Researchers offer guidance on monitoring Parkinson’s with wearable sensors, machine learning

The exploratory study suggests that only a single sensor attached to the hand is necessary to collect the necessary data for automated symptom detection.
By Dave Muoio
Share

Monitoring and automatically identifying the symptoms of Parkinson’s disease patients can be achieved using just a single soft wearable sensor placed on the back of the hand, according to an investigation recently published in NPJ Digital Medicine that sought to optimize such methods of data collection.

According to the researchers — who employed 20 participants, six different sensors and convolutional neural networks (CNNs) in their study — fine motor tasks and walking are each suitable for automated bradykinesia and tremor symptom detection in the upper limbs using just a single hand-attached sensor. Combined data from sensors placed on both sides of the body did not improve symptom detection, and dynamic analyses models based on CNNs that were trained on raw sensor data performed comparably to models based on pre-engineered features.

Interestingly, the researchers found that increasing the number of subjects used to train their detection models improved their performance, yet also noted that model accuracy decreased when including more than 10 individuals. This finding led them to suggest smaller, more targeted training sets to develop “semi-personalized” models based on a patient’s characteristics. Similarly, repeated data collection sessions improved the model’s symptom detection performance, albeit with diminishing returns.

Why it matters

Although examination and observation by a trained physician is currently the gold standard of Parkinson’s symptom evaluation, automated methods using wearable sensors and machine learning models are being employed on a prospective basis, the researchers wrote. So far, however, the methodology of such approaches have greatly varied from implementation to implementation.

“There is no consensus on the minimal data collection procedure, including number of sensors and sensor locations, or number of repeated clinical assessments and type of activities required to develop an accurate model while maximizing patient compliance,” the researchers wrote in the study. “This information is critical towards facilitating the translation of this approach to real-world symptom detection and monitoring in the clinic, as well as in the home and community.”

What’s the trend

Wearables have long been a target for Parkinson’s disease research — Global Kinetics Corporation’s Personal KinetiGraph was cleared by the FDA back in 2014, while the Michael J. Fox Foundation, the Parkinson’s Centre of Excellence and other organizations focused on the condition have launched research projects involving wearables, AI and other digital technologies.