AI-Driven sleep staging from actigraphy and heart rate

May 17, 2023PloS one

Using AI to Determine Sleep Stages from Movement and Heart Rate

AI simplified

Abstract

achieves an overall accuracy of 79% for three-class sleep staging using data from consumer-grade wrist-worn devices.

  • An artificial intelligence technique called sequence-to-sequence LSTM is presented for automated mobile sleep staging.
  • The model classifies sleep into three classes (wake, NREM, REM) and four classes (wake, light, deep, REM) based on wrist-accelerometry and coarse heart rate data.
  • In the MESA cohort, the model achieved a weighted F1 score of 0.80 for three-class staging and 0.72-0.73 for four-class staging.
  • In the MrOS cohort, the model achieved a weighted F1 score of 0.77 for three-class staging and 0.68-0.69 for four-class staging.
  • The method demonstrates high accuracy in predicting the duration of each sleep stage, particularly deep sleep, which is often underrepresented in wearables-derived data.
  • By addressing class imbalance, the model accurately estimates deep sleep time, with results closely matching traditional PSG measurements.

AI simplified

Key numbers

79%
Overall Accuracy for Three-Class Staging
Achieved using data from 808 participants in the MESA cohort.
70-72%
Overall Accuracy for Four-Class Staging
Results from the MESA cohort using low-resolution data.
0.61±0.69 hours
Deep Sleep Duration Prediction
Compared to PSG's ground truth of 0.60±0.60 hours.

Full Text

What this is

  • This research presents a novel AI technique called for sleep staging using data from consumer-grade wearables.
  • classifies sleep into three and four stages using wrist-accelerometry and heart rate data.
  • The method aims to overcome the limitations of traditional polysomnography (PSG) by providing accurate sleep assessments without the need for extensive equipment.

Essence

  • achieves an overall accuracy of 79% for three-class sleep staging and 70-72% for four-class staging using low-resolution data from wearable devices, demonstrating its potential for practical sleep monitoring.

Key takeaways

  • achieves 79% overall accuracy for three-class sleep staging using wrist-accelerometry and heart rate data from 808 participants in the MESA cohort.
  • For four-class sleep staging, reaches 70-72% accuracy in the same cohort, indicating its capability to classify sleep stages beyond simple wake/sleep distinctions.
  • The method effectively estimates deep sleep duration, crucial for monitoring sleep quality, with predicting 0.61±0.69 hours of deep sleep compared to PSG's 0.60±0.60 hours.

Caveats

  • The heart rate data used for validation comes from lab-grade devices, which may not fully represent the capabilities of consumer-grade wearables.
  • The study's reliance on coarse heart rate measures may limit the accuracy of sleep onset latency estimates, which are inherently challenging to assess.

Definitions

  • SLAMSS: An AI technique for automated mobile sleep staging that classifies sleep into three or four stages using activity and heart rate data.

AI simplified

what lands in your inbox each week:

  • 📚7 fresh studies
  • 📝plain-language summaries
  • direct links to original studies
  • 🏅top journal indicators
  • 📅weekly delivery
  • 🧘‍♂️always free