Lecture Notes in Computer Science
Download 12.42 Mb. Pdf ko'rish
|
- Bu sahifa navigatsiya:
- Reproducibility Analysis of Event-Related fMRI Experiments Using Laguerre Polynomials
- Keywords
- 2.1 Reproducibility Analysis
F7
F8 T3 C3 C4 T4 Oz Fp2 O2 Fp1 O1 T5 T6 Fpz P3 P4 Cz Fz F4 F3 Pz 1 2 3 3 4 5 Fig. 3. The 21 electrodes used for EEG recording, distributed according to the 10– 20 international placement system [8]. The clustering into 5 zones is indicated by the colors and dashed lines (1 = frontal, 2 = left temporal, 3 = central, 4 = right temporal and 5 = occipital). over all pairs of zones. (Results for local synchrony and individual frequency bands will be presented in a longer report, including a detailed description of the influence of various parameters such as model order and embedding dimen- sion on the sensitivity.) The p-values, obtained by the Mann-Whitney test, need strictly speaking to be Bonferroni corrected; since we consider many different measures simultaneously, it is likely that a few of those measures have small p-values merely due to stochastic fluctuations (and not due to systematic dif- ference between MCI and control patients). In the most conservative Bonferroni post-correction, the p-values need to be divided by the number of synchrony measures. From the table, it can be seen that only a few measures evince significant differences in EEG synchrony between MCI and control patients: full-frequency DTF and ρ are the most sensitive (for the data set at hand), their p-values remain significant (p corr
< 0.05) after Bonferroni correction. In other words, the effect of MCI and AD on EEG synchrony can be detected, as was reported earlier in the literature; we will expand on this issue in the following section. In other to gain more insight in the relation between the different measures, we calculated the correlation between them (see Fig. 5; red and blue indicate strong correlation and anti-correlation respectively). From this figure, it becomes strikingly clear that the majority of measures are strongly correlated (or anti- correlated) with each other; in other words, the measures can easily be classified in different families. In addition, many measures are strongly (anti-)correlated with the classical cross-correlation coefficient r, the most basic measure; as a result, they do not provide much additional information regarding EEG syn- chrony. Measures that are only weakly correlated with the cross-correlation co- efficient include the phase synchrony indices, Granger causality measures, and stochastic-event synchrony measures; interestingly, those three families of syn- chrony measures are mutually uncorrelated, and as a consequence, they each seem to capture a specific kind of interdependence. A Comparative Study of Synchrony Measures for the Early Detection of AD 121
Table 1. Sensitivity of synchrony measures for early prediction of AD (p-values for Mann-Whitney test; * and ** indicate p < 0.05 and p < 0.005 respectively) Measure Cross-correlation Coherence Phase Coherence Corr-entropy Wave-entropy p-value 0.028
∗ 0.060
0.72 0.27
0.012 ∗ References [8] [9]
Measure Granger coherence Partial Coherence PDC DTF
ffDTF dDTF
p-value 0.15
0.16 0.60
0.34 0.0012
∗∗ 0.030
∗ References [4] Measure
Kullback-Leibler R´ enyi Jensen-Shannon Jensen-R´ enyi I
I p-value
0.072 0.076
0.084 0.12
0.060 0.080
References [15]
[14] Measure
N k S k H k S-estimator p-value
0.032 ∗ 0.29 0.090 0.33
References [6]
[13] Measure
Hilbert Phase Wavelet Phase Evolution Map Instantaneous Period p-value
0.15 0.082
0.072 0.020
∗ References [6] [12]
Measure s t ρ p-value
0.92 0.00029
∗∗ In Fig. 4, we combine the two most sensitive synchrony measures (for the data set at hand), i.e., full-frequency DTF and ρ. In this figure, the MCI patients are fairly well distinguishable from the control patients. As such, the separation is not sufficiently strong to yield reliable early prediction of AD. For this purpose, the two features need to be combined with complementary features, for example, derived from the slowing effect of AD on EEG, or perhaps from different modal- ities such as PET, MRI, DTI, or biochemical indicators. On the other hand, we remind the reader of the fact that in the data set at hand, patients did not carry out any specific task; moreover, the recordings were short (only 20s). It is plausible that the sensitivity of EEG synchrony could be further improved by increasing the length of the recordings and by recording the EEG before, while, and after patients carry out specific tasks, e.g., working memory tasks. 0.045
