Introduction: Polysomnography (PSG) recordings comprise multiple physiological signals with complex temporal structures. Modeling these data requires capturing both fine-grained signal morphology and long-range dependencies across entire sleep studies. Transformer architectures are well-suited to this setting, as self-attention enables modeling of long-range temporal relationships and supports cross-modal integration.
Methods: We propose a hierarchical transformer-based framework that learns multiscale representations of PSG data at the time-segment and patient levels. Each PSG recording is decomposed into fixed-length segments (e.g., 30 seconds), preserving clinically meaningful temporal resolution. A channel-level encoding transformer processes raw multichannel segments to learn embeddings that capture both intra-signal morphology and cross-modal interactions. These embeddings are then aggregated by a patient-level transformer to model long-range temporal dependencies across the full recording duration. The resulting patient representation is fused with clinical metadata and passed to a downstream classifier to predict future cognitive impairment.
Current Progress: We implemented a preliminary baseline using engineered features over 30-second windows. These segment-level predictions were aggregated to the patient level via majority voting. This approach achieved an AUROC of 0.599, AUPRC of 0.077, accuracy of 0.739, and F-measure of 0.093 on the unofficial validation set, indicating modest but consistent performance above chance. These results support the feasibility of segment-level modeling followed by patient-level aggregation for this task.
Next Steps: We will replace engineered features with learned representations from the channel-level transformer encoder and substitute the voting-based aggregation with a patient-level transformer for end-to-end temporal modeling. To ensure the learned representations are physiologically meaningful, in subsequent iterations we will introduce a physics-informed auxiliary objective based on wavelet-domain transformations that encourages the preservation of time-frequency characteristics inherent to sleep. Additionally, we will explore multimodal attention mechanisms to better capture interactions across modalities.