New software built by Princeton University researchers may determine whether someone is infected with COVID-19 in minutes by combining questions about a person’s health with data from wristwatch sensors.

The study led by Niraj Jha, a professor of electrical and computer engineering at Princeton University, resulted in this new diagnostic tool. His group is working on AI technology for COVID-19 detection and diagnosis and monitoring of chronic illnesses, including depression, bipolar disorder, schizophrenia, diabetes, and sickle cell disease.

NeuTigers, a firm created to commercialize Jha’s work, appealed to the FDA seeking permission for its CovidDeep product under the agency’s provision for “software as a medical device.” A Ph.D. student in Jha’s group, Shayan Hassantabar, is the primary author of an article published in IEEE Transactions on Consumer Electronics that describes the creation and testing of CovidDeep. The program combines heart rate, skin temperature, and galvanic skin reaction measurements from wristwatch sensors with blood pressure and oxygen saturation levels, as well as a COVID symptom questionnaire.

Jha’s Princeton research group has long worked on adapting deep learning, a kind of AI that is often energy-intensive, to run on low-power electronic devices like phones and watches rather than centralized cloud computing centers. As it is known, Edge AI has the extra benefit of helping to protect users’ privacy while also increasing security. Reduced neural networks (the “neu” in NeuTigers) that imitate human brain growth are a crucial breakthrough.

According to the researchers, the framework is very generalizable. Smart health care is only one of the applications. The researchers are also using it in cybersecurity and other IoT applications. Machine learning methods, similar to preventative medical interventions, might recognize abnormal trends and assist in correcting software vulnerabilities before a hack begins.

Image Description: Schematic diagram of the CovidDeep framework (GSR: Galvanic skin response, Ox.: oxygen saturation, BP: blood pressure, DT/RF: decision tree/random forest, NN: neural network, KB: knowledge-base, MND: multi-variate Normal distribution, GMM: Gaussian mixture model, KDE: kernel density estimation).
Image Source: CovidDeep: SARS-CoV-2/COVID-19 Test Based on Wearable Medical Sensors and Efficient Neural Networks

Jha’s team has been investigating edge AI for health care applications such as noninvasive diabetes and mental health issue diagnosis using wristwatch and smartphone sensor data in recent years.

Former pharmaceutical executive Adel Laoui audited Jha’s “Machine Learning for Predictive Data Analytics” class in the fall of 2017 and was enthralled by the technology. After the workshop, Laoui approached Jha, who had expertise in designing and deploying novel technology for disease management. NeuTigers was introduced in June 2018 after more conversations with Jha’s Ph.D. students.

“We saw a lot of intersections in our interests,” said Jha. “Smart health care with the help of edge AI was taking off, so it was an opportune moment for a startup in this area. Adel had a lot of connections with angel investors, and so it ramped up very quickly.”

Jha’s lab has licensed many patented technologies to NeuTigers, including diabetes and mental health diagnosis methodologies, as well as security vulnerability identification in internet-of-things systems.

When the COVID-19 pandemic was proclaimed in March 2020, Jha wondered if his team’s deep learning algorithms could be used to detect the virus — particularly in persons who could spread COVID-19 without showing any symptoms, which is a massive concern for disease prevention.

“The hypothesis was that the disease leaves a unique signature on the physiological signals emanating from our body,” said Jha. “This hypothesis seems to have been true, at least for the few diseases we had looked at, so my idea was to see if we could diagnose COVID-19 this way.”

The clinical researchers gathered information from 87 people, 30 of whom tested negative for COVID by PCR, 30 tested positive and were symptomatic, and 27 tested positive but were asymptomatic.

The data was separated into 15-second intervals and included 60 minutes of wristwatch sensor measurements on heart rate, skin temperature, and galvanic skin response (a measure of sweat gland activity). Separately, the professionals took blood pressure and oxygen saturation readings from the patients and a questionnaire to see if they experienced shortness of breath, cough, fever, or any of the other eight symptoms.

Image Description: An illustration of the CovidDeep processing pipeline to generate predictions from data inputs.
Image Source: CovidDeep: SARS-CoV-2/COVID-19 Test Based on Wearable Medical Sensors and Efficient Neural Networks

A part of this data was used to train neural network models to predict a patient’s COVID-19 status. In contrast, another fraction was used to evaluate the models developed by the Princeton researchers, led by Ph.D. student Hassantabar. The scientists discovered that their COVID-19 detection models were 98.1 percent accurate.

The insertion of synthetic data based on the probability distribution of the actual data was one strategy Hassantabar utilized to improve the models’ accuracy — a widely applicable technique that Jha’s group originally employed for other applications. Another approach he used was based on Jha’s group’s grow-and-prune neural network synthesis paradigm.

The strategy has now been verified in a larger field experiment in France, and CovidDeep has been piloted by health groups in the United States and Algeria. NeuTigers is trying to make CovidDeep compatible with select Samsung, Fitbit, and Apple smartwatches, which will also combine blood pressure and pulse oximeter measures to increase CovidDeep’s adoption.

According to Marino, who handled clinical data collection for the study, manually inputting clinical data into a smartphone app might be another helpful screening strategy in many contexts, especially because cellphones are significantly more ubiquitous worldwide than smartwatches. However, according to Laoui, many people may choose to use a smartwatch alone. The researchers are attempting to modify the neural network models to operate with a wristwatch’s limited computational capability.

“I think this could be far superior [to rapid antigen tests], because the accuracy of a rapid test that you do by yourself at home is limited,” said Marino,  a professor of surgery at Thomas Jefferson University in Philadelphia and the executive director of the Jefferson Italy Center. “Pushing the swab up in the nose is obviously a discomfort, and I don’t know if people are going to do that as accurately as they need to. But if you have a device on your wrist that is not invasive and is totally independent from a physical human maneuver, I think that is much better.”

Marino also showed optimism that the technique might aid in the early detection of common diseases like diabetes. In the United States, millions of people have early diabetes, which is almost always curable, yet they are unaware of it. According to the researchers, new information that leads to early diagnosis might lead to “a highly successful treatment,” according to the researchers.

Laoui wants to see if the technologies can be used to diagnose and monitor other ailments, such as cardiovascular issues and sepsis infections, which are becoming more common as the world’s population ages.

“We’re going to have a library of disease models embedded in the watch, and from time to time, we’re going to run the sensors’ information through these disease models, which will be personalized,” says Laoui. “If there is something wrong or something unusual, we are going to inform you in a meaningful way. I believe this new era of smart health care will be powered by edge AI applications and will redefine healthcare delivery and consumer well-being.”

Story Source: S. Hassantabar et al., “CovidDeep: SARS-CoV-2/COVID-19 Test Based on Wearable Medical Sensors and Efficient Neural Networks,” in IEEE Transactions on Consumer Electronics, vol. 67, no. 4, pp. 244-256, Nov. 2021, DOI: 10.1109/TCE.2021.3130228. Background Image Source:


Please enter your comment!
Please enter your name here