0.05 0.055
0.06 0.15
0.2 0.25
0.3 0.35
0.4 0.45
0.5 MCI
CTR ρ F 2 ij Fig. 4. ρ vs. ffDTF 122 J. Dauwels, F. Vialatte, and A. Cichocki 5 10
20 25 30 5 10 15 20 25 30 −0.8 −0.6
−0.4 −0.2
0 0.2
0.4 0.6
0.8 state space corr/coh mut inf
divergence Granger
SES phase
N k (X |Y ) N k (Y |X)
S k (X |Y ) S k (Y |X)
H k (X |Y ) H k (Y |X)
S est
r c r E w E I W I γ H γ W φ EMA IPA K(Y
|X) K(X
|Y ) K D α J J α K ij C ij P ij γ ij F ij χ ij s t ρ Fig. 5. Correlation between the synchrony measures 4 Conclusions In previous studies, brain dynamics in AD and MCI patients were mainly in- vestigated using coherence (cf. Section 2.2) or state space based measures of synchrony (cf. Section 2.7). During working memory tasks, coherence shows sig- nificant effects in AD and MCI groups [26] [27]; in resting condition, however, coherence does not show such differences in low frequencies (below 30Hz), nei- ther between AD and controls [28] nor between MCI and controls [27]. These results are consistent with our observations. In the gamma range, coherence seems to decrease with AD [29]; we did not investigate this frequency range, however, since the EEG signals analyzed here were band pass filtered between 4 and 30Hz. Synchronization likelihood, a state space based synchronization measure simi- lar to the non-linear interdependence measures S k , H
k , and N
k (cf. Section 2.7), is believed to be more sensitive than coherence to detect changes in AD pa- tients [28]. Using state space based synchrony methods, significant differences were found between AD and control in rest conditions [28] [30] [32] [33]. State space based synchrony failed to retrieve significant differences between MCI patient and control subjects on a global level [32] [33], but significant effects were observed locally: fronto-parietal electrode synchronization likelihood pro- gressively decreased through MCI and mild AD groups [30]. We report here a lower p-value for the state space based synchrony measure N k (p = 0.032) than for coherence (p = 0.06); those low p-values, however, would not be statistically significant after Bonferroni correction. A Comparative Study of Synchrony Measures for the Early Detection of AD 123
By means of Global Field Synchronization, a phase synchrony measure similar to the ones we considered in this paper, Koenig et al. [31] observed a general decrease of synchronization in correlation with cognitive decline and AD. In our study, we analyzed five different phase synchrony measures: Hilbert and wavelet based phase synchrony, phase coherence, evolution map approach (EMA), and instantaneous period approach (IPA). The p-value of the latter is low (p=0.020), in agreement with the results of [31], but it would be non-significant after Bonferroni correction. The strongest observed effect is a significantly higher degree of local asyn- chronous activity (ρ) in MCI patients, more specifically, a high number of non- coincident, asynchronous oscillatory events (p = 0.00029). Interestingly, we did not observe a significant effect on the timing jitter s t of the coincident events (p = 0.92). In other words, our results seem to indicate that there is significantly more non-coincident background activity, while the coincident activity remains well synchronized. On the one hand, this observation is in agreement with pre- vious studies that report a general decrease of neural synchrony in MCI and AD patients; on the other hand, it goes beyond previous results, since it yields a more subtle description of EEG synchrony in MCI and AD patients: it suggests that the loss of coherence is mostly due to an increase of (local) non-coincident background activity, whereas the locked (coincident) activity remains equally well synchronized. In future work, we will verify this conjecture by means of other data sets. References 1. Jong, J.: EEG Dynamics in Patients with Alzheimer’s Disease. Clinical Neurophys- iology 115, 1490–1505 (2004) 2. Pereda, E., Quiroga, R.Q., Bhattacharya, J.: Nonlinear Multivariate Analysis of Neurophsyiological Signals. Progress in Neurobiology 77, 1–37 (2005) 3. Breakspear, M.: Dynamic Connectivity in Neural Systems: Theoretical and Em- pirical Considerations. Neuroinformatics 2(2) (2004) 4. Kami´
