Year-long Research At Nanyang Technological University - Our Progress So Far
Updated: Jul 31
In March 2022, I was selected as one of the fifty research candidates in Singapore to conduct an AI and Robotics Research with Jingpeng from St Joseph Institution under Dr Lim. Our research is about using a new type of data - body movement, to innovate sleep staging process in research.
Sleep staging is the process of categorizing sleep into different stages based on specific patterns of brain waves and other physiological measures. There are five stages of sleep: four stages of non-rapid eye movement (NREM) sleep and one stage of rapid eye movement (REM) sleep. Each stage is characterized by distinct brain wave patterns and other physiological changes, and they typically occur in a cyclical pattern throughout the night.
Understanding sleep staging is important because it helps us understand the different functions and benefits of sleep and how to optimize sleep for optimal health and well-being. It also allows researchers and healthcare professionals to assess sleep health and diagnose sleep disorders, such as insomnia and sleep apnea. Overall, sleep staging is an important tool for understanding the complex and multifaceted nature of sleep and its role in our overall health and well-being.
The conventional sleep staging method is by using equipment that can detect neuron activities in the brain to determine the active level of the brain, hence the depth of the sleep. Over the years, a newer, quicker way to sleep stage is PSG, which measures plenty of signals such as heart rate and breathing rate to determine the sleep stage a patient/a subject is at. Our way of sleep staging, however, utilizes body movement recorded by a pillow with pressure sensers, requires zero equipment that require specific medical knowledge from the users nor any wearable detectors.
1. Factors (highlighted means detected by our self-designed machine)
Sound (background noise)
Body movement (pressure sensor)
Frontiers | Rapid Eye Movement Sleep, Sleep Continuity and Slow Wave Sleep as Predictors of Cognition, Mood, and Subjective Sleep Quality in Healthy Men and Women, Aged 20–84 Years | Psychiatry (frontiersin.org)
Blue light from electric devices
Controlled variable (should be banned x min before the sleep)
User's rating (comparison to the day before)
2. Sleep staging past papers (wake/N1–N2/N3/rapid eye movement REM)
Understanding sleep staging https://www.ncbi.nlm.nih.gov/books/NBK526132/
Using Heart rate + breathing rate: 1909.11141.pdf (arxiv.org)
Using Heart rate + body movement: Automatic sleep staging using heart rate variability, body movements, and recurrent neural networks in a sleep disordered population | SLEEP | Oxford Academic (oup.com)
The main methodological problem associated with the validity of actigraphic wake/sleep scoring is the relatively low ability to detect wakefulness during sleep periods.
In insomnia patients, for example, actigraphy highly overestimates the amount of sleep when patients try to fall asleep by lying in bed motionless for extended periods of time. Therefore, it is recommended to use actigraphy with complementary objective and subjective measures of sleep, especially in people with poor sleep quality
3. 3 Datasets:
4. Logic of our ai
The first one uses the body movement + sound + heart rate (possible if installing an app called sleep++ on apple watch) to identify the four sleep stages – n1, n2, n3, n2, REM, then repeat. The idea of developing one’s sleep quality is to test the user’s sleep stages without clinical experiments with EEG or PSG (heavy stupid uncomfortable machines detecting brain cell activity). We used Markov chain forecasting. The other papers mostly used long short-term memory (LSTM) & recurrent neural networks (RNN),
The second one uses a simple algorithm to account for the rest of the factors: light, humidity, temperature, background noise level. We used grey relations analysis to determine the contribution of each factor to the subjective level, then use a simple multi-variable function (also called fuzzy comprehension evaluation, ‘y=a1x1+a2x2+……’) to account for the weight (contribution level). We then optimized our results using bp neural networks.