A Unique EEG-Hyfusion Fully Automated Stacked Model for Classification of Alzheimer's Disease and Fronto-Temporal Dementia

Summary

This article presents an automated feature extraction pipeline that incorporates Nyquist-shift enhancement and machine-learning ensemble models to distinguish between Alzheimer's disease, frontotemporal dementia, and Healthy classes without requiring human intervention.

Abstract

Alzheimer's disease (AD) and fronto-temporal dementia (FTD) are common neurodegenerative disorders that impair memory, cognitive function, and executive processing. The purpose of this study is to develop a fully automated machine learning pipeline for predicting Alzheimer's disease at an early stage without requiring any clinical intervention. This methodology proposes a quantitative analysis of subtle neuro-activity shifts. The goal is to develop a reliable, fully automated system that utilizes EEG data to classify patients into AD, FTD, and Healthy Control groups, eliminating the need for human intervention or clinical assessments. A major innovation of the system lies in its signal processing approach and automated feature pipeline. Specifically, the strategic modification of the Nyquist frequency is used to enhance EEG signal resolution, in combination with a hybrid fusion layer that integrates multi-domain EEG features and demographic data. Subsequently, a two-way ANOVA-based feature selection refines this hybrid set. This enhancement facilitates more effective feature extraction, contributing to higher classification accuracy. In the proposed method, frequencies are epoched to enrich the training dataset. And thereby the standard random forest model gives 99.72% training accuracy. To ensure the robustness and generalizability of the method, a hybrid fusion model is proposed.

Introduction

The common symptoms of Alzheimer's disease are memory loss, cognitive impairment, sleep insomnia, confusion, disorientation, personality changes, etc. Therefore, early prediction can ensure patients do not degrade to the worst levels, where they need help 24/7. There are various ways to diagnose this disease, such as MRI¹, CT, PET², and EEG³. In particular, the EEG signals in combination with MMSE have proven to be a non-invasive method for the prediction of Alzheimer's disease. The EEG signal frequencies are varied as alpha, beta, theta, gamma, and delta. An alpha wave is detected when a ....

Protocol

The proposed method segments EEG signals strategically into the frequency domain to improve spectral resolution. Subsequently, the captured EEG signals are epoched into 10 segments to enhance temporal resolution and consistency in feature computation. The proposed method implements an RF, XGBoost, and SVM-based classifier on derived EEG features, thereby contributing to the development of robust, scalable methods for EEG analysis. The RF model with certain hyperparameter changes has proved to be the best-performing among other standard models. Therefore, a hybrid, fused, stacked metalearner model is proposed that combines RF and XG-Boost to further improve prediction ....

Results

To justify the strength of the epoch method, the RF model's performance is compared without epoching the dataset. In the non-epoched condition, features were extracted from the entire continuous recording and given to the RF model. The non-epoching signal yielded 53% accuracy. This demonstrates that epoching provides a more stable informative method for model creation. Therefore, ANOVA-based features with p < 0.05 are considered to be significant. For instance, alpha_rms exhibited a significant Group effect (p = 9.007.......

Discussion

In this proposed study, the motivation for using the Hy-Fusion stacked architecture is to combine the complementary strengths of tree-based learners, such as RF, and to reduce prediction variance through a logistic-regression meta-learner. The stacking strategy is commonly adopted when base learners have different inductive biases and error patterns. The meta-learner learns to correct systematic errors of base classifiers and to produce more calibrated ensemble outputs.

The process of creating.......

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Bandpass filtering tools	MNE-Python	Built-in filters	Used for preprocessing
Computer workstation	—	Windows/Linux	Analysis computation
EEG Acquisition System	Provided by dataset creators	—	Not used by authors (dataset provided)
Epoching functions	MNE-Python	Built-in functions	Used for segmentation
GitHub or Google Drive	—	—	Storage for code/data
Google Colab	Google	Online	Cloud computing environment
MATLAB (used for signal checks)	MathWorks	R202x	Optional
Mendeley EEG Dataset	Mendeley Data		External validation dataset
MNE-Python	https://mne.tools	v1.x	EEG preprocessing/analysis
NumPy	NumPy developers	Latest	Array computation
OpenNeuro EEG Dataset	OpenNeuro	ds004504 (v1.0.8)	Primary dataset for AD/FTD/HC EEG
Python	Python Foundation	v3.10+	Programming environment
scikit-learn	sklearn developers	v1.x	Machine learning models
SciPy	SciPy developers	Latest	Signal processing