Core Concepts
This study explores the potential of utilizing sleep-related electroencephalography (EEG) signals acquired through polysomnography (PSG) for the early detection of Alzheimer's disease (AD), focusing on semi-supervised deep learning techniques to overcome the challenge of limited labeled data.
Abstract
This study investigates the use of sleep EEG signals for the early detection of Alzheimer's disease (AD) using deep learning techniques. The key highlights and insights are:
The study leverages four fully labeled databases, including data from AD patients and healthy controls, to analyze the potential of PSG signals for AD detection.
The preprocessing steps involve standardizing hypnograms, removing artifacts, filtering, normalizing, and segmenting the EEG signals into uniform 10-second fragments for the different sleep stages (N1, N2, N3, and REM).
The study employs a range of deep learning models, including semi-supervised (SMATE and TapNet), supervised (XCM), and unsupervised (Hidden Markov Models) approaches, to analyze the multivariate time series EEG data.
The SMATE semi-supervised model demonstrates stable and consistent performance across all sleep stages, achieving up to 90% accuracy in its supervised form. It outperforms the TapNet semi-supervised model and the unsupervised Hidden Markov Models.
Ablation tests highlight the critical role of spatial and temporal feature extraction in the semi-supervised models' predictive performance, emphasizing the importance of preserving the spatio-temporal characteristics of the EEG signals.
Visualization using t-SNE confirms the SMATE model's ability to effectively distinguish between healthy and AD cases, validating the presence of disease-specific patterns in the sleep EEG data.
The study underscores the potential of semi-supervised learning in addressing the challenges associated with the scarcity of labeled data, a common issue in clinical settings, and its ability to leverage unlabeled data for improved AD detection.
Overall, this research contributes to the advancement of early AD detection through innovative deep learning approaches, highlighting the crucial role of semi-supervised learning in overcoming data limitations and the potential of sleep EEG signals as viable biomarkers for Alzheimer's disease.
Stats
The study utilized the following key metrics and figures:
Accuracy and standard deviation of the different models (supervised, semi-supervised, and unsupervised) across the four sleep stages (N1, N2, N3, and REM).
ROC/AUC curves for the models, demonstrating the trade-off between True Positive Rate and False Positive Rate.
Ablation test results, highlighting the impact of removing the spatial and temporal feature extraction blocks on the models' performance.