Abstract
Alzheimer's disease (AD) is a chronic disability that occurs due to the loss of neurons. The traditional methods to detect AD involve questionnaires and expensive neuro-imaging tests, which are time-consuming, subjective, and inconvenient to the target population. To overcome these limitations, Electroencephalogram (EEG) based methods have been developed to classify AD patients from normal controlled (NC) and mild cognitive impairment (MCI) subjects. Most of the EEG-based methods involved entropy-based feature extraction and discrete wavelet transform. However, the existing AD classification methods failed to provide promising classification accuracy. Here, we proposed a wavelet-machine learning (ML) framework to detect AD using a newly designed biorthogonal filter bank by optimization of frequency and time localization of triplet halfband filter banks (OTFL-THFB). The OTFL-THFB decomposes EEG signals into various EEG sub- bands. Hjorth Parameters (HP) and Higuchi's Fractal Dimension (HFD) have been investigated to extract features from each EEG subband. Subsequently, ML models are trained and tested using different features such as OTFL-THFB with HFD, OTFL-THFB with HP, and OTFL-THFB with HFD and HP used for detecting AD with 10-fold cross-validation. This method was applied to two publicly available datasets. Our model achieved an accuracy of 98.91% for AD versus NC and 98.65% for AD versus MCI versus NC using the least square support vector machine. Results indicate that this framework surpassed existing state-of-the-art techniques for classifying AD from NC.