EEG processing based on IFAST system and Artificial

EEG processing based on IFAST system and Artificial Neural Networks for early detection of Alzheimer’s disease

Introduction • • Quantitative EEG analysis has been used for many years to understand functional brain modifications associated with dementia and particularly with Alzheimer Disease (AD). In the most of the studies, EEG signal was analyzed into time-frequency domain, using Fast Fourier Transform, Wavelet, etc. For the blind classification test, linear and non linear models was used. Overall accuracy of these researches generally goes from 80% to 85%. * AD severe ** AD mild

AIMS Evaluating a new type of ‘dedicated’ neural network (labelled i. FAST ) able to segregate clusters of data and to investigate whether it was correct in discriminating healthy ageing subjects from MCI and early Alzheimer disease demented patients on the only basis of a brief epoch of EEG recording (1 minute)

Electroencephalography (EEG) • EEG data were recorded in resting subjects (eyes-closed) by specialized clinical units • EEG data were recorded from 19 electrodes positioned according to the international 10 -20 system EEG data Resting subjects 10 -20 system Clinical units

Example : EEG Track Subject 1

Database is composed of the following classes of subjects : • 172 AD : MMSE=23(mean). Age 77. 4(mean) • 163 Normal : MMSE=28. 3 (mean). Age=71. 9. • 115 MCI : MMSE=25. 2(mean). Age=76. 9 (mean). For each subject a minute of a 19 channels EEG at 128 hertz was recorded.

A New Method for Data Analysis • This method was invented by Massimo Buscema at Semeion Research Center (April 2005). • One minute of EEG at 128 Hertz of any subject is recorded as a matrix with 19 columns (the channels) and 7680 rows (the time). This amount of data is very huge and we need a mathematical algorithm able to compress a bi-dimensional matrix into a one-dimensional vector, preserving all the key information. In rough words, we need to squash a multivariate sequence of data, representing the EEG of a subject, into a finite number of parameters. • This new method was named I. F. A. S. T. (Implicit Function As Squashing Time). • I. F. A. S. T. algorithm process the all the EEG channels of a subject simultaneously, through a Non Linear Auto-Associative Neural Networks. At the end of the processing phase, the trained connections of the ANN are used as a representative set of parameters (variables) of the EEG sequence.

IFAST SW –Data Processing

An Example of I. F. A. S. T. Algorithm I. F. A. S. T. algorithm works into 2 phases: • Squashing phase : from 19 x 16480 values distributed along space and time to around 400 values into the space domain. • Noise elimination : from around 400 values to around 100 Squashing Noise Reduction Classification 19 Channels 399 weights … … ……. …. . . Same random weights for any EEG dataset 60/128 input … … …………… ………. …. . .

IFAST RESULTS Results obtained using different Non Linear Auto-Associative Neural Networks within IFAST method (Back Propagation, New Recirculation, Multi-layer Perceptron and Recurrent Back Propagation)

CONCLUSIONS • With I. F. A. S. T. algorithm the correct automatic classification rate reached 93. 6% for Nold vs. MCI and 92. 6% for MCI vs. AD. Here we tested the hypothesis that a correct automatic classification of Nold, MCI, and AD subjects can be obtained extracting spatial information content of the resting EEG voltage by ANNs. These results represent the basis for research aimed at integrating spatial and temporal information content of the EEG by ANNs