Abstract
Electroencephalography (EEG) is the measurement of differences in electrical potential between points on the scalp. Electroencephalography is primarily utilized for scanning and recording the electrical activity of a patient’s brain in response to stimuli and can be a useful tool in monitoring changes to a patient’s brain as they age. With modern population growth and an increase of the average lifespan, more patients are becoming affected by neurodegenerative diseases such as dementia and Alzheimer’s. Patients with histories of epileptic seizures have a much larger risk of developing Alzheimer’s and other neurodegenerative diseases. As humans age, the types, or ranges, of frequencies read by these electroencephalograms can change. One in ten individuals over 65 years old are diagnosed with Alzheimer's disease, and its prevalence only increases with age. As people get older, their ability to multitask goes down, as well as having lower situational processing speeds. Alzheimer's and other similar neurodegenerative disease related deaths had an increase of over 200% between the years 1999 and 2017, and as such, the need for a method to combat the changes in patient brain age earlier via medicine and treatment has also increased. Utilizing recordings of patient EEG signals obtained from the Temple University Abnormal EEG Corpus, the use of Deep Learning Recurrent Neural Networks for the classification and prediction of patient brain age versus chronological age is explored. The EEG data from the Temple University Abnormal EEG Corpus contains data labeled as Normal, data that does not contain large unordinary differences, and Abnormal from a variety of patients with ages ranging from 2 - 88 years of age. The EEG data was obtained using the 10-20 System of EEG node placement, with 87% of the data being recorded at a frequency of 250 Hz. As EEG recordings are categorized as a time series, the use of Recurrent Neural Networks is a focus. The preprocessing of the utilized EEG data is detailed both as requirements for the training of deep learning models, but also as it has a history of requiring much more evaluation than other types of data due to the large number of differences in brain signals per individual, as well as their complex and non-linear nature. Utilizing Fast Fourier Transform as well as Discrete Wavelet. Transform, one of the five major brain wave frequency bands are utilized to train the deep learning models, as certain frequency bands are more prevalent while a patient is performing certain tasks. Three recurrent neural network deep learning models are evaluated: a Long-Short Term Memory, a Gated Recurrent Unit, a Bidirectional Long Short-Term Memory, and a Bidirectional Gated Recurrent Unit. The most promising Recurrent Neural Network model achieved a test accuracy of 90% for a proposed 6 age classes, with the least promising model utilizing Fast Fourier Transform preprocessed data achieving a test accuracy of 17%. This low accuracy could be due in part to the major differences in individual EEG recordings, causing the Recurrent Neural Network model to achieve less than desirable results after testing. For regression analysis of patient EEG data for age prediction, the most promising model achieves a mean absolute error value of 7 years, which is over a benchmark utilizing the same dataset with a Convolutional Neural Network. The adoption of the Recurrent Neural Network to solve problems, even those involving time series data, involves much more than may be believed and other machine learning approaches may be more applicable in some instances.