A Multimodal Multi-Stage Deep Learning Model for the Diagnosis of Alzheimer's Disease Using EEG Measurements

Background/Objectives: Alzheimer's disease (AD) is a progressively debilitating neurodegenerative disorder characterized by the accumulation of abnormal proteins, such as amyloid-beta plaques and tau tangles, leading to disruptions in memory storage and neuronal degeneration. Despite its portability, non-invasiveness, and cost-effectiveness, electroencephalography (EEG) as a diagnostic tool for AD faces challenges due to its susceptibility to noise and the complexity involved in the analysis. Methods: This study introduces a novel methodology employing three distinct stages for data-driven AD diagnosis: signal pre-processing, frame-level classification, and subject-level classification. At the frame level, convolutional neural networks (CNNs) are employed to extract features from spectrograms, scalograms, and Hilbert spectra. These features undergo fusion and are then fed into another CNN for feature selection and subsequent frame-level classification. After each frame for a subject is classified, a procedure is devised to determine if the subject has AD or not. Results: The proposed model demonstrates commendable performance, achieving over 80% accuracy, 82.5% sensitivity, and 81.3% specificity in distinguishing AD patients from healthy individuals at the subject level. Conclusions: This performance enables early and accurate diagnosis with significant clinical implications, offering substantial benefits over the existing methods through reduced misdiagnosis rates and improved patient outcomes, potentially revolutionizing AD screening and diagnostic practices. However, the model's efficacy diminishes when presented with data from frontotemporal dementia (FTD) patients, emphasizing the need for further model refinement to address the intricate nuances associated with the simultaneous detection of various neurodegenerative disorders alongside AD.