We developed a hybrid framework that integrates handcrafted physiological descriptors with pretrained PSG-derived representations under the challenge training constraints. The handcrafted branch comprises demographic variables, physiological summary-statistics features, and algorithmic annotation features. The learned branch uses a pretrained sleep representation model to generate a pooled record-level embedding from available PSG modalities. These complementary representations are fused by feature concatenation and modeled with a compact downstream classifier.
Controlled local ablation supported the proposed design. On a balanced local holdout, the encoder-only model achieved AUROC 0.432, whereas the hybrid model achieved AUROC 0.698. This indicates that frozen pretrained pooled embeddings were not sufficiently discriminative in isolation, but contributed useful predictive information when integrated with handcrafted physiological features.
Our strongest unofficial-phase submission is a refined hybrid variant with quality-aware feature handling that preserves the same frozen-embedding fusion framework. For team OUS_IVS, this best unofficial entry achieved an unseen validation AUROC of 0.586, establishing the hybrid family as the current best-performing direction.
Current work extends this framework by retaining the handcrafted branch, pooled representation design, and simple fusion mechanism while introducing limited task-specific adaptation in the upper temporal portion of the pretrained encoder, trained jointly with a compact neural prediction head. This next step is intended to determine whether narrowly constrained encoder adaptation can improve generalization beyond the current frozen-hybrid formulation while preserving computational tractability, modeling discipline, and submission robustness.