REENTRANT NETWORK CLASSIFICATION OF EEG IN DEMENTIA

我们提出了一种先进的神经网络识别技术

We propose an advanced neural network technology for recognizing the abnormal brain electrical activity of dementia. Digital EEG (electroencephalographic) studies have successfully differentiated Alzheimer's from multi-infarct dementia, using coherence data from a convential 19-channel EEG. Recent hardware advances by our company have made dense sensor array (64 to 256 channel) EEG inexpensive and convenient for routine clinical evaluation. However, effective clinical use of this advance in measurement requires a significant innovation in classification methodology. Previous classification research with EEG has applied back- propagation or self-organizing map networks. These architectures are inherently limited in their ability to characterize the dynamic properties of multi-channel time-series data, including EEG coherence. In order to surmount this limitation, we propose to apply the nested reentrant and recurrent Helmholtz machine recently developed in our laboratory. The dynamics of this network implement a complex high-dimensional Kalman filter, extracting a minimum-description-length (MDL) model of parameters defining the time-series data. In this Phase I project, the EEG classification performance of the Helmholtz machine architecture will be compared with the performance of a standard back-propagation network. Both time series measures and multivariate coherence are used to reduce the dimensionality of the dense array EEG prior to classification. PROPOSED COMMERCIAL APPLICATION: As drug and other therapies for dementia provide significant improvements in functioning, an inexpensive, repeatable assessment of neurological function could provide quantitative data on therapeutic efficacy in each patient. A brief memory test, accompanied by a quick, comfortable EEG, could become a criterion measure of neurological health to guide the effective medical management of each aged person.