Classification of 16 adult sleep patterns based on large-scale sleep analysis
A research group of The University of Tokyo led by Professor Hiroki Ueda (also a Riken team leader) and Machiko Katori, and Assistant Professor Shoi Shi (RIKEN) used ACCEL(1), an original machine learning algorithm developed by their research laboratory, to determine sleep and waking states based on arm acceleration and converted the acceleration data of approximately 100,000 people in the UK Biobank(2) into sleep data, which was then analyzed in detail. They found that the sleep patterns of these 100,000 people could be classified into 16 different types.
The research group first focused on the arm acceleration data of approximately 100,000 people in the UK Biobank. This data was obtained from men and women in their 30s to 60s, mainly in the UK, who were measured for up to 7 days using wristband-type accelerometers. Using an algorithm (ACCEL) they had developed in 2022, the research group generated sleep data(3) for approximately 100,000 people from the acceleration data. The obtained sleep data were converted into 21 sleep indicators, and then, using dimension reduction(4) and clustering(5) methods, the sleep patterns were classified into 8 different clusters. These included clusters related to “social jet lag” and clusters characterized by mid-onset awakenings and considered insomnia, enabling the extraction of clusters related to lifestyles and to sleep disorders. Next, in order to examine sleep patterns associated with sleep disorders in more detail, the research group focused on 6 of the 21 sleep indicators, including sleep duration and intermediate waking time, which are known to be closely related to sleep disorders. By applying the same analysis to data where one indicator deviated significantly from general sleep (data in the upper 2.28th percentile or higher or the lower 2.28th percentile or lower (6) in the overall distribution), they were able to classify the data into 8 clusters. These included clusters related to morning-types and evening-types. They also identified several clusters associated with insomnia, and were able, along with the clustering using the entire dataset, to classify 7 types of sleep patterns associated with insomnia.
Thus, by analyzing sleep on a large scale, they have revealed the landscape of human sleep phenotype. This study has made it possible to quantitatively classify clusters related to lifestyle such as “social jet lag” and morning/evening types, which are usually difficult to determine with short-term PSG measurements(7), In addition, detailed analysis of outlier and classification of sleep patterns revealed 7 clusters related to insomnia. These clusters are classified based on new indicators differing from conventional methods, and are expected to be useful in the construction of new methods in terms of diagnosing insomnia and proposing treatment methods.
These results were obtained through the “Ueda Biological Timing Project,” ERATO Program funded by the Japan Science and Technology Agency (JST). In this project, JST develops “systems biology that contributes to understanding human beings,” using sleep-wake rhythms as a model system, and aims to understand in human sleep-wake behavior the “biological time” information that extends from molecules to individual humans living in society.
Notes:
(1) ACCEL : An original sleep determination algorithm developed by the research team. For details, refer to the following paper. “A jerk-based algorithm ACCEL for the accurate classification of sleep-wake states from arm acceleration” DOI: 10.1016/j.isci.2021.103727
(2) UK Biobank: A large research database containing genetic and health information on approximately 500,000 British participants. This study uses acceleration data for approximately 100,000 people as well as the linked gender and age data.
(3) Sleep data: Time-series data with intervals of 30 seconds labeled as sleeping or waking. PSG measurement uses diverse data measured by specialist technicians to create sleep data. In this study, sleep data was obtained by applying ACCEL to accelerometers.
(4) Dimension reduction method: A method to reduce the number of dimensions of data. This makes it possible to extract important information from the data and to capture the characteristics of the data. In this study, UMAP (Uniform Manifold Approximation and Projection) is used.
(5) Clustering method: A method of classifying data into clusters based on similarities among the data. There are two types of clustering methods: supervised clustering, which uses correct data for clustering, and unsupervised clustering, which does not. In this study, the unsupervised clustering method, DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is used.
(6) Upper and lower percentiles:?The value in any given percent when the values are arranged in descending order is called the upper percentile. Conversely, the value in any given percent when the values are listed in ascending order is called the lower percentile. For instance, data above the upper 2.28th percentile or below the lower 2.28th percentile in normal distribution refers to data deviating from the mean by more than twice the standard deviation (2SD).
(7) Polysomnography (PSG): In PSG measurements, multiple electrodes and sensors are attached to the examinee to measure brain waves, eye movements, respiratory status, and electrocardiogram status. It is currently the most accurate measurement method used to determine human sleep patterns. It is also used to diagnose sleep disorders.