EEG Complexity Measures for Alzheimer's and Frontotemporal Dementia Classification Using Explainable Machine...

Area of Science:

• Neuroscience
• Medical Informatics
• Biomedical Engineering

Background:

• Alzheimer's disease (AD) and frontotemporal dementia (FTD) are progressive neurodegenerative disorders.
• Accurate and early differentiation of AD and FTD from healthy controls (CN) is crucial for effective management.
• Current diagnostic methods can be invasive or lack accessibility.

Purpose of the Study:

• To classify patients with AD and FTD from CN using non-linear electroencephalography (EEG) features.
• To evaluate the performance of various machine learning models in these classification tasks.
• To identify key EEG features and brain regions important for disease differentiation.

Main Methods:

• Utilized a dataset of 88 subjects (36 AD, 29 CN, 23 FTD).
• Extracted non-linear EEG features and applied machine learning models (XGBoost, MLP, KNN, SVM).
• Employed explainable AI (XAI) with SHAP analysis to interpret model decisions.

Main Results:

• Achieved 100% accuracy in CN vs. AD classification and high Area Under the Curve (AUC) values (0.99) for most classifiers.
• Identified the occipital electrode O2 as crucial for AD vs. CN differentiation.
• Frontal and temporal electrode features were important for FTD vs. AD and CN vs. FTD classifications.
• Multi-class classification (AD, FTD, CN) accuracy was 82%.

Conclusions:

• Non-linear EEG features combined with machine learning offer a promising approach for diagnosing AD and FTD.
• The methodology demonstrates potential for a non-invasive, cost-effective tool for early disease detection and differentiation.
• Findings support the clinical relevance of EEG in neurodegenerative disease diagnosis and monitoring.