Alzheimer’s disease dementia and frontotemporal dementia are two of the most common types of neurodegenerative diseases worldwide. Patients suffering from dementia experience a progressive decline in cognition, memory and physical movement; the brain’s health continues to deteriorate and could eventually cause the patient’s death. Global efforts are being put into researching and understanding these diseases, but the problem at hand remains very complex, and there is a need for more robust and accurate diagnosis and prognosis methods. With the introduction of the non-invasive neuroimaging technique electroencephalogram (EEG), medical professionals were able to obtain signals from the neurons which allowed the analysis of patterns and trends in individuals’ brains. State-of-the-art machine learning techniques are being utilised in many fields, and self-supervised learning (SSL) could be very beneficial in the case of learning from neurological data. The technique particularly beneficial where access to large, annotated EEG datasets is quite expensive and limited, and obtaining informative features from the data is very challenging. This project explores the use of the SSL algorithm EEG2rep with neurodegenerative disease data in order to obtain more complex and representative EEG features for classifying Alzheimer’s diseases and Frontotemporal Dementia from a healthy control group. Although the model performed slightly worse than other state-of-the-art methods such as Detach-ROCKET when evaluated on individual instances (short windows of brain activity), it achieved a higher accuracy at subject-level classification through majority vote, which is the most clinically relevant metric.
Shafeek Zakko (Wed,) studied this question.