Kaouther Selmi

University of Monastir, Tunisia

Title: Integrative Analysis of EEG-Based Functional Connectivity Alterations in Early Alzheimer’s Disease Using Custom Bessel-Activation Neural Networks

---

Biography

Kaouther Selmi is an Assistant Professor of Electronics and Computational Neuroscience at the University of Monastir, Tunisia. Her research lies at the intersection of electronics, nonlinear dynamics, and neural modeling, with a particular focus on simulating brain behavior through mathematical and electronic circuit approaches. She has authored and co-authored several scientific papers on topics such as chaotic neural systems, neuron–astrocyte–synapse interactions, and the application of Bessel functions in modeling multidendritic neuron behavior. Her current work explores neuromorphic and bio-inspired architectures, aiming to bridge artificial intelligence and biological signal analysis for understanding and classifying complex brain dynamics. She also mentors graduate students in the areas of machine learning, biomedical signal processing, and the development of intelligent embedded systems.

Abstract

Background:

Alzheimer’s disease (AD) is characterised by progressive cognitive decline and is associate with aberrant neural network dynamics and connectivity. Early detection remains a major clinical challenge. In this study, we propose an innovative signal-processing and machine-learning pipeline to identify functional connectivity alterations in early AD patients using electroencephalography (EEG) data. Methods: EEG recordings were collected from n= 30 early-stage AD patients and n= 30 age-matched healthy controls during a resting-state protocol. Signals were pre-processed (band-pass filtered 1-45 Hz, artefact removal) and functional connectivity was estimated via phase-locking value (PLV) and coherence metrics across canonical frequency bands. A custom multilayer neural network was designed in MATLAB embedding a novel activation function

????(????)=(?????1)2exp (?1/2(?????1) 2)

to better capture non-linear interactions arising from dendritic-type coupling. Network input features included PLV/coherence matrices, Bessel-function-based radial decompositions of potential distributions (inspired by multi-dendritic neuron models), and classical spectral indices (e.g., theta/gamma ratio). Model training employed stratified 10-fold cross-validation; performance metrics included accuracy, sensitivity, specificity, AUC. Results: The proposed approach achieved mean accuracy of 88 % (sensitivity = 85 %, specificity = 91 %, AUC = 0.93) in discriminating early-stage AD from controls. Feature-importance analysis revealed dominant contributions from gamma-band PLV reductions in fronto-parietal networks and elevated theta/gamma ratios. The neural network’s custom activation improved convergence speed by ~25 % and reduced overfitting compared to ReLU and tanh baselines. Conclusions: These findings demonstrate the feasibility of combining advanced EEG connectivity metrics with bespoke neural network architectures for early AD detection. The use of the custom activation function appears particularly promising for capturing subtle non-linear neural dynamics associated with neurodegeneration. Further work will extend this pipeline to longitudinal data and multi-modal integration (e.g., ECG, MRI).

Keywords:

Alzheimer’s disease; EEG functional connectivity; neural network; custom activation function; Bessel functions; early diagnosis.