Researchers at Washington State University have developed an algorithm to increase the accuracy of wearable health devices.

It works by replacing missing health data from wearable sensors with estimated data.

The scientists recently presented their work on the algorithm at the International Joint Conference on Artificial Intelligence (www.ijcai.org/proceedings/2024/807) in Jeju, South Korea. They hope the work could have practical applications, especially for people who are having their health monitored in remote or underserved areas.

Wearable devices are increasingly popular for important health applications, such as for vital sign monitoring, rehabilitation and movement disorders. Especially in rural areas, a wearable device can provide continuous and important health information when a doctor is far away. The wearable devices use sensors to collect information on users’ health and then make decisions using machine-learning algorithms.

However, while the machine-learning algorithms assume data from all sensors are available, that is often not the case, said Ganapati Bhat from WSU’s School of Electrical Engineering & Computer Science, who led the work.

Because of user error, energy limitations or a malfunctioning sensor, the technology can often have missing or incomplete data, which creates inaccuracy in diagnostics, especially in communities where power is only available intermittently.

“Missing data can lead to a significant drop in performance of the health algorithms,” said Bhat. “In the worst case, it can miss catastrophic cases like falls, which impact user health.”

In their work, the WSU (wsu.edu) researchers developed a way to represent the missing data in wearable health applications while maintaining accuracy. The approach aims to represent missing data in an energy-efficient manner since wearable devices are typically constrained by their small batteries.

In addition to Bhat, the work was led by graduate students Dina Hussein and Taha Belkhouja along with associate professor Jana Doppa.

“The key insight is that we do not need the exact representation of the missing sensor data if we can maintain high predictive accuracy for the health task,” said Bhat.

The researchers validated their approach with several wearable health applications, such as when an assistive device is used for paralysed patients, and found their method was highly accurate even when multiple sensors were missing. The researchers now plan to work with the WSU School of Medicine to test their work in gesture and activity recognition applications in real-world settings. The team also plans to apply their work in other application domains, such as environmental monitoring.

The work was funded by Bhat’s National Science Foundation Career award.