Background: Fusing high-dimensional Polysomnography (PSG) signals with clinical metadata is crucial for predicting sleep-related cognitive impairments. However, effectively modeling the massive length and complex temporal dynamics of sleep data in a unified pipeline remains computationally challenging.
Methods: We propose a novel end-to-end multimodal deep learning architecture. We designed a Hierarchical Transformer to process raw multichannel PSG signals. First, to standardize the input length, an Epoch Encoder processes a fixed sequence of 384 30-second epochs per patient to capture intra-epoch micro-patterns. Next, a Sequence Transformer models the inter-epoch macro-dynamics across the entire sleep duration. Crucially, rather than relying on standard pooling, we utilized Attentive Statistics Pooling (ASP) to dynamically aggregate the sequence representations, selectively focusing on the most critical temporal segments indicative of cognitive impairment. These aggregated features are natively fused with auxiliary data (e.g., sleep stage embeddings and global metadata) to output the final classification in a single forward pass.
Results: Evaluated via 5-fold cross-validation, our end-to-end model achieved an internal Area Under the ROC Curve (AUROC) of 0.683. In the unofficial phase of the challenge, our team BAPORLab achieved an AUROC of 0.745, securing 3rd place. The hierarchical ASP strategy successfully captured localized pathological patterns within the massive PSG signals without the need for decoupled processing.
Conclusion: Leveraging this end-to-end framework provides a highly scalable and accurate solution for analyzing long time-series PSG data. This approach demonstrates strong potential for utilizing extensive sleep databases in advanced clinical diagnostics.