Identifying Alzheimer’s Disease Dementia Through Ensemble Learning of Channel and Source Level Electroencephalogram Features

Introduction: Alzheimer's disease dementia (ADD) has now become a crucial concern for modern society as a result of increased life expectancy. However, it is often difficult for a majority of the population to afford expensive medical imaging tests for accurate diagnosis. As a solution, quantitative analysis of electroencephalography (EEG) that aids in a sufficient description of brain activities can be employed as a cost-effective, safe and objective diagnostic tool. In the presented research, we employed diverse QEEG features at both channel- and source-level to enhance the robustness of our previously established artificial intelligence (AI) model that distinguishes non-ADD (NADD) data from ADD data. Method 594 NADD and 137 ADD subjects’ EEG data were employed for the presented research. artifact-free data were obtained through the application of independent component analysis (ICA) and bad epoch rejection. Absolute and relative power spectra at 19 channels were first computed, followed by the estimation of source-level power spectra through standardized low-resolution brain electromagnetic tomography (s-LORETA). Through further feature engineering, functional brain networks were also obtained. The established channel-level features were transformed into images that spatially allocate absolute and relative spectral powers, which were utilized for the training of deep neural network structures. Moreover, source-level spectral powers and functional brain networks were adopted for the training of a tree-based machine learning algorithm. Prediction probabilities of the established classification models were ensembled through the voting method and returned the final classification result. Results The best classification accuracies of the absolute and relative channel-level spectral power image-based deep neural network models were 85.3% and 86.5% respectively. The tree-based model that has been trained with source-level features resulted in an accuracy of 87.7%. The accuracy of the ensemble model was 88.5%, which demonstrates the compensatory interaction among the models. Conclusions The promising classification results indicate the potential behind EEG-AI models for the analysis of neurodegenerative disorders. Through continuous analysis of several independent QEEG features of varying aspects, we may soon be able to more aptly diagnose several neurological disorders.