Alzheimer's disease detection in PSG signals

TL;DR

The study addresses early Alzheimer's disease detection using sleep EEG recorded via polysomnography, tackling the challenge of limited labeled data with semi-supervised learning. It compares SMATE, TapNet, XCM, and HMM, with independent analyses per sleep stage and comprehensive ablation experiments to reveal the importance of spatio-temporal features. Findings show SMATE performs robustly with scarce labels, while XCM achieves the best accuracy in fully supervised scenarios; ablations confirm the critical role of spatial and temporal components. The work demonstrates the potential of semi-supervised sleep-EEG biomarkers for early AD, enabling practical clinical tools despite data scarcity and emphasizing stage-specific patterns for more accurate detection.

Abstract

Alzheimer's disease (AD) and sleep disorders exhibit a close association, where disruptions in sleep patterns often precede the onset of Mild Cognitive Impairment (MCI) and early-stage AD. This study delves into the potential of utilizing sleep-related electroencephalography (EEG) signals acquired through polysomnography (PSG) for the early detection of AD. Our primary focus is on exploring semi-supervised Deep Learning techniques for the classification of EEG signals due to the clinical scenario characterized by the limited data availability. The methodology entails testing and comparing the performance of semi-supervised SMATE and TapNet models, benchmarked against the supervised XCM model, and unsupervised Hidden Markov Models (HMMs). The study highlights the significance of spatial and temporal analysis capabilities, conducting independent analyses of each sleep stage. Results demonstrate the effectiveness of SMATE in leveraging limited labeled data, achieving stable metrics across all sleep stages, and reaching 90% accuracy in its supervised form. Comparative analyses reveal SMATE's superior performance over TapNet and HMM, while XCM excels in supervised scenarios with an accuracy range of 92 - 94%. These findings underscore the potential of semi-supervised models in early AD detection, particularly in overcoming the challenges associated with the scarcity of labeled data. Ablation tests affirm the critical role of spatio-temporal feature extraction in semi-supervised predictive performance, and t-SNE visualizations validate the model's proficiency in distinguishing AD patterns. Overall, this research contributes to the advancement of AD detection through innovative Deep Learning approaches, highlighting the crucial role of semi-supervised learning in addressing data limitations.

Figures (7)

• Figure 1: (a) Block representation of the acquisition of PSG recordings and hypnograms, highlighting the spatial information extracted followed by the preprocessing block and the training and prediction of the different models in each sleep stage signals. (b) Types of models tested, representing the amount of data and the labels used, together with the block of predictions obtained for each supervised, semisupervised, and unsupervised model.
• Figure 2: Workflow of EEG Signal Preprocessing. This figure illustrates the step-by-step transformation of raw EEG data from various databases into standardized, harmonized segments. It outlines the processes of hypnogram standardization, artifact removal, signal filtering, normalization, segmentation by sleep stages, and final resampling and segmentation into uniform 10-second segments. This flowchart demonstrates how disparate data sources are methodically processed to ensure consistency and reliability for subsequent analysis.
• Figure 3: Illustrative Overview of Model Structures Employed. This figure provides a schematic representation of each model's architecture. (a) The SMATE model, focusing on its integration of spatiotemporal features, detailed regularization processes, and the decoding block for final classification smate; (b) The TapNet model, showcasing its sequential information extraction process, the innovative random dimension permutation, the encoding of multivariate time series, and its unique attentional prototype learning mechanism tapnet; (c) The XCM model, emphasizing its dual approach with 1D temporal feature extraction and 2D spatial feature extraction, offering a holistic view of the data xcm; (d) The structure of the HMM model, describing its approach to PSG feature extraction, and detailing the implementation of a 5-state Hidden Markov Model to keep a balance between performance and computational cost, including state transitions for sequence analysis.
• Figure 4: Spatial Modeling Block (SMB) architecture (adapted from smate).
• Figure 5: Comparative analysis of SMATE supervised, SMATE with 10% labeled samples and TapNet supervised model performances. This figure displays the accuracy, F1-score metrics (with their standard deviations), and ROC curves for each model across different sleep stages. Each panel, from left to right, represents the results of each fold for the corresponding model: (a) Depicts the SMATE fully supervised model, showcasing its performance consistency across sleep stages; (b) Illustratres the SMATE 10% supervised model, showcasing its variability across different stages while still outperforming (c), which represents the TapNet fully supervised model, revealing greater variability compared to SMATE across the sleep stages.