Machine learning classification of mild Alzheimer's disease using EEG during emotional processing

Background Mild Alzheimer's disease (AD) is associated with alterations in brain activity, which can be detected using electroencephalography (EEG). Investigating these changes during emotional processing may help identify neurophysiological patterns that differentiate patients with mild AD from healthy controls (HCs). Objective To investigate whether EEG responses elicited during emotional processing provide biomarkers capable of discriminating patients with AD from HCs. Additionally, to identify the emotional contexts, brain regions, frequency bands, and machine learning models that maximize this discriminative capacity. Methods A sample of 39 AD patients and 54 HCs watched brief movie clips designed to elicit tenderness, amusement, anger, fear, and sadness, along with an Alzheimer's-related clip, together with neutral clips used as control, baseline, and recovery conditions, while cortical activity was recorded using EEG. The signals were then classified across emotional contexts using LASSO, Random Forest, SVM-RBF, XGBoost, and CatBoost ML models. Results EEG signals allowed for discrimination between AD patients and HCs across different emotional contexts, including individual emotions, grouped by valence or arousal, and the Alzheimer's-related stimulus. LASSO achieved the best performance for positive and neutral conditions in the parietal and posterior gamma bands, whereas CatBoost performed best for high-arousal negative emotions such as anger and fear, particularly in frontal gamma and theta bands, respectively. Conclusions Patients with mild AD show EEG signal patterns that differ from those of HCs across different emotional contexts. These findings highlight the potential of EEG recorded during emotional processing to support the development of objective biomarkers for mild AD.