How Detecting Asymptomatic Covid-19 Infections Just Got Easier
Using an AI voice-recording tool to predict COVID-19 in people who are not suspected to have the virus of up to 98.5% accuracy.
On October 29, 2020 MIT News released an article in their daily newsletter with incredible news: “Those who are asymptomatic may not be entirely free of changes wrought by the virus.” Researchers at the Institute have noticed differences in speech and coughs between those perfectly healthy and people asymptomatic with the pathogen. It’s clear these minor differences can’t be picked up by the human ear, but for Artificial Intelligence, it’s a piece of cake.
The main way COVID-19 is detected is through early symptoms, which is then confirmed with a test. This is what everyone says: If you show symptoms go test for the virus and/or self-isolate to see if those symptoms vanish. In the absence of a treatment vaccine and cure, early detection of symptoms is vital to cut down on transmission rates through consistent monitoring.
But we all know it’s not that simple. Viruses, especially COVID-19 can be extremely unpredictable and infection from person to person varies. Those with no signs of COVID-19 may pass the infection onto others just as effectively as those tested positive with symptoms. There have even been cases when results that should clearly have returned positive, yielded negative, which goes to show that even professional testing has wormholes (a staggering 20% of tests are false negatives). From there, the lines become more blurry as we try to pinpoint who exactly may have COVID-19 or who may not. The potential of this new AI Algorithm could finally make the picture clear.
This seems like a real shot.
In a paper published recently in the IEEE Journal of Engineering in Medicine and Biology, researchers report on an AI model that distinguishes asymptomatic people from healthy individuals through forced-cough recordings, which people voluntarily submitted through web browsers and devices such as cellphones and laptops.
The model was trained with tens of thousands of samples of coughs and spoken sounds (which included both with those asymptomatic and non-infected patients) so when new data was fed into the algorithm it was able to identify those confirmed with COVID-19.
“When they fed the model new cough recordings, it accurately identified 98.5 percent of coughs from people who were confirmed to have Covid-19, including 100 percent of coughs from asymptomatics — who reported they did not have symptoms but had tested positive for the virus.” — MIT News
And this isn’t the first time voice diagnostic has been so accurate.
This strategy to detect disease/infection isn’t new. At the University of Toronto, a research team developed a deep learning machine (Ludwig), which uses proprietary voice analysis algorithms to pick up on subtle cues that reveal signs of progressive neurological disorders.
“Our platform can analyze natural speech to detect and monitor dementia, aphasia, and various cognitive conditions. Using a short one-minute sample of speech, Ludwig can characterize the speaker’s cognitive, acoustic and linguistic state, including lexical diversity, syntactic complexity, semantic content, and articulation.”
Alongside Ludwig, many other similar technologies have been developed as well that are now fully functional within the health industry. This natural speech analyzation could detect patients potentially with neurodegenerative conditions based on symptoms such as vocal cord strength, sentiment, lung and respiratory performance, and muscle degradation far before other major side effects appear. Many of these symptoms can only be noticed with the aid of AI technology, their abilities far exceeding ours.
The mechanics of voice analysis
Of course, we need to understand how this model was built and why it's so accurate.
This AI algorithm was primarily based on three neural networks:
Measuring vocal cord strength.
First was a trained machine learning algorithm which had the ability to discern different vocal cord powers according to various vocal sounds. For instance, certain noises like “mmm” were more accurate in predicting the weakness of an individual’s vocal cords. Brain Subirana, a research scientist at MIT’s Auto-ID Laboratory, took this network and trained it for over 1,000 hours on a pre-recorded audio dataset to pick up specific words with the sound “mmm,” such as “them,” and then later have the model measure each “mmm’s” intensity.
Speech emotion.
The team then trained a second neural network that could sense the person’s emotional state in their speech. Those with neurological decline conveyed very similar sentiments such as frustration or apathy much more frequently than more common emotions including happiness and being neutral. The model was trained by feeding through different emotional states and their tone of voice in speech as input and the algorithm worked itself out to classify them apart from one another.
Respiratory presentation.
The third algorithm learned to perceive alterations in lung and respiratory performance and was trained by using a database of recorded coughs of different patients.