nski, M., Liang, H.: Causal Influence: Advances in Neurosignal Analysis. Crit- ical Review in Biomedical Engineering 33(4), 347–430 (2005) 5. Stam, C.J.: Nonlinear Dynamical Analysis of EEG and MEG: Review of an Emerg- ing Field. Clinical Neurophysiology 116, 2266–2301 (2005) 6. Quiroga, R.Q., Kraskov, A., Kreuz, T., Grassberger, P.: Performance of Different Synchronization Measures in Real Data: A Case Study on EEG Signals. Physical Review E 65 (2002) 7. Sakkalis, V., Giurc˘ aneacu, C.D., Xanthopoulos, P., Zervakis, M., Tsiaras, V.: As- sessment of Linear and Non-Linear EEG Synchronization Measures for Evaluating Mild Epileptic Signal Patterns. In: Proc. of ITAB 2006, Ioannina-Epirus, Greece, October 26–28 (2006) 8. Nunez, P., Srinivasan, R.: Electric Fields of the Brain: The Neurophysics of EEG. Oxford University Press, Oxford (2006) 9. Xu, J.-W., Bakardjian, H., Cichocki, A., Principe, J.C.: EEG Synchronization Measure: a Reproducing Kernel Hilbert Space Approach. IEEE Transactions on Biomedical Engineering Letters (submitted to, September 2006)
124 J. Dauwels, F. Vialatte, and A. Cichocki 10. Herrmann, C.S., Grigutsch, M., Busch, N.A.: EEG Oscillations and Wavelet Anal- ysis. In: Handy, T. (ed.) Event-Related Potentials: a Methods Handbook, pp. 229– 259. MIT Press, Cambridge (2005) 11. Lachaux, J.-P., Rodriguez, E., Martinerie, J., Varela, F.J.: Measuring Phase Syn- chrony in Brain Signals. Human Brain Mapping 8, 194–208 (1999) 12. Rosenblum, M.G., Cimponeriu, L., Bezerianos, A., Patzak, A., Mrowka, R.: Identi- fication of Coupling Direction: Application to Cardiorespiratory Interaction. Phys- ical Review E, 65 041909 (2002) 13. Carmeli, C., Knyazeva, M.G., Innocenti, G.M., De Feo, O.: Assessment of EEG Synchronization Based on State-Space Analysis. Neuroimage 25, 339–354 (2005) 14. Kraskov, A., St¨ ogbauer, H., Grassberger, P.: Estimating Mutual Information. Phys. Rev. E 69(6), 66138 (2004) 15. Aviyente, S.: A Measure of Mutual Information on the Time-Frequency Plane. In: Proc. of ICASSP 2005, Philadelphia, PA, USA, March 18–23, vol. 4, pp. 481–484 (2005)
16. Aviyente, S.: Information-Theoretic Signal Processing on the Time-Frequency Plane and Applications. In: Proc. of EUSIPCO 2005, Antalya, Turkey, Septem- ber 4–8 (2005) 17. Quiroga, Q.R., Rosso, O., Basar, E.: Wavelet-Entropy: A Measure of Order in Evoked Potentials. Electr. Clin. Neurophysiol (Suppl.) 49, 298–302 (1999) 18. Blanco, S., Quiroga, R.Q., Rosso, O., Kochen, S.: Time-Frequency Analysis of EEG Series. Physical Review E 51, 2624 (1995) 19. Dauwels, J., Vialatte, F., Cichocki, A.: A Novel Measure for Synchrony and Its Application to Neural Signals. In: Honolulu, H.U. (ed.) Proc. IEEE Int. Conf. on Acoustics and Signal Processing (ICASSP), Honolulu, Hawai’i, April 15–20 (2007) 20. Vialatte, F., Martin, C., Dubois, R., Haddad, J., Quenet, B., Gervais, R., Dreyfus, G.: A Machine Learning Approach to the Analysis of Time-Frequency Maps, and Its Application to Neural Dynamics. Neural Networks 20, 194–209 (2007) 21. Chapman, R., et al.: Brain Event-Related Potentials: Diagnosing Early-Stage Alzheimer’s Disease. Neurobiol. Aging 28, 194–201 (2007) 22. Cichocki, A., et al.: EEG Filtering Based on Blind Source Separation (BSS) for Early Detection of Alzheimer’s Disease. Clin. Neurophys. 116, 729–737 (2005) 23. Hogan, M., et al.: Memory-Related EEG Power and Coherence Reductions in Mild Alzheimer’s Disease. Int. J. Psychophysiol. 49 (2003) 24. Musha, T., et al.: A New EEG Method for Estimating Cortical Neuronal Impair- ment that is Sensitive to Early Stage Alzheimer’s Disease. Clin. Neurophys. 