Small changes in daily life may help identify whether risk has increased
Long before a stroke announces itself in a hospital, the body has already begun to speak — in disrupted sleep, in stillness where there was once movement, in the dry air of a room where someone lies awake past midnight. Researchers at KAIST in South Korea have built an AI system that listens to these whispers, analyzing the continuous rhythms of daily life among older adults at home to detect cerebrovascular disease risk weeks before clinical diagnosis. Across 1,224 participants and more than 13,000 two-week windows of lived data, the system achieved 96.53% accuracy in distinguishing imminent risk from baseline — a result that asks medicine to reconsider where prevention truly begins.
- Stroke strikes without warning in the clinic, but KAIST researchers have found it leaves a trail of subtle signals in the weeks before — irregular sleep, late-night restlessness, fading evening activity — that most people never notice and most doctors never see.
- An AI trained on lifelog data from 1,224 older adults identified the four-week window before cerebrovascular diagnosis with 96.53% accuracy, distinguishing imminent risk from a baseline recorded three months earlier.
- The system's power lies not just in detection but in explanation — it surfaces which behavioral and environmental patterns drove the alert, giving clinicians a legible reason to act rather than an opaque score to distrust.
- Low humidity, blurred day-night activity boundaries, and delayed sleep onset emerged as unexpected but coherent markers, suggesting the home environment itself encodes health information medicine has largely ignored.
- The technology is not yet ready for clinical deployment — larger validation studies across diverse populations are required — but the research reframes the question: what if prevention began not at the doctor's office, but at the bedroom door?
A stroke arrives without warning — but what if the warning signs were already there, written in the small rhythms of daily life? Researchers at KAIST, working with colleagues from Sungkyunkwan University and Korea University Anam Hospital, have built an AI system that reads those invisible patterns and flags the approach of cerebrovascular disease weeks before a hospital diagnosis would catch it.
The team analyzed lifelog data — continuous records of movement, rest, and indoor conditions — from 1,224 older adults living in their own homes. The dataset contained more than 13,000 two-week samples, each a window into how a person actually lived, not how they described themselves in a clinic. When the AI compared data from the four weeks before a cerebrovascular diagnosis to data from twelve weeks earlier, it distinguished between the two periods with 96.53% accuracy.
What the system saw told a coherent story. In the early prodromal phase, older adults showed frequent activity between 10 p.m. and 2 a.m., with irregular circadian rhythms and delayed sleep onset. As diagnosis approached, evening activity dropped, inactivity increased, and the air in their homes grew drier. These were not dramatic changes — a person might not notice them, and a doctor asking questions in a brief visit almost certainly would not. But weeks of continuous sensor data made them visible.
The system uses explainable AI, surfacing not just a risk score but the specific patterns that drove it — a transparency essential for clinical trust. It is also well-suited to a population that may struggle to articulate subtle changes in their own condition. A sensor captures what a patient cannot or will not report.
The researchers are careful about their claims: this is an early-warning tool, not a stroke predictor, and it does not replace clinical diagnosis. Larger validation studies are needed before real-world deployment. But the work, published in npj Digital Medicine, stands as proof that the ordinary patterns of life — when someone sleeps, how much they move, the humidity of the air around them — can be read by machines as early signs of serious disease. The question now is whether medicine is ready to listen.
A stroke arrives without warning. One moment a person is fine; the next, their life has fractured. But what if the warning signs were already there, written in the small rhythms of daily life—the time someone got out of bed, how long they sat still, whether they slept when the body expected sleep? Researchers at KAIST, South Korea's leading science and technology institute, have built an artificial intelligence system that reads these invisible patterns and flags the approach of cerebrovascular disease weeks before a hospital diagnosis would catch it.
The work emerged from a collaboration between Lisa Lim's team at KAIST's Department of Civil and Environmental Engineering, Jo Woon Chong from Sungkyunkwan University, and Kyung-Hee Cho from Korea University Anam Hospital. They analyzed lifelog data—continuous records of movement, rest, and environmental conditions—collected from 1,224 older adults living in their own homes. The dataset contained 13,362 two-week samples, each one a window into how a person actually lived, not how they reported living to a doctor in an office visit. The researchers fed this information into an AI system trained to spot the subtle shifts that precede a stroke diagnosis.
