The purpose of this project is to identify and classify Event Related Potential using competitive Neural Net architecture.
ABSTRACT
The purpose of this project is to identify and classify Event Related Potential using competitive Neural Net architecture. The network weights converge to the embedded signal patterns, resulting in the formation of a matched filter bank.. The network performance is analyzed via a simulation study, exploring identification robustness under low SNR conditions. The classifier is applied to real event-related potential data recorded during a classic odd-ball type paradigm, signal patterns are automatically identified, dismissing the strong and limiting requirement of a-priori stimulus-related selective grouping of the recorded data.
BACKGROUND
The human brain is a vast network of microscopic nerve cells connected by synapses. In this network, many extremely weak electrical signals are formed and the collective result of which is a faint electrical signal that can be measured directly from the scalp. This signal is what is called the Electroencephalogram or EEG. It is a well established fact, that there is a direct correlation between the process of thought and the EEG. Subsequently, research has concentrated on time-locked brain activity, related to specific events, external or internal to the subject. This time-locked activity, referred to also as Event Related Potentials (ERP’s), is regarded as a manifestation of brain processes related to preparation for or in response to discrete events.
The ongoing electrical activity of the brain, the EEG, is comprised of relatively slow fluctuations, in the range of 0.1-100 Hz, with magnitudes of 10-100. ERP’s are characterized by overlapping spectra with the EEG, but with significantly lower magnitudes of 0.1-10 . The unfavorable Signal to Noise Ration (SNR) requires filtering of the raw signals to enable analysis of the time-locked signals. The common method used for this purpose is signal averaging, synchronized to repeated occurrences of a specific event. Averaging-based techniques assume a deterministic signal within the averaged session, and thus signal variability can not be modeled unless a-priori stimulus -or response – based categorization is available. It is the purpose of this project to provide an alternative working method to enhance conventional averaging techniques, and thus facilitating identification and analysis of variable brain responses.
BASIC APPROACH
Competitive learning is a well-known branch of the general unsupervised learning theme. The elementary principles of competitive learning yield a learning paradigm where individual units learn to specialize on sets of similar patterns and thus become feature detectors. The detector is supposed to discover statistically salient features of the input population, without a-priori categorization into which the patterns are to be classified. Thus the detector needs to develop its own featural representation of the population of input patterns capturing its most salient features.
The complicated, generally unknown relationships between the stimulus and its associated brain response, and the extremely low SNR of the brain responses which are practically masked by the background brain activity, make the choice of a self organizing structure for post-stimulus epoch analysis most appropriate. The competitive network, having the property that its weights converge to the actual embedded signal patterns while inherently averaging out the additive background EEG, is thus an evident choice. The competition is held among the neurons to respond to the pattern appearing on the input layer of the network. The learning rules are such that the weights of the winning neuron are shifted towards the input pattern; thus assuming zero-mean additive background EEG, once converged, the network operates as a matched filter bank classifier. Thus our goal is achieved.
TOOLS
The simulations were programmed in Matlab for Windows.
CONCLUSIONS
In order to estimate the performance of the competitive network classification system a simulation was carried out The simulation included embedding a deterministic signal, simulating the ERP’s, in a synthesized background activity at different SNR levels, and training the two neurons competitive network with 1000 sweeps (per SNR level). Due to the simulation results, complete classification is obtained for SNR higher than -2dB. Fig. 1 shows the percentage of Identification of the input signals. Then the network was supplied by real 87 ERP’s recorded during a classic odd-ball type paradigm. These ERP’s contain two types of EP embedded in the background activity EEG with SNR of 0dB. The network converged after approximately 400 iterations (per neuron). The automatic identification procedure has provided two categories, with almost perfect matches (about 90%) to the stimulus-related selective averaged signals. The obtained categorization confirms the usage of averaging methods for this classic experiment, and thus presents an important result in itself.
Figure 1: classification & identification percentage for varying SNR levels
ACKNOWLEDGMENTS
We would like to thank our supervisor Danny Lange for his guidance throughout the project. Also we wish to thank the laboratory staff for their technical support. This project was supported by the Ollendorf Center Research Fund.