EEG-Based Systematic Explainable Alzheimer’s Disease and Mild Cognitive Impairment Identification Using Novel Rational Dyadic Biorthogonal Wavelet Filter Banks

• Published in: Circuits, systems, and signal processing Vol. 43; no. 3; pp. 1792 - 1822
• Main Authors: Puri, Digambar V., Nalbalwar, Sanjay L., Ingle, Pallavi P.
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.03.2024; Springer Nature B.V
• Subjects: Alzheimer's disease; Circuits and Systems; Cognitive ability; Complexity; Electrical Engineering; Electroencephalography; Electronics and Microelectronics; Engineering; Filter banks; Impairment; Instrumentation; Low frequencies; Signal,Image and Speech Processing; Support vector machines; Synchronism; Discrete wavelet transform; Halfband filter; Support vector machine; Alzheimer’s disease; Electroencephalogram; Filter banks
• ISSN: 0278-081X; 1531-5878
• DOI: 10.1007/s00034-023-02540-x

Alzheimer’s disease (AD) is a frequently encountered chronic disorder. AD patients suffer from various cognitive dysfunctions. The traditional methods fail to identify AD in the early stage. The presence of AD results in significant changes in electroencephalogram (EEG) signals, including a slowing effect and less synchronization. The important information in EEG is available in low-frequency bands. These bands can be obtained using various wavelet filter banks. This work proposes new, less complex Rational Dyadic Biorthogonal Wavelet Filter Banks (RDBWFBs) with maximum vanishing moments for the decomposition of EEG signals from normal controlled (NC) subjects, mild cognitive impairment (MCI), and AD patients into desired EEG bands. Novel design approaches have been introduced to decrease the complexity associated with current irrational biorthogonal wavelet filter banks. Three different features were calculated from each EEG subband. The importance of these features was determined through the utilization of Kruskal–Walli’s test. The present model achieved an AD detection accuracy of 98.85 % for NC vs. AD using RDBWFB-5 and 96.30 % for NC vs. MCI vs. AD classifications using the RDBWFBs-4 with a support vector machine, respectively. New RDBWFBs are more effective and less complex than existing wavelet filter banks.