James Parkinson first flagged a link between changes in breathing patterns and the debilitating disease that now bears his name. But since his work in the early 19th century, only minimal progress has been made in treating a condition that has become alarmingly prevalent.
A study published Monday offers a glimmer of new hope.
The paper by researchers at the Massachusetts Institute of Technology and several other institutions describes an artificial intelligence tool that can analyze changes in nighttime breathing to detect and track the progression of disease, which causes tremors and other serious issues with movement.
The AI was able to accurately flag Parkinson’s using one night of breathing data collected from a belt worn around the abdomen or from a passive monitoring system that tracks breathing using a low-power radio signal. It was trained on breathing data collected during 12,000 nights of sleep from institutions around the country, and tested on outside data.
Results from 12 patients who didn’t have Parkinson’s during their initial sleep study, and later developed the disease, indicated that the AI may be able to catch Parkinson’s far earlier in the disease process than existing methods. The researchers are working on a new study to confirm that finding.
“That is really the golden question,” said Dina Katabi, a computer scientist and principal investigator from MIT’s Jameel Clinic . “All the indications so far are positive and we hope that we can start detecting Parkinson’s much earlier.”
The ability to do so points to the power of AI to extend the capabilities of human perception to find new signals of disease and shake up the science surrounding some of medicine’s hardest problems. Parkinson’s is especially difficult to treat because there is no objective way to measure the disease. Existing methods rely on a series of largely subjective assessments that must be conducted by highly-trained specialists. But the analysis of breathing patterns offers hope of an objective biomarker that could diagnose and monitor the condition in a person’s home, reducing reliance on visits with far-flung experts.
“Most people never make their way to Harvard,” said Ray Dorsey, a Parkinson’s expert at the University of Rochester who was also a co-author on the study. “It’s a figment of their imagination. That you can measure health outside the clinical setting is really, really powerful, and it opens the door to providing more care in the home.”
A significant portion of the data used in the study was collected via a wireless radio transmitter that could be placed in a subject’s bedroom. Rather than requiring the use of clunky belts and tubes, it measures the reflections off a person’s body, allowing the AI to analyze a surprisingly rich trove of data about human physiology.
Contained in the data is information about a person’s breathing, the pulsing of their blood, and the twitching of their muscles. The information it provides about breathing is continuous, allowing the AI to analyze the entire inhale and exhale movement of the chest.
In teaching the AI to recognize changes related to Parkinson’s, the researchers also created a tool to identify the data that it relied on to make its predictions, allowing the model to offer much richer information about a key symptom of the disease.
It has long been known that people with Parkinson’s, like those with Alzheimer’s, suffer from disturbed sleep. In aligning the data with information collected by an electroencephalogram (EEG), a device that measures brain activity, the researchers found that the model’s detection of Parkinson’s was associated with instances when the alpha and beta bands in the EEG signal were at high power. Conversely, in finding a person was healthy, it placed greater emphasis on the delta band.
“Our input, of course, is not EEG. It’s breathing,” Katabi said. “But you can align the instances in time and ask, what was the state of the EEG in those instances when the model decided that something was important for its decision? And you see it was very consistent in finding that when someone has Parkinson’s, it’s really looking at the falling asleep stage.”
The AI model was also able to differentiate between people with Alzheimer’s and Parkinson’s, indicating that it was not simply honing in on a neurologic degeneration, but a specific signature associated with Parkinson’s.
If further testing shows the model can reliably recognize early signs of the condition and characterize its progression, it could help with the recruitment of patients for clinical trials. Such an advance could also allow investigators to more quickly track their response to a new treatment, potentially making trials faster and less expensive.
Neurologic conditions like Parkinson’s and Alzheimer’s are notorious for the high failure rates of clinical trials, mainly because the conditions are so hard to measure and track.
“It’s hard to say whether nocturnal breathing is going to be the measure you’re going to see change in response to treatment. It may be more useful for diagnosis,” Dorsey said. “But I think if you can get objective measures of disease in the real world, this would tell you in a shorter period of time whether a drug works.”
The sheer amount of digital data collected for the study — nearly 12,000 nights of sleep and 120,000 hours of breathing on 757 of Parkinson’s patients — drives home the power of remote, digital monitoring to gather and analyze information about the condition. Many years of additional study will be needed to show the AI can reliably diagnose Parkinson’s at an earlier stage and track its progression, but Dorsey said this first study is a rare sign of progress.
“The number of Americans with Parkinson’s disease has increased 35% in the last 10 years,” he said. “Yet we’ve seen no major breakthroughs for treating Parkinson’s this century, suggesting that the status quo is not working. We need new approaches to measuring the disease. We need new approaches to caring for people and finding treatments.”