Weiterentwicklung von Methoden der Datenanalyse zur räumlichen und zeitlichen Charakterisierung einzelner Quellsignale in MEG- und EEG-Daten bei sensorischer Stimulation
Zusammenfassung der Projektergebnisse
The objective of this work was the assessment and the development of blind source separation algorithms for the decomposition of evoked MEG data. Based on evoked MEG experiments and mathematical modeling, we have analyzed at which dependency level independent component analysis (ICA) is suited to decompose evoked signals and at which level it fails. For the latter case, a new separation method for evoked dependent signals was developed. Furthermore, the separation results were assessed for single-trial analysis in terms of the performance of subsequently estimated single-trial latencies. Our first task was to find signal configurations with adjustable dependencies. We considered a basic but trial-to-trial variant evoked MEG signal model. Specifically, we assumed the signal form of each source to be constant and the latencies to be variant over trials. Indeed, two signal sources that we generated by using equal signal forms and latencies for all trails, resulted in two equal signals, which were maximally dependent. Changing the latencies randomly (normal distribution), we were able to systematically adjust mutual information of signals by changing the standard deviation parameter. For convenience, this parameter was named 'jitter'. The higher the jitter the more independent the signals were and vice versa. Subsequently, established ICA algorithms were extensively assessed using the new simulation tool. We found that the performance of all established ICA methods degraded closely correlated to an increase of mutual information between two signals. Furthermore, for closely spaced dipoles, each estimated field pattem showed a dipolar structure, which is commonly interpreted as a successful decomposition; it was found to fail. Specifically, an important contribution was that ICA can fail to separate evoked MEG signals, while the user may not be aware of this failure. Next, alternative methods were investigated. The cluster and tree component analysis was found to work well if the dependent signals were distributed following a tree structure. If not, it was shown to also fail. Likewise, frequency band approaches were observed to fail if assumptions were not met. The origin of the problem was observed to be the bi-Iinear model underlying ICA, which is not unique, inherently. Hence, additional assumptions were always needed, which may not be justified by the recorded MEG data. However, structure in time, space and stimulation events come along with all evoked MEG data. Hence, our starting point for a new method was the tri-linear decomposition technique CANDECOMP/PARAFAC (CP). This method uses structure in three modalities and yields a unique decomposition if all modalities enter linearly. No additional assumptions are needed. However, latency shifts violated the tri-linear model assumption and also CP was operated outside its model. Consequently, we aimed at considering latency shifts resulting in the non-linear but unique shifted factor analysis (SFA). Specifically, we used a Taylor series expansion of our assumed MEG model, which resulted in a new algorithm that included the estimation of latency shifts. We showed that our algorithm was robust along a wide range of latency variations. Most significantly, SFA outperformed the ICA methods for dependent and almost independent evoked signals. Furthermore, latency shifts that were inherently estimated with the new SFA method, were improved compared to a maximum likelihood latency estimation method. An audio-visual MEG experiment was designed and high quality data with highly dependent signals were recorded at the PTB laboratory. A novel way of separation assessment was developed. ICA was shown to fail for this data, while SFA gave good results.
Projektbezogene Publikationen (Auswahl)
- Performance of ICA for dependent sources using synthetic stimulus evoked MEG data. DGBMT workshop proceedings, pp. 32-35, 2008
F. Kohl, G. Wübbeler, T. Sander, L. Trahms, D. Kolossa, R. Orglmeister, C. Elster, M. Bär
- Performance of ICA for MEG data generated from subspaces with dependent sources. ECIFMBE 2008 proceedings, vol. 22, pp. 1281-1285, 2008
F. Kohl, G. Wübbeler, D. Kolossa, R. Orglmeister, C. Elster, M. Bär
- Non-independent BSS; a model for evoked MEG signals with controllable dependencies. ICA 2009 proceedings, lect. notes comp. sci. 2009, vol. 5441, pp. 443-450, 2009
F. Kohl, G. Wübbeler, D. Kolossa, C. Elster, M. Bär, R. Orglmeister
- Classifying ICA components of evoked MEG data. Biomed. Tech. 2010 proceedings, vol. 55, 2010
D. Ghaemi, F. Kohl, R.Orglmeister
- Noise adjusted PCA for finding the subspace of evoked dependent signals from MEG data. LVA/ICA 2010 proceedings, lect notes comp. sci. 2010, vol. 6365, pp. 442-449, 2010
F. Kohl, G. Wübbeler, D. Kolossa, C. Elster, M. Bär, R. Orglmeister
- Recent advances in modeling and analysis of bioelectric and biomagnetic sources. Biomed. Tech.., vol. 55, pp. 65-76, 2010
T.H. Sander, T.R. Knösche, A. Schlögl, F. Kohl, CH. Wolters, J. Haueisen, and L. Trahms
- Shifted factor analysis for the separation of evoked dependent MEG signals. Phys. Med. Biol., vol. 55, pp. 4219-4230, 2010
F. Kohl, G. Wübbeler, D. Kolossa, M. Bär, R. Orglmeister and C. Elster
