loading page

Leveraging SVD Entropy and Explainable Machine Learning for Alzheimer's and Frontotemporal Dementia Detection using EEG
  • Utkarsh Lal ,
  • Arjun Vinayak Chikkankod ,
  • Luca Longo
Utkarsh Lal
Manipal Academy of Higher Education, Manipal Academy of Higher Education

Corresponding Author:[email protected]

Author Profile
Arjun Vinayak Chikkankod
Author Profile
Luca Longo
Author Profile

Abstract

Alzheimer’s Disease (AD) and Frontotemporal Dementia (FTD) are debilitating neurodegenerative disorders that often remain undiagnosed until their later stages, where symptoms become nearly irreversible. There is scope to strengthen the ongoing research regarding the optimal feature-extraction methods for capturing relevant AD/FTD biomarkers from Electroencephalography (EEG) data. This study aims to fill this gap by conducting a comparative analysis of various feature-extraction measures: Higuchi’s Fractal Dimension, Singular Value Decomposition Entropy, Zero Crossing Rate, Detrended Fluctuation Analysis, and Hjorth parameters. Our results highlight SVD Entropy as the superior measure, a finding not deeply explored in previous research with respect to AD/FTD. Furthermore, we aim to design an optimal machine learning pipeline incorporating sliding window segmentation, feature-extraction, and a supervised learning algorithm for discriminating AD/FTD patients from healthy controls. To amplify the interpretability of this study, we harnessed Explainable AI to generate feature-importance topographic brain plots, illuminating regions with potentially higher degeneration. The results highlight a model incorporating SVD Entropy, KNN Classifier, and 90% overlapped sliding windows as the best performer, yielding a mean F1-score of 93% with an Accuracy of 91% for AD and HC discrimination, an average F1-score of 92.5% with an Accuracy of 93% for FTD and HC differentiation, and F1-score of 91.5% with an Accuracy of 91% for FTD and AD distinction. This research presents a novel architectural pipeline for detecting AD and FTD from EEG data, addressing the crucial need for accurate differentiation between AD and FTD, especially in the early stages where misdiagnosis is common, while also providing an in-depth comparative analysis of various feature-extraction measures.