Estimating the sleep period time window based on a hip- worn accelerometer collected in children and adults

Sleep is a fundamental rhythm of human life, a daily cycle where the body repairs itself and the mind processes the world. To understand how much sleep people get, and how that amount relates to their health, scientists need to measure the exact window of time between when a person falls asleep and when they wake up. This period is known as the sleep period. For decades, researchers have relied on small devices worn on the wrist, similar to a watch, to track movement and guess when a person is sleeping. These devices work well because the wrist moves noticeably when a person is awake and stays still when they are asleep. However, many large-scale studies that track the health of thousands of people over many years use a different approach. Instead of a wristband, these studies place the sensor on the hip, much like a belt, to measure all physical activity throughout the entire day. While this method captures a complete picture of daily movement, the standard computer programs used to find sleep times were built for wrist data, not hip data. This created a gap: scientists had a wealth of hip-worn data but lacked a reliable way to extract accurate sleep information from it.

A team of researchers set out to bridge this gap by testing and refining two open-source computer programs designed to find sleep windows. They focused on a method that uses raw movement data collected from sensors worn on both the wrist and the hip. The study involved 303 participants, including 109 children between the ages of eight and twelve, and 194 young adults. Each person wore two sensors simultaneously for six nights: one on their wrist and one on their hip. To ensure the computer programs were being tested against a reliable standard, the participants also kept a sleep diary, writing down exactly when they went to bed and when they got up. The researchers used the wrist data, guided by these diaries, as a trusted reference point. They then took the hip data and ran it through two different algorithms, which are sets of rules that tell a computer how to interpret the movement. One algorithm, called HDCZA, looks for long periods where the sensor does not move much. The other, called HorAngle, is a newer method specifically designed for the hip; it searches for the time when the sensor is lying flat, which suggests the person is lying down in bed.

The researchers did not simply run the programs with their default settings. Instead, they carefully tuned the internal rules of both algorithms using a portion of their data to find the settings that produced the most accurate results. They then tested these tuned settings on the remaining participants to see how well they performed. The results showed that both programs could successfully estimate sleep times from the hip sensors. When comparing the computer's guess to the actual time the participant wrote in their diary, the average difference was very small. For the time a person went to sleep, the HDCZA program was off by an average of three minutes, while the HorAngle program was off by zero minutes on average. For the time a person woke up, the differences were similarly small, with HDCZA off by four minutes and HorAngle by four minutes in the opposite direction. The total amount of sleep time estimated by the programs also matched the diary records closely, with differences averaging only a few minutes.

While the average errors were small, the researchers noted that for any single night, the guess could be off by a larger amount, sometimes by several hours. This is expected, as individual nights of sleep can be irregular. However, when the researchers looked at the overall patterns, both algorithms were quite good at ranking people correctly. If one person slept longer than another, the computer programs were able to identify that difference most of the time. The study also found that the accuracy of these estimates did not depend on whether the participant was a child or an adult, or whether they were male or female. The main factor that influenced the accuracy was how long the sensor was not being worn. If the device was taken off during the night, the computer had less information to work with, making the estimate less reliable.

The study concludes that it is possible to accurately estimate sleep windows from hip-worn sensors, provided the right computer algorithms are used. The HDCZA algorithm has the advantage of not needing any special adjustments for the specific type of sensor, making it a flexible tool. The HorAngle algorithm, which relies on detecting when the sensor is horizontal, performed just as well but requires the researcher to correctly identify which axis of the sensor represents the length of the body. Both methods offer a valuable way to unlock sleep data from the massive amounts of hip-worn activity data already collected in large health studies. This means that researchers can now look back at decades of existing data and gain new insights into how sleep patterns affect health across the entire population, without needing to re-measure everyone with a wrist sensor. The work demonstrates that with careful tuning, the same movement data that tracks a person's daily steps can also reveal the quiet hours of their rest.