基于因果历史上下文的轻量级睡眠分期方法

A Lightweight Sleep Staging Method Based on Causal Historical Context

摘要: 针对轻量级单片段睡眠分期难以充分利用跨片段历史信息，以及部分序列模型在连续预测过程中依赖未来上下文的问题，提出一种基于因果历史上下文的轻量级睡眠分期方法MSA-TC-Lite。该方法冻结MSA-CNN small的局部特征编码路径，将当前片段与此前9个片段构成单向历史窗口，并利用轻量时序模块建模历史依赖；连续推理时通过缓存历史特征减少重复编码。在Sleep-EDF-20上采用受试者级十折交叉验证，MSA-TC-Lite在41,957个具有完整历史上下文的测试片段上取得83.89%的Accuracy和78.15%的Macro-F1，较MSA-CNN small分别提高2.18和4.77个百分点，其中N1和REM的F1分别提高14.88和6.50个百分点。No-History对照和历史上下文扰动结果进一步表明，性能改善与真实历史内容及其时间排列的利用具有较一致的关联。缓存式连续推理将稳态计算量由12.164 M MACs降低至1.319 M MACs。结果表明，该方法能够在不使用未来信息的条件下改善轻量级睡眠分期性能，并减少连续推理过程中的重复计算。

Abstract: To address the difficulty of lightweight single-epoch sleep staging in fully exploiting cross-epoch historical information and the reliance of some sequence-based models on future context during continuous prediction, a lightweight sleep staging method based on causal historical context, termed MSA-TC-Lite, is proposed. The method freezes the local feature encoding pathway of MSA-CNN small and constructs a unidirectional historical window consisting of the current epoch and the preceding nine epochs. A lightweight temporal module is then used to model dependencies across historical representations. During continuous inference, previously computed historical features are cached to avoid repeated encoding. Subject-wise ten-fold cross-validation was conducted on Sleep-EDF-20. On 41,957 test epochs with complete historical context, MSA-TC-Lite achieved an Accuracy of 83.89% and a Macro-F1 of 78.15%, outperforming MSA-CNN small by 2.18 and 4.77 percentage points, respectively. The F1 scores of N1 and REM increased by 14.88 and 6.50 percentage points. The No-History control and historical-context perturbation analyses further showed that the performance improvement was consistently associated with the use of real historical content and its temporal arrangement. Cached continuous inference reduced the steady-state computational cost from 12.164 M MACs to 1.319 M MACs. These results indicate that MSA-TC-Lite improves lightweight sleep staging without using future information while reducing redundant computation during continuous inference.