基于多时间粒度深监督的SpO₂睡眠呼吸暂停检测方法

SpO₂ Sleep Apnea Detection Method Based on Multi-Time Granularity Deep Supervision

摘要: 血氧饱和度(SpO₂)信号能够反映睡眠过程中由异常呼吸事件引起的血氧下降与恢复过程，在睡眠呼吸暂停检测中具有重要应用价值。与脑电等生理信号相比，SpO₂获取简便且适合长期连续监测，但现有方法多基于单一时间粒度分析，难以兼顾局部细节与长程时序结构。针对上述局限，文章提出一种基于多时间粒度深监督的SpO₂睡眠呼吸暂停检测方法。该方法以1小时、1 Hz采样的SpO₂序列为输入，构建一维U-Net编码器–解码器结构，并在1 s、5 s、15 s、30 s和60 s五个时间尺度上引入监督分支，通过多尺度标签构建与联合优化增强模型对跨尺度异常模式的表征能力。实验结果表明，所提方法在SpO₂单模态任务上具有良好稳定性，五折交叉验证平均准确率为79.96%，平均F1-Score为77.39%。这证明多时间粒度建模与深监督联合优化可提升模型对呼吸暂停事件的识别能力，为低侵入式睡眠呼吸暂停筛查提供了一种有效方法。

Abstract: Blood oxygen saturation (SpO₂) signals can reflect the desaturation and recovery processes caused by abnormal respiratory events during sleep, and thus have important application value in sleep apnea detection. Compared with physiological signals such as EEG, SpO₂ is easy to acquire and suitable for long-term continuous monitoring; however, existing methods are mostly based on a single temporal resolution, making it difficult to capture both local details and long-range temporal structures. To address this limitation, this paper proposes a multi-temporal granularity deep supervision-based method for SpO₂ sleep apnea detection. The method takes a 1-hour SpO₂ sequence sampled at 1 Hz as input, constructs a one-dimensional U-Net encoder-decoder architecture, and introduces supervision branches at five temporal scales (1 s, 5 s, 15 s, 30 s, and 60 s). A multi-scale label construction and joint optimization strategy is adopted to enhance the model’s ability to represent cross-scale abnormal patterns. Experimental results show that the proposed method achieves good stability in the single-modality SpO₂ task, with an average accuracy of 79.96% and an average F1-Score of 77.39% under five-fold cross-validation. These findings verify that the combination of multi-temporal modeling and deep supervision improves the detection of apnea events, providing an effective approach for low-invasive sleep apnea screening.