PATTERN RECOGNITION OF ELECTRO-
CD 1967016 E&MP44.148
October 24, 1963
At the Tulane Bio-Medical Computing System Center, where Tulane University and International Business Machines Corporaton are engaged in a joint study, the status of an electrocardiogram displayed on a monitor screen is checked against values computed and printed by an IBM 1401 Data Processing System.
B.G. Gilmore of IBM performs the comparison. The computer is used at the Center to detect significant changes in electrocardiogram wave forms that may indicate cardiac disorders.
additional text found
PATTERN RECOGNITION OF ELECTROENCEPHALOGRAMS
Electroencephalograms, or EEGs are weak electrical signals obtained from eletrodes placed on a persons head. These brain wave signals represent the state of cell activity in the brain, and their interpretation is a major analytical problem.
Over the years, physicians and scientists have correlated certain waveforms with the level of an individuals consciousness, with brain damage which might be present or with certain kinds of brain ailments. In the joint Tulane-IBM study, digital computer techniques are used to extract and identify such significant information from the EEG signals.
In some of these experiments, in which a pattern recognition technique is used, a number index is computed automatically to indicate the deviation from a normal EEG pattern that is stored in the computer. This typical pattern is obtained by first processing normal EEG waveforms.
Besides recording EEGs as visible waveforms on a graphic chart, special data acquisition instruments record the same signals on magnetic tape. An experimental device multiplies the low-frequency EEG signals by a factor of 240. This compresses a 10-minute recording to about 2.5 seconds.
The signal is divided into 36 different frequencies by a bank of filters. All of these 36 new symbols are then analyzed simultaneously.
The next step involves identifying and classifying amplitude peaks and valleys. This is done for each of the 36 readings that are available simultaneously. The filter outputs are coded as either peaks or valleys, and they may be high, medium or low. The high peaks are those with maximum amplitude in their respective categories. Included in the high designation are those filters with amplitude within 70 per cent of the highest levels. Those from 30-70 per cent are medium peaks and those under 30 per cent are classed as low. A similar cirteria is used to classify the valleys.
With the EEG information now represented by a sequence of such coding, the information is presented as imput to an experimental adaptive pattern recognition program on an IBM 7094 Data Processing System in Yorktown, N.Y. Two stages are included in this program -- a learning phase which develops a definition of a normal EEG, and a testing phase which indicates how closely an unknown EEG matches the definition of a normal EEG.
In the learning mode, information on peaks and valleys for pairs of frequency channels both within and between time instants is obtained from EEGs previously designated as normal. The definition of normality is refined by adding one normal EEG sample at a time. A computed number gives an approximate indication of information learned. When new samples do not significantly increase this number, enough information has been obtained to determine the variation existing among normal EEGs. While this does not imply that good discrimination between normal and abnormal EEGs will be obtained, it does show that the definition of normality has been stabilized.
In the testing mode, the program produces a number between zero and one that indicates how well an unknown EEG matches the definition of normality established in the leraning mode. If the match number is one, there is a perfect correspondence with this definition.
Original Caption by Science Service
©Tulane University, ©International Business Machines (IBM)