log ORs. In a few studies, the false positive frequency was null (Table 1). Thus, the estimated log OR were biased and are not plotted. Noteworthy, these biases did not contribute to the studies heterogeneity according to the Qdf statistic (see Section 2).
4. Discussion In patients suffering from stroke, hemorrhage, trauma, or metabolic encephalopathy, the odd ratios (ORs) of the N100 (ORN100), MMN (ORMMN), and P300 (ORP300) were found as follows: ORP300 larger than ORMMN larger than ORN100, with a nonsignificant difference between ORP300 and ORMMN.
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
613
none
none
none
Signorino et al, 1997
Mutschler et al, 2001
Naccache et al, 2005
Lew et al, 2003
Kane et al, 1996
Mutschler et al, 2001 Lew et al, 2003 Guérit el al, 1999
Kane et al, 1996
Fischer et al, 2004
Guerit et al, 1999
Fischer et al, 2006
Fischer et al, 2004
Daltrozzo et al
Daltrozzo et al
Daltrozzo et al
0
5
P300 log OR (95 % CI)
Grapperon et al, 2004
0
5
MMN log OR (95 % CI)
0
2
N100 log OR (95 % CI)
Fig. 3. Assessment of Robustness of the odds ratio (OR) estimation for the N100, MMN, and P300. This figure shows the logarithm of the estimated ORs and 95% confidence intervals (CI) (limit lines) for each component of even-related potential and each meta-analysis performed after removal of each individual study. The largest deviation from the meta-analysis of all studies is due to the removal of a high percentage of patients (i.e., 59.8% if Fischer et al., 2004 is removed from the MMN studies and 46.6% if Mutschler et al., 2001 is removed from the P300 studies) or to a low goodness of fit (Hosmer– Lemeshow p < .05 if Fischer et al., 2006 is removed from the MMN studies). Note that the goodness of fit is correct (Hosmer–Lemeshow p > .05) for all other OR estimations.
This trend (ORP300 > ORMMN > ORN100) with an equivalence between ORP300 and ORMMN was confirmed when the ORs of ERPs were estimated on larger patient samples with a regression analysis and the etiology as explanatory variable. Thus, our data do not confirm that the MMN is a better predictor of awakening than the P300 (Kane et al., 2000). The meta-analysis also indicated that the presence of each of the N100, MMN, and P300 components is a highly significant predictor of awakening, the P300 and MMN being significantly better predictors than the N100. However, since the ERP components are not mandatory evoked potentials, even in healthy participants (Gue´rit, 2004), their absence does not predict nonawakening (see Section 2). It could be argued that the comparison of the ORs between six N100 studies, five MMN studies, and six P300 studies is not meaningful because they were estimated on different patient samples. However, these samples largely overlapped. The fact that the estimated ORs for the etiology variable were almost identical between samples (Table 4) indicates that the differences between them were moderate. Further, each OR was estimated on a set of several studies; thus can be considered independent of methodologies. The homogeneity of the included studies was not only warranted by the Q test (Fleiss and Gross, 1991) but also by an assessment of the robustness of the OR estimation (Fig. 3) indicating stability against removal of an individual study.
A few low responsive patients could have a GOS larger than 2 at the time of the ERP recording. When these patients are more likely to have ERPs present, the prognostic value of ERPs becomes probably less for them. The GOS was assessed between 1 and 12 months after the coma onset for most of the selected sources (Table 1). Although this difference may have slightly affected the results, it should be taken into account that the acute coma state rarely lasts longer than two or three weeks, after which the outcome can be defined. The difference between 1-month and 1-year assessments could rather be critical for the outcome of the subsequent vegetative state (i.e., GOS = 2), which was not the issue of the present study. The fact that the N100 tends to be a poorer predictor of awakening than the two later, more ‘‘cognitive’’ ERP components (except perhaps in anoxic patients where ORMMN > ORN100 > ORP300, Table 5) should probably not be discussed in the present article, before this tendency is confirmed by another, larger study. However, the issue has an important methodological aspect, since some investigators look for the late components only in the presence of N100. In a study of the persistent vegetative state, we showed that some patients can demonstrate very clear, and statistically highly significant, late components (including P300) without any visible sign of N100 (Kotchoubey et al., 2005; Fig. 4). Therefore, defining the presence of one component (which may be a better predictor) as dependent on the presence of another component (which
614
J. Daltrozzo et al. / Clinical Neurophysiology 118 (2007) 606–614
