Electroencephalogram signal learning model to improve the detection of absence seizures in neuropediatric infant patients

Abstract

  The objective of this research is to determine the impact 
of applying artificial intelligence-based models to 
electroencephalogram frequencies in order to improve the 
detection of absence seizures in neuropediatric infant patients, 
such as supervised neural networks, SOM neural networks, 
nearest neighbor, decision trees, and random forests, as well 
as to find the relationship between channels at the time of the 
seizure. The methodology used is applied, with an explanatory 
level of research, and the research design is experimental. The 
sample used consists of four absence seizure events in 2,256 
seconds, applying the Gabor filter to the frequencies prior 
to entering the models so that they become input patterns. 
At the moments of absence crisis, a channel coherence of 
0.63 was identified, highlighting that at the moment of 
crisis all channels follow the same common pattern. The 
correlation coefficient R2 = 0.77 and a minimum R2 of 0.57 
were identified, indicating the similarity of frequencies at the 
moment of crisis. A very high standard deviation was identified, 
highlighting the polypoint tails with more than 5 peaks per 
crisis in each channel. In testing the artificial intelligence
based models, the sensitivity, specificity, precision, and accuracy 
values obtained for each model with respect to identifying 
non-crises, crises, and pre-crises were 0.99, 1.0, 0.99, and 0.93 
for the back propagation artificial neural network; 0.99, nan, 
0.99, 0.99, for the nearest neighbor 0.99, 0.0, 0.99, 0.97, for 
decision tree 0.99, 0.0, 0.99, 0.97, and random forest 0.99, nan, 
0.99, 0.97, respectively. Therefore, it concludes that there is 
correct data collection and processing with the learning models 
to identify seizures.