A neural network based mathematical model for P300 generation

The P300 is a late EP (evoked potential) component that is usually obtained using an odd-ball paradigm. It is endogenous in nature that is, its amplitude and latency depend on the amount of information carried by the stimulus rather than its physical characteristics. The two primary determinants of P300 amplitude are the probability of occurrence of the stimulus and the task relevance associated with it. It has been suggested that P300 reflects the 'context updating' mechanism in the human working memory. We have implemented a neural-network based model of the human working memory simulating the learning and forgetting mechanisms of external stimuli that are responsible for P300 generation. A new learning rule has been implemented to govern the weight dynamics of the network. The output of the network is modeled to resemble the P300 and the results from simulations prove that the model output shows a relationship to stimulus probability and task relevance that is similar to experimental results.We tested the sensitivity of model parameters on the model output. Subsequently we added noise to the system which increased the variability in the model-generated response in a way that is consistent with the experimental results. Based on model simulations, we predicted that the P300 amplitude depends on the difference in the frequency complexity of external stimuli used in odd-ball paradigms. We conducted actual EEG experiments using odd-ball paradigms with various combinations of pure tones and a complex tone, and though the hypothesis was not confirmed, we found that complex sounding stimuli generate larger late EP components, which may be related to subject's arousal level. We also conducted experiments using a repeating sequence of two stimuli to study the effect of subject expectation on P300 amplitude. We found that the ability to detect repeating patterns in stimulus sequence and consequently predict their occurrence is probabilistic in nature. Subjects who could not detect the repeating pattern did not show any difference in P300 amplitude as compared with an odd-ball paradigm (with the same GTP---Global Target Probability) whereas subjects who did detect the repeating stimulus sequence generated significantly smaller P300 amplitudes. Correction procedures have been suggested to accommodate for inconsistencies such as the relationship with frequency complexity, the relationship with subject's arousal level and the detection of repeating stimuli patterns. Also, the relationship between P300 amplitude and factors apart from probability and task relevance, such as ISI (Inter-stimulus interval) and TTI (target-target interval), were tested and acceptable results were obtained. Overall, the study resulted in a P300 model that mimics many aspects of the nervous system responsible for P300 generation. Additional work is needed to refine the model and to test the effects of neurological disorders.