MIT Open Voice Model Used to Identify Asymptomatic COVID-19 Patients From Cough Recordings


Image Credit: CC License

MIT researchers collected over 70,000 cough recordings from 5,320 people through a website in April and May of 2020 to see if an AI-based speech model could indicate whether someone had contracted COVID-19 even if they are asymptomatic. The hypothesis outlined in an IEEE Open Journal of Engineering in Medicine and Biology article in late September was, “COVID-19 subjects, especially including asymptomatics, could be accurately discriminated only from a forced-cough cell phone recording using Artificial Intelligence.”

The MIT Open Voice Model was trained using the recordings of 4,256 subjects to identify acoustic biomarkers that could indicate the presence of COVID-19 infection. The model was then tested on the remaining 1,064 subjects to determine efficacy. Models based on a Convolutional Neural Network (CNN) were enhanced with transfer learning from previous data sets designed to identify Alzheimer’s disease.

When the test cough recordings were run through the system, “it accurately identified 98.5% of coughs from people that were confirmed to have COVID-19, including 100 percent of asymptomatic — who reported they did not have symptoms but had tested positive for the virus,” according to reporting by MIT News.

Diagnosis by Vocal Biomarkers

The research group has previously been working on building speech models to diagnose Alzheimer’s, asthma, and pneumonia. This experience suggested that a viral infection of the respiratory system would likely have a discernable impact on the vocal cords. The study included about 2,500 people that had confirmed COVID-19 cases, both those with symptoms and the asymptomatic. According to the MIT News article, “the AI framework originally meant for Alzheimer’s…was able to pick up patterns in the four biomarkers — vocal cord strength, sentiment, lung and respiratory performance, and muscular degration — that are specific to COVID-19.”

Researchers are working on an app and seeking FDA approval. If approved, it could become a non-invasive method for detection that consumers could use on a regular basis to monitor their health status related to COVID-19.

Vocalis Health launched a similar initiative in April and partnered with the Israeli Ministry of Defense to monitor patients with confirmed infections to track the status of their illness. Patients were provided a mobile app they could use to submit daily voice samples. Vocal biomarkers were then used to determine how the disease was progressing and if hospitalization might be required.

Amazon filed a patent in 2018 that specifically referenced the ability for a voice assistant to identify a cough during dialog and use that information to personalize the interaction. That was assumed to have more of a commercial angle than healthcare diagnosis, but there are many applications of voice technology that go well beyond the weather, timers, and asking for your favorite song.

The study by MIT along with work by Vocalis and others suggest early promise that the big investments in speech technology may have immediate applications that could assist with both public health crises and personalize health monitoring. This work was less visible in the past but COVID-19 many have helped elevate its profile. Let’s hope the technology proves successful and widely available very soon.

Israeli Startup Vocalis Health is Partnering with the Government to Refine a Voice Test for Coronavirus

COVID-19 is Accelerating Voice Technology Adoption Despite Accuracy Issues: Adobe Survey

Vocalis CEO Tal Wenderow Discusses Vocal Biomarkers, Healthcare, and Coronavirus – Voicebot Podcast Ep 145