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

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: 29% increase in classification accuracy with ECG pretraining and PPG fine-tuning

• Primary Outcome: Multi-stage sleep classification accuracy using PPG signals

• Nnt Or Sensitivity: Sensitivity and specificity not reported - only overall accuracy improvement mentioned

• Confidence Interval: Not reported

Clinical Application

Potentially feasible for consumer applications using existing wearable technology, but clinical validation needed Could complement sleep disorder services under Medicare if validated clinically, though TGA approval would be required for diagnostic claims. May support RACGP sleep health initiatives and reduce PSG waiting lists if proven accurate Adults using consumer sleep technology devices, though validation limited to healthy participants aged 22-71 years

Abstract

Objective.This work aims to enable adaptive consumer sleep technologies (CSTs) for sleep intervention by developing a deep learning model for sleep stage classification using wearable sensor data.Approach.We propose an end-to-end deep learning approach leveraging Photoplethysmography (PPG) signals, commonly available in CSTs. Model performance is improved by pretraining with electrocardiography (ECG) from the large-scale multi-ethnic study of atherosclerosis dataset datasets. Training and evaluation are conducted with the Dataset for Real-time sleep stage EstimAtion using Multisensor wearable Technology and an additional dataset comprising of 13 participants (aged 22-71 years) without prior known sleep disorders. The dataset contains combined synchronized polysomnography (PSG) and Empatica E4 wearable data, annotated with American Academy of Sleep Medicine (AASM) sleep stages.Main results.The proposed method demonstrates sleep stage classification from minimally processed PPG signals for real-time intervention. While ECG-trained models are not directly transferable to PPG, fine-tuning significantly improves performance, achieving up to a 29% increase in multi-stage classification accuracy.Conclusion.Pretraining with ECG and fine-tuning with PPG increases sleep stage classification for end-to-end deep learning models, exceeding previous efforts particularly in 3-stage sleep classification.Significance.This work contributes to sleep health by developing a sleep stage classification model for minimally processed PPG sensor data and takes a step further towards making adaptive CSTs feasible for use with wearable sensors.