Exploring the potential of explainable deep learning for EEG-based cognitive decline prediction.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Early detection of Alzheimer's disease (AD) is crucial for effective treatment and preventing neuronal damage.
• Current diagnostic methods for AD are often invasive and costly.
• There is a need for cost-effective and non-invasive methods for early cognitive decline detection.

Purpose of the Study:

• To develop and evaluate novel deep learning (DL) algorithms for the early, non-invasive detection of cognitive decline.
• To create a cost-effective diagnostic tool for Mild Cognitive Impairment (MCI) and preclinical AD.
• To identify potential EEG biomarkers associated with early cognitive deficits.

Main Methods:

• Utilized a publicly available dataset of resting-state electroencephalographic (EEG) data from healthy controls and MCI patients.
• Developed two novel DL algorithms incorporating self-attention mechanisms.
• Evaluated model performance in predicting MCI and cognitive decline, comparing against a traditional Convolutional Neural Network (CNN).

Main Results:

• The proposed DL algorithms outperformed a traditional CNN for MCI prediction, with accuracy improvements of 8.5% and 10%.
• Ablation studies confirmed the attention layer's importance, boosting accuracy by 8.5%.
• Analysis revealed beta band frequencies (13-30 Hz) as key indicators for distinguishing MCI; preclinical decline prediction achieved 56.08% accuracy using transfer learning.

Conclusions:

• The developed DL models achieved state-of-the-art results for MCI classification and showed progress in predicting preclinical cognitive decline.
• This study pioneers the use of DL attention models for classifying healthy subjects based on cognitive scores, identifying subtle brain changes.
• The findings offer new avenues for discovering early AD biomarkers and validating AI in healthcare through interpretable attention mechanisms.