Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Bachelor thesis - Generalization of sleep staging to low frequency data
How accurately can we classify sleep stages and, subsequently, estimate the number of sleep cycles a person has gone through using only aggregated low-frequency data derived from datasets that contain blood volume pulse and motion signals?
While asleep, we go through multiple sleep cycles each consisting of 4 different sleep stages. Based on the specific phase, sleep is classified as deep or light sleep, and rapid eye movement (REM) or non-REM sleep. Sleep staging has long been performed in sleep laboratories, also known as Polysomnography (PSG). In recent years, following the development of mobile devices such as fitness trackers, analyzing sleep phases has become less troublesome for participants of respective studies. To some extent, sleep staging has been implemented in common smartwatches. Fitbit, for instance, have implemented alarms aiming to wake you up during the light sleep phase proceeding your originally set alarm to reduce grogginess after waking up.
In this project, we assess how the accuracy of sleep staging techniques changes when the frequency of the supplied data is reduced. In particular, we have access to an unlabeled dataset (MSS dataset) which continuously records heart rate and IMU data at 1Hz. In addition, heart rate variability (HRV) features were continually calculated across 5 minute windows. We will replicate existing techniques on publicly available datasets that recorded motion signals from the wrist and either PPG or ECG signals and evaluate what techniques work best on unseen datasets at low frequencies. Based on the derived sleep cycles, we will analyze changes in HRV features to infer about the autonomic nervous system (ANS).
Check out https://siplab.org/teaching for further information. You should have some experience in data analytics in Python (or R).
While asleep, we go through multiple sleep cycles each consisting of 4 different sleep stages. Based on the specific phase, sleep is classified as deep or light sleep, and rapid eye movement (REM) or non-REM sleep. Sleep staging has long been performed in sleep laboratories, also known as Polysomnography (PSG). In recent years, following the development of mobile devices such as fitness trackers, analyzing sleep phases has become less troublesome for participants of respective studies. To some extent, sleep staging has been implemented in common smartwatches. Fitbit, for instance, have implemented alarms aiming to wake you up during the light sleep phase proceeding your originally set alarm to reduce grogginess after waking up.
In this project, we assess how the accuracy of sleep staging techniques changes when the frequency of the supplied data is reduced. In particular, we have access to an unlabeled dataset (MSS dataset) which continuously records heart rate and IMU data at 1Hz. In addition, heart rate variability (HRV) features were continually calculated across 5 minute windows. We will replicate existing techniques on publicly available datasets that recorded motion signals from the wrist and either PPG or ECG signals and evaluate what techniques work best on unseen datasets at low frequencies. Based on the derived sleep cycles, we will analyze changes in HRV features to infer about the autonomic nervous system (ANS).
Check out https://siplab.org/teaching for further information. You should have some experience in data analytics in Python (or R).
We will answer the following questions:
How well can we classify sleep stages using the supplied low-frequency data?
What are optimal techniques for different signal frequencies (e.g., black-box methods & deep learning versus interpretable statistical techniques)?
How does the ANS behave during different sleep stages (analysis based on our predictions)?
Bonus: Can we include uncertainty estimates?
We will answer the following questions: How well can we classify sleep stages using the supplied low-frequency data? What are optimal techniques for different signal frequencies (e.g., black-box methods & deep learning versus interpretable statistical techniques)? How does the ANS behave during different sleep stages (analysis based on our predictions)? Bonus: Can we include uncertainty estimates?
Max Moebus (email: firstname.lastname@inf.ethz.ch)
Max Moebus (email: firstname.lastname@inf.ethz.ch)