Abstract
Background/Objectives: Alzheimer's disease (AD) is a progressive neurodegenerative disorder advancing through subjective cognitive decline (SCD), mild cognitive impairment (MCI), and dementia, making early diagnosis crucial. Electroencephalography (EEG) is a non-invasive, cost-effective alternative to advanced neuroimaging for detecting early neural changes. While most studies focus on resting-state EEG or handcrafted features with traditional machine learning, deep learning (DL) offers a promising tool for automated EEG analysis. This study classified the AD spectrum (SCD, MCI, AD) using EEG recorded during resting-state and task-based conditions. Specifically, EEG was recorded during a simple yes/no question-answering task, mimicking everyday cognitive activities, and was explored. We hypothesized that brain activity during tasks involving listening, comprehension, and response execution provides diagnostic insights. Methods: We collected 1 min of resting-state EEG and approximately 3 min of task-based EEG from 20, 28, and 10 participants with SCD, MCI, and AD, respectively. Task data included response accuracy and reaction time. After minimal preprocessing, two DL models, attention long short-term memory and EEGConformer, were used for binary (e.g., SCD vs. MCI) and three-class (SCD, MCI, AD) classification. Results: Task-based EEG outperformed resting-state EEG, with a 5-15% improvement in accuracy. The area under the curve (AUC) results consistently demonstrated superior classification performance for task-based EEG compared to resting-state EEG across all group distinctions. No significant performance difference was observed between the two DL models. Conclusions: We proposed a cognitive task-based approach for early AD spectrum diagnosis via EEG, offering greater accuracy by leveraging advanced DL models.