Artificial Intelligence Model‐based Alzheimer’s Disease Dementia Classification Using Electroencephalogram‐acquired Brain Functional Network

Background Alzheimer’s disease dementia (ADD) is a predominant neurodegenerative disorder that results severe abnormalities in cognitive functions. However, most diagnoses rely on assessments made by individual clinical experts. Hence, we aim to establish a novel biomarker which can provide both accuracy and objectivity. Electroencephalogram (EEG) data were utilized for the training of artificial intelligence (AI) models, making use of its cost‐advantage and simple utility. Method 19‐channel EEG data were recorded at Chung‐Ang University hospital in accordance with international 10‐20 system at resting state with eyes being closed. We first converted 19 channels into 68 regions of interest (ROIs) using standardized low resolution brain electromagnetic tomography (sLORETA). Henceforth, imaginary part of coherence (iCoh) between each ROI were calculated, reflecting functional connectivity between the ROIs. The iCoh values were then utilized in the calculation of the brain’s functional brain network. The subjects were divided into non‐ADD and ADD group. The resulting training dataset consisted of 118 non‐ADD and 110 ADD subjects, where the number of non‐ADD training data were adjusted in order to account for data imbalance. The test dataset included 476 non‐ADD and 27 ADD subjects. The classification model was trained using 68 ROI source powers and functional brain network features. Only statistically significant features were selected for feature reduction purposes. Through primary classification via light gradient boosting method (LGBM), top 30 features with high importance values were selected and were employed in the training of our final LGBM classification model. Result Through appropriate hyperparameter tuning, the LGBM classification model resulted in 87.67% accuracy, 88.89% sensitivity and 87.61% specificity. From the receiver operating characteristic (ROC) curve, 0.9506 area under ROC curve (AUC) was achieved. Conclusion The LGBM model resulted in promising classification performance, solely through EEG features. In addition, functional brain network features had higher importance in comparison with source power features, indicating high potential of EEG brain network features in distinguishment of neurodegenerative diseases.