113, 1052–1058 (2002) 25. Vialatte, F., et al.: Blind Source Separation and Sparse Bump Modelling of Time- Frequency Representation of EEG Signals: New Tools for Early Detection of Alzheimer’s Disease. In: IEEE Workshop on Machine Learning for Signal Pro- cessing, pp. 27–32 (2005) 26. Hogan, M.J., Swanwick, G.R., Kaiser, J., Rowan, M., Lawlor, B.: Memory-Related EEG Power and Coherence Reductions in Mild Alzheimer’s Disease. Int. J. Psy- chophysiol. 49(2), 147–163 (2003) 27. Jiang, Z.Y.: Study on EEG Power and Coherence in Patients with Mild Cognitive Impairment During Working Memory Task. J. Zhejiang Univ. Sci. B 6(12), 1213– 1219 (2005) 28. Stam, C.J., van Cappellen van Walsum, A.M., Pijnenburg, Y.A., Berendse, H.W., de Munck, J.C., Scheltens, P., van Dijk, B.W.: Generalized Synchronization of MEG Recordings in Alzheimer’s Disease: Evidence for Involvement of the Gamma Band. J. Clin. Neurophysiol. 19(6), 562–574 (2002) A Comparative Study of Synchrony Measures for the Early Detection of AD 125
29. Herrmann, C.S., Demiralp, T.: Human EEG Gamma Oscillations in Neuropsychi- atric Disorders. Clinical Neurophysiology 116, 2719–2733 (2005) 30. Babiloni, C., Ferri, R., Binetti, G., Cassarino, A., Forno, G.D., Ercolani, M., Ferreri, F., Frisoni, G.B., Lanuzza, B., Miniussi, C., Nobili, F., Rodriguez, G., Rundo, F., Stam, C.J., Musha, T., Vecchio, F., Rossini, P.M.: Fronto-Parietal Coupling of Brain Rhythms in Mild Cognitive Impairment: A Multicentric EEG Study. Brain Res. Bull. 69(1), 63–73 (2006) 31. Koenig, T., Prichep, L., Dierks, T., Hubl, D., Wahlund, L.O., John, E.R., Jelic, V.: Decreased EEG Synchronization in Alzheimer’s Disease and Mild Cognitive Impairment. Neurobiol. Aging 26(2), 165–171 (2005) 32. Pijnenburg, Y.A., Made, Y.v., van Cappellen, A.M., van Walsum, Knol, D.L., Scheltens, P., Stam, C.J.: EEG Synchronization Likelihood in Mild Cognitive Im- pairment and Alzheimer’s Disease During a Working Memory Task. Clin. Neuro- physiol. 115(6), 1332–1339 (2004) 33. Yagyu, T., Wackermann, J., Shigeta, M., Jelic, V., Kinoshita, T., Kochi, K., Julin, P., Almkvist, O., Wahlund, L.O., Kondakor, I., Lehmann, D.: Global dimensional complexity of multichannel EEG in mild Alzheimer’s disease and age-matched cohorts. Dement Geriatr Cogn Disord 8(6), 343–347 (1997) M. Ishikawa et al. (Eds.): ICONIP 2007, Part I, LNCS 4984, pp. 126–134, 2008. © Springer-Verlag Berlin Heidelberg 2008 Reproducibility Analysis of Event-Related fMRI Experiments Using Laguerre Polynomials Hong-Ren Su 1,2 , Michelle Liou 2, *, Philip E. Cheng 2 ,
John A.D. Aston 2 , and Shang-Hong Lai 1
1 Dept. of Computer Science, National Tsing Hua University, Hsinchu, Taiwan 2 Institute of Statistical Science, Academia Sinica, Taipei, Taiwan mliou@stat.sinica.edu.tw
polynomials for analyzing data collected in event-related functional magnetic resonance imaging (fMRI) experiments. This particular family of polynomials has been widely used in the system identification literature and recommended for modeling impulse functions in BOLD-based fMRI experiments. In empirical studies, we applied Laguerre polynomials to analyze data collected in an event- related fMRI study conducted by Scott et al. (2001). The experimental study investigated neural mechanisms of visual attention in a change-detection task. By specifying a few meaningful Laguerre polynomials in the design matrix of a random effect model, we clearly found brain regions associated with trial onset and visual search. The results are consistent with the original findings in Scott et al. (2001). In addition, we found the brain regions related to the mask presence in the parahippocampal, superior frontal gyrus and inferior parietal lobule. Both positive and negative responses were also found in the lingual gyrus, cuneus and precuneus. Keywords: Reproducibility analysis, Event-related fMRI. 1 Introduction We previously proposed a methodology for assessing reproducibility evidence in fMRI studies using an on-and-off paradigm without necessarily conducting replicated experiments, and suggested interpreting SPMs in conjunction with reproducibility evidence (Liou et al., 2003; 2006). Empirical studies have shown that the method is robust to the specification of hemodynamic