The results were striking. When the team looked at data from the four weeks immediately before someone received a cerebrovascular disease diagnosis and compared it to data from twelve weeks earlier, the AI distinguished between the two periods with 96.53% accuracy. The system was not simply flagging risk; it was detecting the moment when risk became imminent. This matters because cerebrovascular disease, if caught early, can be managed or prevented. If caught late, it can leave a person paralyzed, unable to speak, or dead.
What the AI actually saw in the data told a coherent story. Older adults in the early stages of cerebrovascular disease showed a particular pattern: frequent continuous activity between 10 p.m. and 2 a.m., hours when the body should be winding down for sleep. Their circadian rhythms had become irregular. Sleep onset was delayed. The distinction between day and night activity had blurred. As diagnosis approached, another shift occurred. Evening activity from 6 p.m. to 10 p.m. dropped noticeably, while periods of inactivity increased. The air in their homes grew drier, and low humidity emerged as another marker of imminent risk. These were not dramatic changes. A person might not notice them. A doctor asking questions in a clinic would almost certainly miss them. But the AI, reading weeks of continuous data, could see them clearly.
The technology works because it uses what researchers call explainable AI—it does not simply output a risk score but shows which patterns and factors drove that score. This transparency matters for clinical use. A doctor or caregiver can understand why the system flagged someone, not just that it did. The system is also designed for a population that often struggles to articulate their own condition. Older adults may not remember or clearly describe subtle changes in their sleep or activity. A sensor in the home captures what they cannot or will not report.
But the researchers are careful about what they claim. This is not a predictor of stroke onset. It does not replace a clinical diagnosis. It is a tool for early warning and prevention, a way to connect people to medical care before crisis arrives. Lisa Lim framed the contribution plainly: the system detects risk signals in small lifestyle changes at home and helps route patients toward medical attention at the right moment. The vision is a healthcare system that catches disease in its early stages rather than treating it after damage is done.
The work was published in June in npj Digital Medicine, a journal ranked in the top 0.3% of its field. Before the technology moves into actual clinical use, the team will need to validate it in larger patient groups and different settings. That work is underway. For now, the study stands as proof that the patterns of ordinary life—when someone sleeps, how much they move, the humidity of the air around them—can be read by machines as harbingers of serious disease. The question is whether medicine is ready to listen.
Citações Notáveis
The key point is not that AI should replace hospital diagnosis, but that it can detect risk signals in small lifestyle changes at home and help connect patients to medical care at the right time— Professor Lisa Lim, KAIST
A Conversa do Hearth Outra perspectiva sobre a história
Why does this matter more than, say, a person just keeping better track of their own sleep and activity?
Because people don't notice gradual changes in themselves. You don't wake up one morning and think, "My evening activity is down 15 percent." You just feel a little off. The AI sees the pattern across weeks of data that a person would never consciously register.
The accuracy number is 96.53 percent. That's high, but it's not perfect. What happens in the 3.5 percent of cases where it's wrong?
That's the critical question before this goes into hospitals. A false positive sends someone to a doctor unnecessarily—not ideal, but manageable. A false negative misses someone at risk. The team knows this, which is why they're calling this a supporting tool, not a replacement for clinical judgment.
You mentioned low humidity as a marker. That seems almost random. Why would dry air in a home correlate with stroke risk?
We don't know yet. It could be that people at risk change their behavior in ways that affect humidity—maybe they move less, spend more time in certain rooms, or the environment itself reflects something about their health status. The AI found the correlation; understanding the mechanism is the next step.
Who actually benefits from this right now? Is it available to use?
Not yet. It's published research, proof of concept. Before it becomes a tool a hospital or home care service actually deploys, it needs validation in larger groups and different populations. The researchers are honest about that timeline.
What's the shift in thinking this represents?
It's the difference between waiting for someone to have a stroke and then treating it, versus watching for the warning signs and intervening before the stroke happens. That's not new as an idea, but having an objective, continuous way to detect those warnings—that changes what's possible.