A deep-SVM hybrid framework with enhanced EEG feature engineering and SHAP-based explainability for Alzheimer's...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Accurate differentiation of Alzheimer's Disease (AD), Fronto Temporal Dementia (FTD), and Cognitively Normal (CN) individuals using electroencephalography (EEG) is clinically significant yet challenging.
• Existing EEG classification methods often lack interpretability or robustness across different datasets and conditions.

Purpose of the Study:

• To develop and validate a novel, interpretable, and robust multi-class classification framework for distinguishing AD, FTD, and CN from EEG data.
• To fuse spectral/connectivity biomarkers with deep learning temporal embeddings for enhanced classification accuracy.

Main Methods:

• A hybrid classification approach combining interpretable EEG biomarkers (band power, spectral entropy, α-coherence) with temporal features from a lightweight 1D Convolutional Neural Network (1D-CNN).
• Feature fusion followed by Principal Component Analysis (PCA) for dimensionality reduction and Support Vector Machine (SVM) for classification.
• Rigorous methodology to prevent data leakage, including fitting data-dependent steps (SMOTE, z-scoring, PCA) strictly on training folds and employing inner-loop grid sweeps for hyperparameter optimization.

Main Results:

• Achieved high classification accuracy (94.5%), macro-F1 score, and AUC on the OpenNeuro ds004504 dataset (AD/FTD/CN).
• Demonstrated robustness through cross-condition testing on eyes-open recordings (ds006036) and zero-shot transfer to an independent dataset.
• SHapley Additive exPlanations (SHAP) analysis provided clinically interpretable, subject-level attributions consistent with known electrophysiological changes in neurodegenerative diseases.

Conclusions:

• The proposed transparent hybrid feature design offers an accurate, leakage-safe, and explainable method for EEG-based differentiation of AD, FTD, and CN.
• The findings support the clinical utility of fusing interpretable biomarkers with deep learning for neurodegenerative disease classification.
• Preliminary external validation suggests the generalizability of the developed classification framework.