ESSN: An Efficient Sleep Sequence Network for Automatic Sleep Staging.

Area of Science:

• Biomedical Engineering
• Computer Science
• Sleep Medicine

Background:

• Advanced automatic sleep staging algorithms show promise but struggle with limited computational resources and N1 stage confusion.
• Existing methods are not well-suited for portable sleep detection or consumer-level sleep disorder screening.

Purpose of the Study:

• To develop an efficient automatic sleep staging algorithm (ESSN) suitable for low-power computing environments.
• To address the N1 stage confusion issue prevalent in current sleep staging algorithms.
• To achieve high accuracy in automatic sleep staging with reduced computational cost.

Main Methods:

• Proposed an Efficient Sleep Sequence Network (ESSN) with a specialized structure for computational efficiency.
• Introduced a novel N1 structure loss function leveraging N1 transition probability to mitigate confusion.
• Evaluated the ESSN on the SHHS dataset comprising 5,793 subjects.

Main Results:

• ESSN achieved an overall accuracy of 88.0%, macro F1 of 81.2%, and Cohen's kappa of 0.831 on the SHHS dataset.
• The model demonstrated low computational requirements with 0.27M parameters and 0.35G floating-point operations for a 200-sample input.
• ESSN exhibited inference speeds twice as fast as L-SeqSleepNet on the same hardware, alongside superior accuracy.

Conclusions:

• The ESSN offers an efficient and accurate solution for automatic sleep staging, particularly in resource-constrained settings.
• The novel N1 structure loss effectively reduces N1 stage confusion, enhancing diagnostic reliability.
• ESSN presents a competitive advantage over state-of-the-art methods for clinical assistance and consumer applications.