Nonlinear Dynamical Analysis of EEG Signals for Differential Diagnosis of Alzheimer's Disease, Frontotemporal Dementia, and Healthy Controls: A Comprehensive Multiscale Entropy, Chaos, and Fractal Framework

Abstract

Alzheimer's disease (AD) and frontotemporal dementia (FTD) are two major neurodegenerative disorders with overlapping clinical presentations, posing significant challenges for differential diagnosis. This study presents a comprehensive computational framework for distinguishing AD, FTD, and healthy controls (CN) using nonlinear dynamical analysis of resting-state EEG signals from 87 participants. We employed phase-space reconstruction with optimized parameters (embedding dimension m = 6, time delay τ determined via mutual information) to extract a comprehensive set of nonlinear features: largest Lyapunov exponent (LLE) using both Rosenstein and Wolf algorithms , Higuchi fractal dimension (HFD), Katz fractal dimension, multifractal detrended fluctuation analysis (MF-DFA), multiscale entropy (MSE) across 5 scales, sample entropy (SampleEn), spectral entropy (SpecEn), permutation entropy (PermEn) in frontal theta and alpha bands, and bifurcation analysis. Feature fusion was performed using the mRMR algorithm to select the most discriminative biomarkers, followed by manifold learning (t-SNE and Isomap) for nonlinear dimensionality reduction. The mRMR selection identified four optimal biomarkers: TAR Frontal (ANOVA p=0.0003), Alpha ParietoOcc (ANOVA p=0.0006), HiguchiFD (ANOVA p=0.0080), and Tau Delay (ANOVA p=0.0359). Classification using KNN achieved 67.91% accuracy under 5-fold cross-validation, while subject-level generalization using SVM-RBF achieved 64.07% under leave-one-subject-out cross-validation. Channel selection analysis identified PZ (parietal midline) and T7 (left temporal) as the most informative channels. Manifold learning revealed clear structural separation between diagnostic groups. These findings demonstrate that nonlinear EEG features provide robust, interpretable biomarkers for dementia differential diagnosis, with fractal dimension, multifractal spectrum width, and entropy measures capturing complementary aspects of neural dynamics degradation