Abstract
Obstructive sleep apnea (OSA) is characterized by frequent episodes of sleep-disordered breathing (SDB), which are often accompanied by leg movement (LM) events, especially periodic limb movements during sleep (PLMS). Traditional single-event detection methods often overlook the dynamic interactions between SDB and LM, failing to capture their temporal overlap and differences in duration. To address this, we propose Attention-enhanced CRF with U-Net (AttenCRF-U), a novel joint detection framework that integrates multi-head self-attention (MHSA) within an encoder-decoder architecture to model long-range dependencies between overlapping events and employs multi-scale convolutional encoding to extract discriminative features across different temporal scales. The model further incorporates a conditional random field (CRF) to refine event boundaries and enhance temporal continuity. Evaluated on clinical PSG recordings from 125 OSA patients, the model with CRF improved the average F1 score from 0.782 to 0.788 and reduced temporal alignment errors compared with CRF-free baselines. The joint detection strategy distinguished respiratory-related leg movements (RRLMs) from PLMS, boosting the PLMS detection F1 score from 0.756 to 0.778 and the SDB detection F1 score from 0.709 to 0.728. By integrating MHSA into a CRF-augmented U-Net framework and enabling joint detection of multiple event types, this study presents a novel approach to modeling temporal dependencies and event co-occurrence patterns in sleep disorder diagnosis.