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

This study aimed to classify patients with Alzheimer's disease (AD) and frontotemporal dementia (FTD) from healthy controls (CN) using non-linear EEG features. We explored various classification tasks, including binary classifications (CN vs. AD, CN vs. FTD, FTD vs. AD) and multi-class classifications, employing machine learning models such as XGBoost, multi-layer perceptron, k-nearest neighbors, and support vector machines. To understand the model's decision-making process, we employed explainable AI (XAI) using SHAP (SHapley Additive exPlanations) analysis. An EEG dataset of 88 subjects: 36 with AD, 29 controls, and 23 with FTD, was used. The occipital electrode O2 played a crucial role in differentiating AD from controls. In both FTD vs. AD and CN vs. FTD classifications, features from the frontal and temporal electrodes exhibited the highest importance. The results showed that XGB and MLP perform best across all classification tasks, with 100% accuracy achieved in CN vs. AD classification and area under the curve values of 0.99 for most classifiers. Distinguishing unhealthy patients (AD and FTD) from healthy controls yielded lower performance, potentially due to the differential EEG signal alterations in these conditions. The multi-class classification of AD, FTD, and controls achieved accuracy of 82%, lower than the binary classification tasks. The study proposed a novel methodology combining non-linear EEG features and machine learning models, offering the potential for improved disease detection.Clinical relevance- This study offers the potential to provide a non-invasive, efficient method for early detection and differentiation of AD and FTD from healthy controls. The ability to classify these neurodegenerative diseases using EEG, a widely accessible and cost-effective tool, could significantly aid in the timely diagnosis and monitoring of disease progression. This could lead to more personalized treatment plans and improved patient care.