Can a Phone App Detect COVID?
Artificial intelligence can potentially detect it in your cough, even if you show no symptoms.
I was browsing Reddit one evening when a headline grabbed my attention. It was an article from Gizmodo about a neural network ‘100 percent accurate’ at detecting COVID in asymptomatic individuals by listening to a phone recording of a cough.
https://gizmodo.com/this-ai-can-tell-if-you-have-covid-19-just-by-listening-1845540851
Non-invasive and potentially deployable as a phone app — in the late half of 2020 mid quarantine, a tool like this would be very handy indeed. If this model can truly accurately detect COVID, especially asymptomatic cases, it would save an enormous amount of resources in testing equipment, and serve as a prescreening tool to hit a wider range of the population on a regular basis.
A claim like 100 percent accurate raises some flags, and I treat anything i find coming from Reddit with scrutiny, so I followed the links and looked into authors’ paper to see what this is about.
The MIT and Harvard authors J. Laguarta et al (2020) hypothesized that minute differences in coughs between healthy and infected individuals could be detected by artificial intelligence. This boils down to a binary classification problem where the target is either healthy or infected. They collected the data by making a website that takes an audio sample, followed by a short list of questions such as if they were self diagnosed or officially diagnosed, their age, sex, region, etc. They collected 5,320 samples of audio recordings of individuals coughing to train their model.
They used a combination of Poisson and ResNet50 neural networks to train on 80% of the samples, and then tested the model on the remaining 20%. The audio was broken down with a type of Fourier analysis called MFCC (Mel Frequency Cepstral Coefficient), and multiple layers of ResNet50 were trained on biomarkers with features of muscular degradation, vocal chords, subject sentiment, and lungs/respiratory tract.
The conclusion of their results was a sensitivity of 98.5% and a specificity of 94.2%. They then further conclude that for asymptomatic individuals it has a sensitivity of 100% and specificity of 83.2%. Let’s break down what these numbers mean. A sensitivity of 98.5% means that out of 1000 people who have COVID, this model would correctly identify 985 of them.
The 94.2% specificity part means that out of 1000 people who don’t have COVID, 942 will be correctly identified as not having COVID, while 58 will actually be positive. For a non-invasive diagnostic tool that can potentially be downloaded on everyone’s phone, it’s a good start.
For asymptomatic individuals the breakdown is a bit harsher. If you have COVID but don’t show symptoms, this model claims it can 100% predict that you are a positive. However it’s rate of false positives is much higher leading to a specificity of 83.2%. A quick look at the ROC curve shows an AUC of 0.95, indicating a strong ability to correctly classify COVID. In the paper, the authors disclose that even a Fortune 100 company has picked up their model as part of their COVID response.
It will be very interesting to see how this model performs when deployed in the general population. I would be shocked to see that it maintains such a high statistic for classifying asymptomatic individuals. My initial response was that it must be an example of overfitting the model to the data, but I haven’t worked with neural networks yet to know the specifics. Still, this could be a solid layer in the weave of defenses society employs to protect ourselves. For example businesses or schools could require a phone cough check every day to quickly and cheaply screen for asymptomatic infected.
Citation:
J. Laguarta, F. Hueto and B. Subirana, “COVID-19 Artificial Intelligence Diagnosis using only Cough Recordings,” in IEEE Open Journal of Engineering in Medicine and Biology, doi: 10.1109/OJEMB.2020.3026928.
