Towards real-time sleep stage classification: a deep learning approach leveraging PPG and ECG

Area of Science:

• Biomedical Engineering
• Sleep Science
• Artificial Intelligence

Background:

• Consumer Sleep Technologies (CSTs) lack robust sleep intervention capabilities.
• Accurate sleep stage classification from wearable sensors is crucial for personalized sleep health.

Purpose of the Study:

• To develop a deep learning model for sleep stage classification using Photoplethysmography (PPG) signals from wearable sensors.
• To enhance model performance through pretraining with Electrocardiography (ECG) data.

Main Methods:

• An end-to-end deep learning model was developed using PPG signals.
• Model pretraining utilized Electrocardiography (ECG) data from the Multi-Ethnic Study of Atherosclerosis (MESA) dataset.
• Training and evaluation were performed on the Dataset for Real-time sleep stage EstimAtion using Multisensor wearable Technology (DREAMT) and a custom dataset with synchronized polysomnography (PSG) and Empatica E4 wearable data.

Main Results:

• The model achieved sleep stage classification from minimally processed PPG signals suitable for real-time intervention.
• Fine-tuning ECG-pretrained models on PPG data significantly improved multi-stage classification accuracy by up to 29%.

Conclusions:

• Pretraining with ECG and fine-tuning with PPG enhances deep learning models for sleep stage classification.
• This approach surpasses previous efforts, especially for 3-stage sleep classification, advancing adaptive CSTs.