Inexpensive AI-outfitted box looks, listens for respiratory infection in crowds

Researchers at the University of Massachusetts, Amherst, have developed a portable device that combines sensors with AI-based analysis to estimate how many people within a crowd seem to have a respiratory virus such as the seasonal flu or COVID-19.

The team’s idea is to initially place the devices in medical waiting areas, from where it would help prepare staff for caseload ebbs and flows. Later it might be set in larger public spaces, helping to monitor epidemiological trends at the population level.

The device, which the inventors are calling FluSense, processes data from thermal images of people in groups and from audio captures of such sounds as coughing, sneezing and, presumably, groaning with discomfort.

According to the study, running in the March edition of the Proceedings of the Association for Computing Machinery on Interactive, Mobile, Wearable and Ubiquitous Technologies, senior author Tauhidur Rahman, PhD, and team fed their platform more than 350,000 thermal images and 21 million non-speech audio samples from hospital waiting areas at UMass’s University Health Services clinic.

The researchers found the setup could predict daily patient counts with a high level of accuracy.

Further, comparing predictions from the devices against lab-confirmed flu cases, they found their sensor-based features “strongly correlated with laboratory-confirmed influenza trends.”

The device stores no speech, sound or image data that could be used to identify individuals. Its deep neural network draws boundaries around thermal images representing group members, while its audio analyzer distinguishes between speech and non-speech sounds associated with respiratory illness.

The study’s lead author, PhD candidate Forsad Al Hossain, tells the university’s news division FluSense showcases the power of AI combined with edge computing done at or near the source of the data.

“We are trying to bring machine-learning systems to the edge,” Al Hossain says. “All the processing happens right here. These systems are becoming cheaper and more powerful.”

UMass epidemiologist Andrew Lover, PhD, MPH, says the study proves the concept that specific coughing sounds correspond with flu-related illness. “Now we want to validate it beyond this specific hospital setting,” he adds, “and show that we can generalize across locations.”

Click here to access the study and here to read the full UMass-Amherst news item.