response functions (HRFs). Recently, BOLD-based event-related fMRI experiments have been widely used as an advanced alternative to the on-and-off design for studies on human brain functions. In event- related fMRI experiments, the duration of stimulus presentation is generally longer and there are no obvious contrasts between the experimental and control conditions to be used in data analyses. In order to detect possible brain activations during stimulus presentation and task performance, there have been a variety of event-related HRFs proposed in the literature. In this study, we introduce the use of orthogonal causal
* Corresponding author.
Reproducibility Analysis of Event-Related fMRI Experiments 127 Laguerre polynomials for modeling response functions. This particular family of polynomials has been widely used in the system identification literature and was recommended for modeling impulse functions in fMRI experiments (Saha et al., 2004). In the empirical study, we applied Laguerre polynomials to analyze data in the study by Scott et al. (2001). The dataset was published by the US fMRI Data Center and is available for public access. The original experiment involved 10 human subjects and investigated brain functions associated with a change-detection task. In the experimental task, subjects look attentively at two versions of the same picture in alternation, separated by a brief mask interval. The experiment additionally analyzed behavioral responses that subjects detected something changing between pictures and pressed a button with hands. In our reproducibility analysis, a few meaningful Laguerre polynomials matching the experimental design were inserted into a random effect model and reproducibility analyses were conducted based on the selected polynomials. In the analyses, we successfully located brain regions associated with the visual change-detection task similar to those found in Scott et al.. Additionally, we found other interesting brain regions that were not included in the previous study.
In this section, we will briefly describe the method for investigating the reproducibility evidence in fMRI experiments, and outline the family of Laguerre polynomials including those used in our empirical study. 2.1 Reproducibility Analysis In the SPM generalized linear model, the fMRI responses in the i th run can be expressed as i i i i e X y + = β , (1) where y i is the vector of image intensity after pre-whitening, X i is the design matrix, and
β is the vector containing the unknown regression parameters. In the random effect model, the regression parameters i β are additionally assumed to be random from a multivariate Gaussian distribution with common mean μ and variance Ω . The empirical Bayes estimate of i β in the random effect model would shrink all estimates toward the mean μ , with greater shrinkage at noisy runs. In fMRI studies, the true status of each voxel is unknown, but can be estimated using the t-values (i.e., standardized β estimates) within individual runs derived from the random effect model along with the maximum likehood estimation method. By specifying a mixed multinomial model, the receiver-operation characteristic (ROC) curve can be estimated using the maximum likelihood estimation method and t-values of all image voxels. The curve is simply a bivariate plot of sensitivity versus the false alarm rate. The threshold (or the operational point) on the ROC curve for classifying voxels into the active/inactive status was found by maximizing the kappa value. We follow the 128 H.-R. Su et al. same definition in Liou et al. (2006) to categorize voxels according to reproducibility, that is, a voxel is strongly reproducible if its active status remains the same in at least 90% of the runs, moderately reproducible in 70-90% of the runs, weakly reproducible in 50-70% of the runs, and otherwise not reproducible. The brain activation maps are constructed on the basis of strongly reproducible voxels, but include voxels that are moderately reproducible and spatially proximal to those strongly reproducible voxels. Download 12.42 Mb. Do'stlaringiz bilan baham: |
ma'muriyatiga murojaat qiling