Researchers in the United States have developed an artificial intelligence system that could detect Parkinson’s disease in patients earlier than currently possible by analyzing a person’s breathing pattern. The tool could improve the diagnosis and treatment of the ailment, which eludes a cure.

In a medical paper released in the journal Nature Medicine last week, scientists from MIT said they have developed a tool called a neural network — algorithms that mimic the way a human brain works — that can identify whether someone has Parkinson’s disease from how they breathe while sleeping.

Already, engineers and researchers are trying to develop various forms of technology — from iPhone apps to watches — to detect Parkinson’s disease earlier in patients, which is notoriously difficult for doctors to do.

“For diseases like Parkinson’s … one of the biggest challenges is that we need to get to [it] very early on, before the damage has mostly happened in the brain,” said Dina Katabi, an author of the study and a professor of electrical engineering and computer science at MIT. “So being able to detect Parkinson’s early is essential.”

Still, medical ethicists said, the algorithm underlines a broader worry in health care: that technological advances are being used to bolster claims computers should fuel more medical decision-making without yet having significant evidence to back it up. They said algorithms could be helpful in detecting Parkinson’s, but they urge more testing as they worry the technology could create false-positive diagnoses.

“If you read about AI, there’s a vast amount of overselling … that AI is going to solve vast amounts of practical problems,” said Torbjørn Gundersen, who researches the use of algorithms in medicine at Oslo Metropolitan University in Norway. “It hasn’t really proved that yet.”


Parkinson’s is a neurological disease that reduces the amount of dopamine neurons released in the part of the brain that controls movement. As it progresses, people can suffer from tremors, limb stiffness and general slowness. Roughly 60,000 Americans are diagnosed with the disease every year, according to the Parkinson’s Foundation, with nearly 10 million people living with it globally.

Despite the disease’s prevalence, doctors don’t have a widely accepted way to screen for Parkinson’s in patients, said James Beck, the chief scientific officer at the Parkinson’s Foundation. This often results in doctors misdiagnosing the disease or catching it much later in its progression, when tremors may already be apparent.

“It’s really hard,” Beck said. “There’s no blood test. There’s no brain scan. There’s no objective way of saying someone has Parkinson’s disease or not. It requires a skilled clinician.”

Katabi and Yuzhe Yang, an MIT researcher and the study’s lead author, set about trying to solve this problem using machine learning. They trained algorithms on sleep data collected from over 7,600 people, of which roughly 750 had Parkinson’s disease.

To collect the data, researchers developed a tool — similar to the shape of a small box — that could be put into a study participant’s room and gather breathing patterns from people wirelessly while they sleep. Some data was also culled from existing data sets collected at academic sleep centers.

The data was used to train a neural network that ended up predicting with high accuracy whether a person had Parkinson’s or not. It was 90% accurate based on data from one night’s sleep. The model improved to 95% accuracy when analyzing 12 nights of breathing patterns. The neural network could also track how severe Parkinson’s was in a patient.


Katabi said the AI model could provide a host of benefits. Pharmaceutical companies trying to create drugs to treat and cure Parkinson’s could use the tool to better track the severity of the disease in patients enrolled in their clinical trials, speeding up the drug creation process, she said. People who live in remote places, far away from neurologists, could have a way to detect and track the disease without having to make lengthy drives.

“Most of the people who have Parkinson, they tend to live away from these medical centers,” she said. “So they end up not receiving the proper treatment and care from an expert.”

Katabi added that the tool, called an Emerald device, is being used by large pharmaceutical and biotechnology companies working on Parkinson’s treatments, but he declined to name which companies, citing confidentiality agreements.

Beck, of the Parkinson’s Foundation, said the AI tool is simply one of many ways scientists are racing to better detect and track Parkinson’s disease. He said these tools should not replace a physician’s diagnosis, but they should be used as part of a broader strategy that helps doctors identify the disease earlier on.

“This shouldn’t supplement or replace a clinical diagnosis,” he said. “It should assist in it … until we can come up with [a test] that’s a little more biologically based.”

Gundersen, of the Oslo Metropolitan University, said that while the research study is promising, more needs to be done. He said there are numerous studies coming out which show artificial intelligence having an advantage over humans in performing certain medical tasks, such as disease diagnosis, but he noted there are fewer studies showing whether these algorithms improve health outcomes when used in a clinical setting.


Katabi agreed. “We need more data,” she said. “We have just started to produce these results, and we need more evidence.”

Gundersen added that ethically, AI algorithms in health care bring up a larger issue: Who is to blame if computers get a diagnosis wrong?

“If we think that holding people accountable is something of value in society,” he said, “the use of AI [would] challenge this.”