Parkinson’s Predicted From Smartwatch Data

Machine-learning model could see signs as many as seven years before clinical diagnosis

3 min read
older man looking at smartwatch on wrist and nature in the background

Data from wrist-worn accelerometers showed hidden signs of Parkinson's.


Scientists at Cardiff University, in Wales, have discovered a way to accurately predict the occurrence of Parkinson’s disease years before clinical symptoms show up by analyzing motion data available from common smartwatches. They showed that a type of machine learning trained on wrist-worn accelerometer data was able to accurately determine Parkinson’s risk up to seven years prior to clinical diagnosis.

Parkinson’s disease affects more than 10 million people worldwide. The disease is characterized by a sharp loss of dopamine-producing neurons in the substantia nigra, a structure in the lower brain. Parkinson’s causes those afflicted to gradually lose control of their bodies. This initially presents as an overall decrease in physical activity and later progresses to uncontrollable tremors and the loss of autonomic activities like breathing and heart function.

While Parkinson’s is treatable with drugs that bolster dopamine activity in an affected patient’s brain, there is unfortunately no known cure for the disease. Furthermore, because current clinical tests are costly, diagnosing the disease in its prodromal stages—before tremor onset—can be difficult. This has prevented early diagnosis for patients, and it has also necessarily limited the amount of research done to study Parkinson’s in its early stages. Developing treatments that prevent neuron loss earlier in the disease’s progression may keep a Parkinson’s patient from developing its more debilitating symptoms.

The scientists, working with the UK Dementia Research Institute, based their research on the UK Biobank project (UKBB), a dataset of health records and data collected from more than 500,000 people. From the dataset, scientists drew accelerometry data from participants that later developed Parkinson’s disease. This data, collected from wrist-mounted accelerometers worn over a week-long period, provided time-varying activity measurements that the researchers hypothesized would indicate the later development of Parkinson’s.

“It is very unique that we have this seven-year observation time after data was collected so that we can do retroactive analysis,” says Ph.D. student Ann-Kathrin Schalkamp, who worked on the study with fellow scientists Kathryn J. Peall, Neil A. Harrison, and Cynthia Sandor.

To test the hypothesis, the researchers trained a machine-learning model to predict Parkinson’s disease outcomes from a range of measurements, including accelerometry data, genetic screens, lifestyle descriptions, and blood chemistry. Their analyses showed that the logistic regression models they trained on all the data sources were highly accurate when predicting Parkinson’s. However, they also showed that models trained on accelerometry data alone predicted diagnoses with an average accuracy that was not significantly different, demonstrating that accelerometry is more indicative than any other feature the researchers consider.

In one case, a participant was accurately predicted to have developed the disease from accelerometry data they provided seven years before they were diagnosed. “Because we only had updates from 2016 to 2021, we have a maximum of seven years, but the model was able to predict it,” says Schalkamp. “If we wait another five, 10 years or so, we can then check to see if the data can predict even further” into the future, she says. The researchers plan to expand on their prediction results by using other machine-learning models that may provide even more accurate Parkinson’s disease predictions.

Early Parkinson’s detection and more

The results of this experiment may help drive research into prodromal Parkinson’s disease. Schalkamp believes that earlier detection of prodromal disease symptoms could help patients get care at earlier stages of the disease. Using these results, individuals could link accelerometry data from fitness monitors or smartwatches to their health-care providers, enabling them to accurately assess their own risk of developing Parkinson’s. This service could take a form similar to another existing smartwatch-based application for Parkinson’s patients: Last year, the U.S. Food and Drug Administration approved an application developed by Rune Labs that uses Apple Watch biometric data to provide patients with up-to-date assessments of their disease symptoms.

While early detection of Parkinson’s could benefit patients, the researchers note that their models could create liability or data-privacy concerns if they’re used to control access to health care. “People often have concerns with these health applications, that this data might be shared with insurance companies,” says Schalkamp, who noted that insurers could “adjust their rates or just not insure you if they think you are at high risk of developing something that will cost them a lot of money.” To prevent misuse, Schalkamp suggests that predictions be tied directly to patients’ electronic health records: “The first step would be to not share the raw data. There should be secure storage of the private, identifiable data, and only give out derived phenotypes like the risk of Parkinson’s to certified health-care staff.” Limiting raw data access and model outputs could maintain the benefits of this technology while mitigating some of the risks.

The researchers recently published their results in Nature Medicine.

The Conversation (0)