Machine Learning Prediction of Memory Capacity from Brain Signals

Sammanfattning

Assessing memory capacity accurately and objectively remains challenging due to the absence of easily measurable biological markers. Such an assessment could greatly benefit neuroscience, particularly in diagnosing conditions like Alzheimer’s disease. This thesis aims to identify brain markers of memory by utilizing artificial intelligence. To do so, 20 healthy subjects underwent EEG scans while listening to familiar and unfamiliar songs interspersed with 2-second silent intervals. The EEG data collected during these silences were used to reflect the subjects' ability to remember the songs. In this thesis, we aim to develop an artificial intelligence model to determine whether a song was known or not, based on EEG signals during the silences to potentially identify memory markers. To achieve this, three conventional models (SVM, logistic regression, and random forest), three Riemannian models (MDM, FgMDM, and Tangent Space Mapping), and a transfer learning model were implemented to classify the silent periods. Results show that the Tangent Space Mapping model achieved the highest classification performance with 80% accuracy. The best-performing conventional model, logistic regression, achieved 71% accuracy. Further analysis revealed that the right and left auditory cortexes were the most important and discriminating brain areas for this classification. These findings indicate that these brain areas contain information relevant to familiarity and, by extension, memory capacity.