It’s quite clear all three of these machine learning networks work in very similar fashion, like the fundamentals of most neural networks:
- They first learn to classify one situation from another by using pre-recorded databases and later build its own algorithm according to it’s observations.
- The more data that is repetitively fed through, the more the algorithm can perfect itself in the classification process, thus leading to more accurate results (recurrent neural network).
- When new data is placed as input, the network would be able to easily classify it, after enough training, even if it’s not present in the training database.
Combining all three
With their new AI framework, when the team looped in audio recordings of those diagnosed with Alzheimer's and other similar mental disorders, it could detect the disease faster and to a higher degree of accuracy than pre-existing methods.
Currently, this model’s accuracy has surpassed all those of existing algorithms for finding whether an individual has the disease.
The three models were specifically designed to measure factors that were easter eggs for diagnosing the disease. Afterwards, a fourth algorithm was layered over top to detect any muscular degradation to further enhance the model. Voila, an almost perfect formula of diagnosis.
The link between Alzhemier’s and the pathogen
Yes, those with Alzheimer’s and other diseases of the sort experience neurological symptoms such as temporary neuromuscular impairment, but the interesting piece is that there has been increasing evidence those diagnosed with COVID-19 do so as well.
Could this be a new and improved strategy to properly diagnose the virus, even in asymptomatic hosts?
“The sounds of talking and coughing are both influenced by the vocal cords and surrounding organs. This means that when you talk, part of your talking is like coughing, and vice versa. There’s so much we can find from a simple cough, such as sentiment, vocal cord strength, one’s mother tongue…So we thought, why don’t we try these Alzheimer’s biomarkers for Covid?” — Brian Subirana, Research Scientist in MIT’s Auto-ID Laboratory
This observation clearly ties in with the earlier sample experiment. This means one thing: we have a clear shot of discerning asymptomatic and healthy patients from one another not just more efficiently, but also with incredibly high accuracy.
Tying it in with COVID-19 Detection
As an experiment conducted by researchers at MIT, they collected 70,000 recordings of forced-cough samples, which people voluntarily submitted through a website the team established. Ultimately, the submissions amounted to roughly 200,000 coughs in total, plenty enough to put the algorithm to proper test.
Out of the 70,000 people, around 2,500 had been tested positive, including those who were asymptomatic. The team grasped those specific 2,500 recordings along with another 2,500 more that were randomly drawn out from the batch. 4,000 of the total 5,000 were dedicated to train the neural network while the remaining 1,000 were set aside. They would be later fed into the model after training to see if it could accurately discern the coughs of healthy patients versus those infected, regardless if they were asymptomatic or not.
And the results?
Although the structure was initially intended for individuals with Alzheimer’s and other related diseases, it worked almost perfectly in detecting COVID when those final 1,000 were inputed. Just as Subirana predicted, there were striking similarities between covid patients and healthy people in the four specific biomarkers (vocal cord strength, speech emotion, respiratory presentation, and muscle degradation), which the AI was able to pick up on.
Surprisingly, as the researchers write in their paper, their efforts have revealed “a striking similarity between Alzheimer’s and Covid discrimination.” — MIT News
Without much tweaking necessary, through training, the networks were able to track patterns within these four specific areas and build an algorithm of classification. Not only that, but it’s accuracy rate was a staggering 98.5%, and was able to pinpoint all coughs from even asymptomatic patients!
From nose cotton swabs to voice recordings.
While common tests used to diagnose an infection with the corona virus are almost 100% effective, the same can’t be said of tests to those who’ve gotten COVID more than once and have developed antibodies. Even the first time round of infections there could still be complications, and considering COVID’s drastic rate of infection, we need that 100% guarantee.
While voice recordings are currently only 98.5% accurate, through further modelling the framework, achieving that 100% is certainly not far out of reach. Not only that, the testing process would move much more quickly and the thought of only having to force-cough into a device rather than driving out to get a massive cotton swab jammed up your nose, sounds much more reassuring. Knowing that this method is still novel and has only recently been discovered, there’s so much potential.
That being said, until the vaccine is administered to everyone needed and herd immunity has been established, it’s crucial that we find a more effective way of testing, especially for asymptomatic patients. Now, it seems like voice recordings might offer another way around.