is possibly a worse predictor) would decrease the predictive value of the ERP method as a whole. Our data are in line with most literature indicating that traumatic and post-operative etiologies are related to the best chance of awakening whereas the lowest rate of awakening is obtained for anoxia and metabolic encephalopathy. Thus, Fischer et al. (2004) reported a significantly smaller OR for anoxia compared to traumatic and post-operative etiologies. This consistency could be expected since a big portion of the data used for the estimation of the ORMMN and ORN100 (60% and 51%, respectively) stems from Fischer et al. (2004). Obviously, traumatic and postoperative etiologies often relate to focal brain injuries while anoxia and metabolic encephalopathy are related to very diffuse injuries. Thus, our results confirm that the spatial extension of the brain injury in coma is markedly related to the probability of awakening. For certain etiologies like metabolic encephalopathy or post-operative, the number of patients is small (Table 3). Thus, the reliability of the estimated OR for these etiologies is questionable. 5. Conclusion This study is the first to compare the prognostic power of three ERP components: the N100, the MMN, and the P300, using a statistical measure (the odd ratio), on the basis of a large number of patients (313–548). The analysis reveals that the presence of each of these ERP components significantly predicts awakening. The P300 and MMN had significantly larger predictive powers than the N100. This trend was confirmed in patient samples of stroke and hemorrhage, trauma, and metabolic encephalopathy etiology but not with anoxic. Our data confirm that the etiology is a strong indicator of outcome, the likelihood of subsequent awakening being higher in trauma and post-operative etiologies than in anoxia and metabolic encephalopathy. Acknowledgements This study was supported by the Deutsche Forschungsgemeinshaft (SFB 550 to B.K.) and the French Ministry of Health (PHRC 2004 R-03-03). References De Giorgio CM, Rabinowicz AL, Gott PS. Predictive value of P300 eventrelated potentials compared with EEG and somatosensory evoked potentials in non-traumatic coma. Acta Neurol Scand 1993;87:423–7. Egger M, Davey Smith G, Schneider M, Minder C. Biais in meta-analysis detected by a simple, graphical test. BMJ 1997;315:629–34. Fischer C, Luaute´ J, Ne´moz C, Morlet D, Kirkorian G, Mauguie`re F. Improved prediction of awakening or nonawakening from severe
anoxic coma using tree-based classification analysis. Crit Care Med 2006;34(5):1520–4. Fischer C, Luaute´ J, Adeleine P, Morlet D. Predictive value of sensory and cognitive evoked potentials for awakening from coma. Neurology 2004;63:669–73. Fleiss JL, Gross AJ. Meta-analysis in epidemiology, with special reference to studies of the association between exposure to environmental tobacco smoke and lung cancer: A critique. J Clin Epidemiol 1991;44(2):127–39. Gott PS, Rabinowicz AL, De Giorgio CM. P300 auditory event-related potentials in nontraumatic coma. Association with Glasgow Coma Score and awakening. Arch Neurol 1991;48:1267–70. Grapperon J, Vidal F, Bruschera D, Cantais E, Salinier L, Costes O, et al. Auditive and somato-sensory event-related potentials in comas: Prognostic value for awakening, social and professional reinstatement. Ann Franc¸aises Anesth Re´anim 2004;23:102–8. Gue´rit JM. Prognostic contribution for potentials evoked in unit of intensive care. Ann Franc¸aises Anesth Re´anim 2004;23:99–101. Gue´rit JM, Verougstraete D, De Tourtchaninoff M, Debatisse D, Witdoeckt C. ERPs obtained with the auditory oddball paradigm in coma and alterated states of consciousness: Clinical relationships, pronostic value, and origin of components. Clin Neurophysiol 1999;110:1260–9. Jennett B, Bond M. Assessment of outcome after severe brain damage. Lancet 1975;1(7905):480–4. Kane NM, Butler SR, Simpson T. Coma outcome prediction using eventrelated potentials: P3 and mismatch negativity. Audiol Neurol 2000;5:188–91. Kane NM, Curry SH, Rowlands CA, Manara AR, Lewis T, Moss T, et al. Event related potentials – neurophysiological tools for predicting emergence and early outcome from traumatic coma. Intensive Care Med 1996;22:39–46. Kane NM, Curry SH, Butler SR, Cummins BH. Electrophysiological indicator of awakening from coma. Lancet 1993;341:688. Kotchoubey B, Lang S, Mezger G, Schmalohr D, Schneck M, Semmler A, et al. Information processing in severe disorders of consciousness: Vegetative state and minimally conscious state. Clin Neurophysiol 2005;116:2441–53. Lew HL, Dikmen S, Slimp J, Temkin N, Lee EH, Newell D, et al. Use of somatosensory-evoked potentials and cognitive event-related potentials in predicting outcomes of patients with severe traumatic brain injury. Am J Phys Med Rehabil 2003;82:53–61. Mutschler V, Wioland N, Delabranche X, Le Gourrier L, Calon B. Valeur pronostique des potentiels evoques exogenes et endogenes dans le coma chez l’adulte. In: Gue´rit JM, editor. L‘e´valuation neurophysiologique des comas, de la mort ence´phalique et des e´tats ve´ge´tatifs chroniques. Marseille, France: Solal; 2001. p. 183–94. Naccache L, Puybasset L, Gaillard R, Serve E, Willera JC. Auditory mismatch negativity is a good predictor of awakening in comatose patients: a fast and reliable procedure. Clin Neurophysiol 2005;116:988–9. Reuter PS, Linke DB. P300 and coma. In: Maurer K, editor. Topographic Brain Mapping of EEG and Evoked Potentials. Berlin, Heidelberg, New York, Tokyo: Springer; 1989. p. 192–6. Signorino M, D’acunto S, Cercaci S, Pietropaoli P, Angeleri F. The P300 in traumatic coma: Conditioning of the odd-ball paradigm. J Psychophysiol 1997;11:59–70. Sinson G, D’acunto S, Cercaci S, Pietropaoli P, Angeleri F. Magnetization transfer imaging and proton MR spectroscopy in the evaluation of axonal injury: Correlation with clinical outcome after traumatic brain injury. Am J Neuroradiol 2001;22:143–51. Thompson SG, Sharp SJ. Explaining heterogeneity in meta-analysis: A comparison of methods. Statist Med 1999;18:2693–708.