Multimodal Ambulatory Sleep Detection Using LSTM Recurrent Neural Networks

Area of Science:

• Biomedical Engineering
• Computer Science
• Sleep Medicine

Background:

• Accurate long-term sleep monitoring is crucial for health but current ambulatory methods often rely on unreliable self-reported data.
• Existing methods face challenges with data accuracy and participant compliance.

Purpose of the Study:

• To develop and validate an unobtrusive, accurate ambulatory method for monitoring long-term sleep patterns.
• To improve sleep/wake state detection and sleep onset/offset determination using multimodal sensor data.

Main Methods:

• Developed a novel algorithm employing long-short-term memory (LSTM) recurrent neural networks to synthesize temporal information from multimodal data.
• Collected 5580 days of data from 186 participants using smartphones and wearable devices.
• Compared the LSTM algorithm against non-temporal machine learning methods and state-of-the-art actigraphy software.

Main Results:

• The LSTM method achieved 96.5% accuracy in sleep/wake classification.
• Sleep onset and offset detection yielded F1 scores of 0.86 and 0.84, respectively, with low mean absolute errors (5.0 and 5.5 min).
• LSTM performance was statistically superior to existing non-temporal algorithms and actigraphy software.

Conclusions:

• The proposed LSTM-based algorithm offers a highly accurate and unobtrusive solution for ambulatory sleep monitoring.
• The method demonstrates good generalization and potential for near real-time application with minimal performance reduction.
• This advancement can significantly improve long-term health monitoring through objective sleep pattern analysis.