A Wireless Box Detects Parkinson's From Breathing While You Sleep

Researchers at MIT have built an artificial intelligence model that can detect Parkinson's disease simply by reading a person's breathing patterns while they sleep. The system, developed by Dina Katabi and her team at the MIT Jameel Clinic, does not require blood tests, spinal taps, or brain scans. A device that looks like a home internet router sits in the bedroom, emits radio signals, and analyzes how those signals bounce off the sleeping body to extract breathing patterns. A neural network then assesses whether the person shows signs of Parkinson's, how severe it is, and whether it is getting worse over time. The research was published in Nature Medicine in 2022.

This matters because Parkinson's is notoriously hard to diagnose early. Doctors rely primarily on motor symptoms such as tremors, stiffness, and slowness, but those symptoms often appear several years after the disease has already taken hold. A test that runs passively every night, at home, with zero effort from the patient, could shift diagnosis dramatically earlier and give researchers an objective way to track the disease.

A Clue From 1817

The idea of detecting Parkinson's from breathing is not as strange as it sounds. A relationship between the disease and breathing was noted as early as 1817, in the writings of Dr. James Parkinson himself, the physician who first described the condition. Modern medical studies have shown that respiratory symptoms can manifest years before motor symptoms, which means breathing patterns may carry an early warning signal that clinicians have never had a practical way to read at scale.

Katabi's team connected that old clinical observation to modern machine learning. Yuzhe Yang, then an MIT PhD student, and postdoc Yuan Yuan trained a neural network, a series of connected algorithms that loosely mimic how a brain works, to recognize the breathing signatures associated with Parkinson's. The model was evaluated on a large dataset comprising 7,687 individuals, using data from several hospitals in the United States as well as multiple public datasets. Of those, 757 were Parkinson's patients.

Why Existing Methods Fall Short

Over the years, researchers have investigated detecting Parkinson's using cerebrospinal fluid and neuroimaging. The problem is that these methods are invasive, costly, and require access to specialized medical centers. That makes them unsuitable for the frequent, repeated testing that early diagnosis and continuous tracking of disease progression would demand. A spinal tap is not something anyone wants every month.

The scale of the problem is enormous. Parkinson's is the fastest growing neurological disease in the world and the second most common neurological disorder after Alzheimer's. In the United States alone it afflicts over 1 million people and carries an annual economic burden of 51.9 billion dollars. Yet diagnosis still often waits until tremors appear, and clinical trials for new drugs struggle because measuring disease progression in the real world is so difficult.

A Router That Reads Breath

The hardware is deliberately mundane. The device has the appearance of a home internet router, but instead of providing network access it emits radio signals and analyzes their reflections off the surrounding environment. From those reflections it extracts the subject's breathing patterns without any bodily contact. The breathing signal is fed to the neural network, which assesses Parkinson's in a fully passive manner. There is zero effort required from the patient or a caregiver, and the assessment can happen every single night.

The system also accepts breathing input from a simple breathing belt worn during sleep, which means the AI model itself can be used with existing sleep study equipment in hospitals. The research was performed in collaboration with the University of Rochester, the Mayo Clinic, and Massachusetts General Hospital, with co-authors from Rutgers University and Boston University as well.

Why Drug Developers Care

Beyond earlier diagnosis, the tool has significant implications for Parkinson's drug development. In terms of clinical trials, the results can enable trials with a significantly shorter duration and fewer participants, because the AI provides an objective, nightly measure of disease progression rather than occasional clinic visits. That could accelerate the development of new therapies in a field that has seen no therapeutic breakthroughs this century.

Ray Dorsey, a professor of neurology at the University of Rochester and a co-author of the paper, put the current state of assessment bluntly. He compared existing Parkinson's evaluations to a street lamp at night, showing only a very small segment of the disease's real behavior, and said the contactless sensor helps illuminate the darkness by providing objective, real world assessments of how people are doing at home. He also noted the study is likely one of the largest sleep studies ever conducted on Parkinson's.

Reaching Patients the System Currently Misses

In terms of clinical care, the approach can help assess Parkinson's patients in traditionally underserved communities, including those who live in rural areas and those with difficulty leaving home because of limited mobility or cognitive impairment. A diagnosis that once required traveling to a specialist can happen passively in a bedroom. The research was sponsored by the National Institutes of Health, with partial support from the National Science Foundation and the Michael J. Fox Foundation.

A box on a nightstand. Radio waves nobody can feel. A neural network trained on thousands of nights of breathing. Together they turn one of medicine's oldest observations about Parkinson's into a passive, nightly test that could catch the disease years before tremors ever appear.