AI Device Converts Coughs Into Flu and Pandemic Predictions

FluSense uses AI to detect flu and novel viral outbreaks like COVID-19.

Posted Mar 28, 2020

Source: Mohamed_Hassan/Pixabay

Imagine having the ability to predict future viral outbreaks and pandemic trends based on real-time analysis of coughs or sneezes in public spaces where people gather. Last week, a newly revealed invention realized this vision.

Through the combination of state-of-the-art machine learning and edge-computing, researchers at the University of Massachusetts Amherst invented a portable device called FluSense that can monitor, track, and forecast seasonal influenza and viral respiratory illnesses, using sound and heat sensing.

Inventors Forsad Al Hossain, lead author, and Tauhidur Rahman, co-author and assistant professor of computer and information sciences, worked in collaboration with Andrew Lover, epidemiologist and assistant professor in the School of Public Health and Health Sciences, George Corey, M.D., executive director of University Health Services, and Nicholas Reich, biostatistician and director of the UMass-based CDC Influenza Forecasting Center of Excellence. The team published their ground-breaking device in the Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies on March 18, 2020.

The concept of the FluSense portable device is based on AI edge-computing. Machine learning is deployed in a distributed computing paradigm where data is processed closer to the data source. This decentralized computing is achieved through its architecture that consists of a neural computing engine using an Intel Movidius Myriad X Vision Processing Unit (VPU) to run deep learning models on the edge, a microphone array of four microphones to capture sound input, a low-cost Seek CompactPRO thermal camera for heat detection, and an inexpensive, single-board Raspberry PI computer to synchronize the various components.

The device does not store speech nor personally identifiable images. FluSense is able to run real-time deep learning algorithms for sound and heat imaging-based crowd density estimation. The device is able to detect and count the number of coughs and people, data that can be used to forecast viral respiratory pandemics and seasonal flu alike.

The team tested FluSense in a field study in four public health services waiting rooms at the University of Massachusetts Amherst. From December 10, 2018, to July 12, 2019, the researchers collected and partially annotated 21,230,450 no-speech audio snippets and over 350,000 thermal images to create a dataset.

Patient counts were determined by a thermal imaging-based crowd estimation algorithm. From the total thermal images collected, four days were randomly selected and labeled with bounding boxes for shapes of people in order to create a thermal imaging dataset with 2,100 images for crowd density estimation.

The researchers created a cough recognition algorithm using a convolutional neural network (CNN or ConvNet), which is a neural network that uses the mathematical operation of convolution instead of general matrix multiplication in one or more of the layers. Convolutional neural networks are deep neural networks that are often used in computer vision to analyze images. The neurons in the layers of a CNN have the three dimensions of width, height, and depth, and each layer transforms three-dimensional input volume to three-dimensional output volume through a differentiable function. Convolutional neural networks are able to input images, assign weights and biases to determine importance to different objects in the image, and distinguish images.

Drawing upon neuroscience, convolutional neural networks are somewhat like the visual cortex of the human brain, where individual neurons respond to stimuli in the receptive field. Each neuron is connected to only a local region of the input volume versus connecting neurons to all neurons in the previous volume, which is a more efficient approach when processing high-dimensional input data such as images.

"FluSense can accurately predict daily patient counts with a Pearson correlation coefficient of 0.95," the researchers wrote in their study. "We also compared signals from FluSense with the gold standard laboratory-confirmed influenza case data obtained in the same facility and found that our sensor-based features are strongly correlated with laboratory-confirmed influenza trends."

With this proof-of-concept established, the researchers' next plan is to validate FluSense in different public areas and locations beyond hospital settings. In the future, FluSense may prove to be a useful tool to assist public health officials, policymakers, first responders, hospital intensive care units, and health care decision-makers by providing real-time analysis and predictions of outbreaks of the seasonal flu and various other viral respiratory illnesses, including COVID-19, a disease caused by the novel SARS-CoV-2 coronavirus.

Copyright © 2020 Cami Rosso All rights reserved.