Machine learning algorithm helps Bluetooth contact tracing systems discern per-event risk of COVID-19 transmission

A new proof-of-concept machine learning system outlined in npj digital medicine could provide Bluetooth-based contact tracing tools with the means to quantify the risk of coronavirus transmission when device owners are in close proximity.

The approach, researchers from the Fraunhofer Heinrich Hertz Institute in Berlin write, measures whether the received signal strength of a second party’s outbound Bluetooth signal exceeds a cutoff “high risk” value.

And although that signal strength is subject to substantial variability, the researchers found that their trained model performed reliably in experiments conducted among 48 participants, and had high correlation with ground truth risk when tested among 392 contacts held out from the researchers’ validation data.

The team wrote that their model’s performance among subjects in a controlled environment could likely be improved further if trained using a sample of “true infection events” data donated by willing participants.

While this approach does not account for certain relevant factors such as wearing a face mask, the researchers wrote that the performance of such a system could be augmented by introducing “additional sources of data, like user questionnaires or the phone’s GPS and gyroscope sensor. These are interesting directions of future research.”

WHY IT MATTERS

Bluetooth-based contact tracing systems are being designed and deployed worldwide at the federal, state and organization level. While these systems can identify events in which a proximity contact likely occurred, they have little ability to discern between high-risk contact, or those likely to result in infection, and those in which the risk was minimal.

Incorporating these capabilities into a widespread Bluetooth contact tracing network could help better direct individuals as they decide whether to self-quarantine or release their personal information for contact tracing, and it could help public health authorities direct their manual contact tracing effort toward the highest risk contact events.

“To make this approach practically applicable, i.e., to avoid that every short time or distant encounter raises an alarm, it is crucial to reliably estimate the risk of infection transmission from the BLE signal strength measurements,” the researchers wrote.

THE LARGER TREND

Bluetooth technology is at the heart of Apple and Google’s now widespread Exposure Notifications system. As of the companies’ most recent check-in last month, the framework had been adopted by more than 20 countries in regions worldwide, and was being explored by 25 U.S. states. That update also included new features that would make the system easier for public health authorities to deploy, while cutting out the need for device owners to download a dedicated app when opting in.

These effectiveness of such systems, however, is highly impacted by how many individuals in a given community have opted in – which has proven to be something of a roadblock as consumers become increasingly careful with their personal data as they consider digital COVID-19 tools.