Improving the multi-class classification of Alzheimer's disease with machine learning-based techniques: An EEG-fNIRS hybridization study

Abstract

BACKGROUND: As an incurable neurodegenerative disorder, Alzheimer's Disease (AD) is being projected to become one of the most expensive diseases for understanding the pathophysiological mechanism. Exploring alterations in the AD brain network is therefore of central importance for early and accurate diagnosis of cognitive deficits levels, yet diagnostic uncertainties exist merely confirmed by postmortem brain examination. The hybrid approach of two modalities, Electroencephalography (EEG) and functional Near-Infrared Spectroscopy (fNIRS) reveals an ideal technique-derived diagnosis due to the inexpensive and noninvasive experiments coupled with the reliability and versatility that make it highly desirable for AD multi-class classification tasks. METHODS: In this study, the hybrid EEG-fNIRS was employed in constructing the classification machine learning (ML)-based models to categorize four subject groups, including healthy controls (HC) and three AD patient classes. First, a concurrent EEG-fNIRS setup was utilized to record data from 41 subjects during the Oddball - a cognitive ability test and 1-back - a memory ability test. Second, while fNIRS features were computed and extracted, the event-related spectral perturbation (ERSP) features measuring the average dynamic changes in amplitude of three main EEG frequency bands were extracted. Third, a conventional neural network and a hybrid deep Convolutional Neural Network (CNN) - Long Short-Term Memory Network (LSTM) was built for the multi-class classification using the fNIRS and EEG features, respectively. To boost the classification accuracy, the output of different models was combined by majority voting methods. RESULTS: The hybrid EEG-fNIRS feature set was able to achieve a higher accuracy (71.01% ± 0.004) by combing their complementary properties, compared to using EEG (66.35% ± 0.016) or fNIRS (67.15% ± 0.017) alone. In addition, fNIRS and ERSP feature patterns displayed a significant difference between subject groups during the Oddball and 1-back experimental tasks. Thus, the feature-level ensemble shows a great ability to understand neurophysiological AD status and its related symptoms and develop a syndrome-specific diagnostic tool for our future works. CONCLUSIONS: These findings demonstrate the potential of the hybrid EEG-fNIRS systems to enhance the AD diagnosis and assessment process and the capability of ML techniques to facilitate further AD classification studies.