Introduction
Early diagnosis of Alzheimer’s Disease (AD) remains a critical challenge in geriatric medicine. Conventional diagnostic methods often rely on clinical symptoms that manifest only after significant neurodegeneration has occurred. Recent research indicates that sleep disruptions, detectable via polysomnography (PSG), may precede cognitive decline, offering a window for earlier intervention. However, the application of Deep Learning (DL) to PSG data is hindered by the scarcity of expert-labeled clinical data from AD patients. This study investigates a semi-supervised learning framework to bridge this gap, leveraging the SMATE model to achieve high diagnostic accuracy even with limited labeled data.
Methodology
The researchers evaluated four modeling approaches to classify EEG signals as "Healthy" or "AD":
- SMATE (Semi-Supervised Spatio-Temporal Representation Learning): A framework utilizing an autoencoder to map signals into an embedding space, followed by a three-step regularization process to cluster AD and healthy patterns using both labeled and unlabeled data.
- TapNet: A semi-supervised attentional prototypical network serving as a secondary benchmark.
- XCM: A supervised model employing 1D and 2D convolutions to extract temporal and spatial features simultaneously.
- Hidden Markov Models (HMM): An unsupervised statistical baseline.
A central technical contribution is the integration of spatio-temporal analysis. By examining both the timing and the scalp location of EEG signals, the models capture complex inter-regional brain dynamics often disrupted in AD.
Datasets and Preprocessing
To ensure robust performance, the study combined four distinct databases:
- AD Patient Database: Clinical data from 58 individuals with mild-to-moderate AD.
- Healthy Controls: Data aggregated from ISRUC-SLEEP (10 subjects), Dream Open Dataset-Healthy (25 subjects), and the Sleep Disorders Research Center (11 subjects).
Data harmonization involved standardizing sleep stages (N1, N2, N3, REM) and resampling all EEG signals to 128 Hz to ensure consistency across disparate sources.
Key Insights
- Optimal Sleep Stage: The N1 sleep stage (transition from wakefulness to sleep) consistently yielded the most stable and accurate classification results.
- Robustness to Label Scarcity: The SMATE model maintained high performance even when trained with only 10% of available labels, significantly outperforming TapNet in data-constrained scenarios.
- Criticality of Spatial Features: Ablation studies confirmed that spatial modeling is essential. Removing the ability to analyze inter-channel relationships led to a sharp decline in accuracy, highlighting the importance of regional brain communication patterns in AD detection.
- Age-Related Confounding: The authors implemented specific tests to distinguish between normal age-related EEG changes and AD-specific signatures, ensuring the models were not simply identifying the age of the subjects.
Results Evaluation
The supervised XCM model achieved the highest overall accuracy, ranging between 92% and 94% across all sleep stages. The SMATE model reached 90% accuracy in its supervised configuration and demonstrated remarkable stability in semi-supervised settings, making it highly suitable for clinical environments where labeled data is rare.
Visual validation via t-SNE embeddings confirmed that the SMATE model successfully forced AD and healthy samples into distinct, separable clusters, proving the model's ability to learn meaningful physiological signatures of the disease.
Conclusion and Future Work
This research demonstrates the potential of semi-supervised Deep Learning to transform routine sleep studies into early-warning systems for Alzheimer’s Disease. While the results are promising, future efforts should focus on age-matched datasets to further decouple aging from neurodegeneration. Additionally, automating sleep stage classification and improving robustness to signal artifacts (e.g., muscle movement) are necessary steps toward deploying these diagnostic tools in real-world clinical settings.
Summary updated