Sleep deprived? New study says your performance will suffer

It’s true: A good night’s sleep really does help us do our best the next day, and a couple of bad nights of sleep could hurt us for days to come. That’s according to a new study from Microsoft’s research organization, which analyzed anonymized data on people’s online activities and sleep behavior to show how sleep quality impacts our ability to type queries on a search engine and click on the results.

The research reinforces the importance of catching ZZZs and highlights the negative influence of sleep deprivation on our ability to think and act.

“When you don’t sleep well, it affects your cognitive performance, which means your work performance and lots of other things,” said Tim Althoff, who led the research during a summer 2016 internship with Microsoft’s research organization in Redmond, Washington.

Eric Horvitz
Eric Horvitz

For example, the study shows that people who sleep less than six hours for two consecutive nights are sluggish for the next six days. While the researchers suspect other factors in addition to the two nights of poor sleep may contribute to the long recovery, the finding points to an intriguing direction for further studies.

The research also demonstrates a new method for using tens of millions of data points from people’s online activities to advance research in physiology.

“This kind of population-scale work has not been done before in sleep research,” said study co-author Eric Horvitz, technical fellow and director of Microsoft’s research lab in Redmond.

Keystrokes and clicks

The researchers mined data from 75 million keystrokes and clicks on Microsoft’s Bing search engine made by more than 30,000 consenting individuals wearing a fitness device that records bodily activity including sleep. Users of the device, the Microsoft Band, can opt to link the Band data with their Microsoft account, including their calendar and search activity, to gain personalized health insights.

“We can tell you things like if you have a packed day, you tend to sleep less well the night after,” said Ryen White, chief technology officer for health intelligence at Microsoft Health in Redmond. These findings are shared with users who opt in to the service.

For the new study, White and colleagues analyzed anonymized search engine interactions from this dataset to determine if they could infer information about Band wearers’ sleep quality and the daily rhythm of their cognitive performance.

Ryen White
Ryen White

Searching the web requires your brain to do a few complex tasks: Figure out what terms to search on, type the query and then process the results to decide which one to click, explained Jamie Zeitzer, a study co-author and assistant professor of psychiatry and behavioral sciences at Stanford University School of Medicine in Palo Alto, California.

“Even small differences in the amount of time it would take you to click on the result are indicative of how rapidly you are processing that information,” he said. “The idea is people have slower processing speeds as they get more tired.”

Overlap, and new directions

The web-scale study provides insight into the impact of sleep deprivation in the real world, where people compensate for lost sleep with extra coffee and naps, and otherwise adapt to life circumstances that limit pillow time, Zeitzer said. The findings largely overlap with results from small and controlled lab-scale studies, where participants are systematically sleep deprived and assessed on standardized tests.

For example, the web-scale study confirms previous lab-based research that shows individual cognitive performance varies throughout the day, aligning with circadian rhythms and so-called chronotype traits – early birds achieve peak performance earlier in the day than night owls. Whenever people wake up, they are sluggish for the first hour or two of their day, a phenomenon known as sleep inertia.

Across the board, keystroke and click speeds were slowest – by up to 31 percent – at 4 o’clock in the morning, an hour that previous research has shown that “your body thinks it should be sleeping,” said Althoff, who is pursuing a PhD in computer science at Stanford University.

Keystroke timing varies throughout the day. The fastest keystrokes occur a few hours after waking and the slowest occur during habitual sleep times. The shape of the curve is consistent with findings from controlled, lab-scale studies that show how cognitive performance varies throughout the day.

“The alignment with lab-based studies on circadian rhythms around sleep and performance is stunning and suggests that we can take sleep research forward and learn new things,” Horvitz said.

The methods can be used to explore patterns of sleep deprivation and recovery, for example.

The study also shows that staying up an extra hour, even if followed by a full night’s sleep, is correlated with slower performance the next day. Going to bed an hour earlier than normal, however, has a negligible effect.

“There is something not only about the duration of sleep, but also the timing,” said Althoff, who cautioned that further research is needed to better understand the phenomenon.

Additional ongoing research, Horvitz added, is examining the links between sleep deprivation and statistics on safety and accidents.

Without band aids

The Microsoft Band provided the researchers a way to establish so-called “ground truth” about their inferences on sleep patterns made from the keystrokes and clicks on Bing. Going forward, White said, the hope is to apply the technique to all web users based purely on their web interactions. No Band required.

White and Horvitz have already pioneered methods for using Bing search logs and other data to gain health insights on side effects from prescription drugs, the phenomenon of cyberchondria, in which health concerns are escalated by information on the web, as well as indicators that a person is likely to be diagnosed with pancreatic or lung cancer.

“Now we are showing how we can harness the web for doing large-scale physiological studies, which opens up new directions for understanding health and illness,” Horvitz said.

The methodology of mining computer interaction data could potentially be extended to study other kinds of diseases, such as the onset of neurodegenerative disorders from patterns of cursor movements and scrolling behavior, White noted.

“It all comes back to this notion of everyday interactions and being able to observe those in ways that allow us to make some really compelling inferences,” he said

The research will be presented at the World Wide Web Conference, 2017, which runs April 3-7 in Perth, Australia.


John Roach writes about Microsoft research and innovation. Follow him on Twitter.