Everyone sleeps, but we have few tools to measure the amount of sleep the world is getting. AI and sleep can help us study global shocks in near real time.
It is not uncommon to hear people complaining of fatigue several times a day, but why? Sleep is fundamental to human health, but because it is so private, there are few tools to measure how much sleep each person gets. Existing methods use time diaries, sleep surveys, sleep laboratories, or more recently wearable technology to measure sleep. But none of these approaches is ready to tackle the global sleep-loss epidemic.
a Monash University The study asked a different question: Does the massive amount of data we all generate as we connect and disconnect from the Internet help researchers understand sleep?
As Internet addresses go online and offline throughout the day, they track daily Human behavior cycle: A trough in the early hours, followed by increasing activity during the day reaching a peak in the evening, and then a sharp decline overnight.
And yet, no two cycles are the same: the day of the week matters (heading downtown on a Friday night slows down Internet activity), stay-at-home orders certainly matter (we go online earlier and longer), and even activity drops at prayer times. It is traditionally observed in Muslim regions during Ramadan.
This American Time Use Survey (ATUS) asks Americans about their previous day’s activities, including when they woke up and when they went to bed. This Monash The study used survey data from 81 US cities over a six-year period to calculate when residents went to sleep and when they woke up each year, then used Internet activity-data to make the same calculation.
The researchers then trained a machine-learning algorithm to track how changes in internet use over a day were related to the average wake-up and sleep times in each city.
When the algorithm was asked to predict the expected average sleep duration for a city it had never seen before, it was accurate to within 20 minutes. When estimating the average time to wake up in the morning, it was accurate to within nine minutes.
The researchers replicated this result when using daily electricity-demand data instead of Internet data to predict sleep. But there’s something fundamentally different about internet-activity measurements compared to electricity-demand data: global availability. The US have highly functional electricity bureaucracies, but not all countries do. Internet-activity measurement, on the other hand, can be measured remotely and continuously for any Internet-connected device on the planet.
This suggests that how much sleep we are collectively getting can be estimated in near real time for any (Internet-connected) city on Earth.
This type of research has a wide range of applications, including impact mapping during natural disasters, documenting Internet shutdowns associated with human-rights violations, and providing Internet-availability assessments during the Russo-Ukrainian war.
Whether this approach can be applied globally remains to be seen. Americans’ technology and sleep habits can be unique. If so, the artificial intelligence (AI) model learning the internet-sleep association in the US will fall outside its borders. Another potential pitfall is that internet patterns can be affected by the mix of technologies – the internet signature of a ‘mobile-first’ continent like Africa can be very different from that of North America, which relies heavily on fixed broadband internet.
As with many challenges in the application of AI in the health sciences, the answer to both obstacles lies with widening the training pool for the model. The more measurements researchers have from traditional sleep studies in more countries, cultures, and technology contexts, the more confident they can be in any model’s predictions.
Population health and sleep scientists would probably benefit the most if a global sleep observatory (from Internet measurements) were created.
If large changes in Internet use show similar changes in sleep patterns, researchers can come out in the field and use more specific tools to investigate further. Likewise, significant global shocks such as pandemics and recessions can be studied in near real time for their impact on our sleep, prompting appropriate public health messages about mental health and sleep, better technology and app design, and timely education about the importance of sleep. In times of stress.
