QEEG‐based discernment of dementia pathologies through machine learning: Lewy body and Alzheimer’s disease

Abstract

BackgroundAlzheimer’s disease is the most common cause of dementia, which destroys nerve cells in the brain through the formation and accumulation of amyloid plaques. However, various other pathologies of dementia also exist, such as Lewy body pathology where the destruction is caused by the accumulation of Lewy body peptides. Differing dementia pathologies carry distinguishing symptoms and treatment methods; hence we cannot disregard the importance of correct identification of the pathology.Currently, the correct pathology can be identified through positron emission tomography (PET), which is expensive and results in exposure to harmful ionizing radiation. Therefore, the present study aims to overcome the disadvantages of current screening methods and distinguish the quantitative electroencephalography (QEEG) of two dementia pathologies through machine learning.MethodEEG data employed in the present study were recorded at the 19 electrode locations defined by the international 10‐20 system, in eyes‐closed resting‐state condition. The acquired data were then re‐referenced to a common average reference. Hereafter, bandpass filtering at 0.1‐45.5Hz, bad epoch rejection and independent component analysis (ICA) was performed to remove noise components from the data. Standardized low resolution brain electromagnetic tomography (sLORETA) was further used for mathematical estimation of source cortical activity, yielding data for 68 cortical regions based on the Desikan‐Killiany atlas.Gamma wave (30‐45Hz) features were excluded from the data due to is vulnerability to electromyography. Further feature reduction was also performed, through the designated criteria on p‐value threshold (0.05) and feature importance threshold determined by shapely values. The final dataset (N = 104; 30 ADD, 74 DLB) was established, where 20% of the data (N =21; 6 ADD, 15 DLB) were randomly selected and excluded as test data for model verification.ResultThe final machine learning model trained and tested through the dataset yielded test accuracy at 85.7% with ADD sensitivity at 83.3%, DLB sensitivity at 86.7%. The 5 fold cross validation accuracy was at 82.1%.ConclusionThe model developed in the present study yielded a promising classification performance, through EEG data which is cheap and harmless to record. Such QEEG‐based classification models may sufficiently replace the PET in future, resolving current disadvantages.