Detecting Sleep States
Rank: 23rd out of 1925
Detect sleep onset and wake from wrist-worn accelerometer data
About the competition
The competition hosted by Child Mind Institute (CMI) aims to improve researchers’ ability to analyze accelerometer data for sleep monitoring of children. This method will enable large-scale studies of sleep. Based on a dataset provided by CMI, Epoch has to detect when a child falls asleep and when they wake up.
Child Mind Institute
The CMI, based in New York, is an independent, national nonprofit organization treating children with mental health and learning disorders. The institute globally serves as a hub for education, training, and convening.
Relevance
Sleep is crucial in regulating mood, emotions, and behaviour. Currently, Polysomnography (PSG) is the medical approach to sleep studies, so why is a new approach with AI necessary? PSG requires a dedicated physical space to conduct sleep assessments and on-site overnight staff to apply and monitor. Furthermore, the discomfort of the patients and financial costs should also be taken into account.
Wrist actigraphy bands, such as those from the CMI competition, are wireless, portable, and can be worn in a free-living environment. This makes them feasible for use in large-scale population research and makes them a good alternative for multi-day real-life sleep monitoring.
Currently, Dutch general practitioner Joyce Goris de Geus uses MotionWatch 8, similar to the wristwatch used in this competition, to monitor the sleep activity of children. Data in the form of actigraphs has to be processed manually. This way of analysing the is very time-consuming. AI could help speed up the process.
Technical details
The engineers receive time series with lengths of several weeks, and have to predict when children wake up and fall asleep. This is challenging because of the amount of data. Furthermore, some sensor data is invalid due to children taking off the wristband. Thus it is not only necessary to make predictions but also to learn when to withhold from making predictions. Tackling this problem requires a combination of handcrafted algorithms and time series models, such as transformers and GRUs.
UN Sustainable Development Goals
The competition was chosen based on the UN sustainable development goals. This competition links to goal #3: good health and well-being. With the research results doctors could efficiently interpret actigraphy data and quickly find the right treatment for their patients.
