Machine Learning From 8.2 Million Mayo Clinic COVID-19 Clinical Notes Identifies Early Symptoms
On April 7, I reported the results of a study using a mobile app to identify which self-reported symptoms best predict a confirmable case of COVID-19, giving central importance to the loss of smell (anosmia) or the loss of taste (ageusia), with lesser importance to fever, loss of appetite, cough, diarrhea, and fatigue. The study estimated the chances that someone with anosmia or ageusia has COVID-19 are 62%. I suggested this should be the upper bound and that 16% should be the lower bound. Anosmia and aguesia were common symptoms: 59% of those who tested positive had at least one of them.
Today, the Mayo Clinic released as a preprint* the results of a machine learning analysis of 8.2 million clinical notes from the electronic health records of 14,967 patients who were tested for COVID-19, 272 (1.2%) of whom tested positive.
Machine learning is a subset of artificial intelligence where a computer program uses an algorithm to learn from the data it is fed to continuously improve its performance. The analysis focused on the week leading up to the COVID-19 test, which helps identify early predictors of a positive test.
On the one hand, the analysis confirms that anosmia and ageusia are dramatically more common among COVID-19 cases: these symptoms were 28 times more common in those who tested positive than in those who tested negative. On the other hand, the symptoms were not commonly noted: they were found in 3% of positive cases and 0.1% of negative cases.
In the week leading up to the COVID-19 test, a number of other symptoms were also more abundant in cases that turned out to be positive:
15.8% of patients who tested positive had diarrhea, compared to only 5.6% of those who tested negative; diarrhea was 2.8 times more common in positive cases.
7% of positives had a change in appetite, compared to only 3% of negatives; change in appetite was just over 2 times more common in positive cases.
11% of positives had the sweats, compared to only 6% of negatives; the sweats were about twice as common in positive cases.
13% of positives had headaches, compared to 7% of negatives; headaches were 1.85 times as common in positive cases.
14% of positives had fatigue, compared to 9% of negatives; fatigue was about 1.6 times as common in positive cases.
Respiratory failure, which can involve shortness of breath or loss of consciousness, was twice as likely in positives, but not very common (4% in positives, 2% in negatives) and the difference was not statistically significant.
Other signs and symptoms thought to be common were not very predictive:
Although 25% of positives had fever and chills, and 25% had cough, these were only 33% more common in positives than in negatives.
Although muscle and joint pain occurred in 16% of positives, they were only 18% more common in positives than in negatives.
A productive (phlegm-producing) cough was 4% in positives and negatives and had no predictive value.
Some of the symptoms were more common when combined:
Fever and cough together were found in 13.2% of positives and only 3.3% of negatives; the combination was four times more likely in positives.
Sweating and diarrhea together were found in 7.7% of positives and only 1.4% of negatives; the combination was 5.6 times more likely in positives.
It is unclear why the loss of the senses of smell and taste are extremely common when self-reported (59%) but very uncommon in electronic health records (3%). A preprint released yesterday found that in COVID-19 cases that were self-reported on Twitter, 26% had loss of smell and 24% had loss of taste, whereas these were not reported in any clinical studies at the time of the analysis.
It is possible that anosmia and ageusia occur in different proportions in different populations. For example, the high proportion self-reported in the mobile app was from the UK, while the Mayo Clinic is in the US and all the early clinical studies came from China. Twitter, however, is international.
It is also conceivable that anosmia and ageusia are more common in cases that do not require hospital visits and therefore do not generate electronic health records.
However, I wonder if these symptoms simply got lost early on in the panic over managing the more life-threatening aspects of the disease, creating an enduring bias. If the initial reports from China did not capture this information, the early clinical studies they published would not note them. Physicians in other countries would not be looking for them as signs of COVID-19, so may not ask about them; the patients may not think to mention them, and if they do, they may not make it into the clinical notes if the physician doesn't think they are significant. Perhaps if physicians are more proactive in asking about this symptom, its occurrence in the electronic health records will become as common as it is among self-reported cases.
Only this latter explanation seems to explain why even 19% of COVID-19 negatives self-reported loss of smell or taste in the UK mobile app, but only 0.1% of COVID-19-negative patients had these recorded as symptoms in their electronic health records. It seems the doctors just aren't noting it.
The bottom line for now is that none of these symptoms occur in everyone, but:
Anosmia and ageusia, even in the absence of other symptoms, are strong predictors of COVID-19.
Fever and cough are not very predictive by themselves, but are strongly predictive in combination.
Sweating and diarrhea are both predictive on their own, but are more powerful in combination.
Headaches, fatigue, and a change in appetite are the next most useful predictors.
Not having one of these predictors does not mean someone doesn't have COVID-19. It's just less likely. Testing is still necessary for confirmation.
Stay safe,
Chris
Please Support This Service
These research updates are made possible by purchases of The Food and Supplement Guide to the Coronavirus. The guide contains my most up-to-date conclusions about what we should be doing for nutritional and herbal support on top of hygiene and social distancing for added protection. Due to the absence of randomized controlled trials testing nutritional or herbal prevention, these are my best guesses for what is likely to work without significant risk of harm, based on the existing science.
Many people have asked why I am charging for this instead of giving it away for free, given that this is a time of crisis and people are in need.
Unfortunately, I have not been immune to the effects this virus has had on the economy. My revenue from my other offerings started falling in February, and by mid-March I had days where my revenue was zero. I have three people who work for me full-time, and I am doing everything I can to avoid laying any of them off. By mid-March I had depleted 75% of my emergency fund in order to avoid any layoffs, and without charging for the guide I would not have been able to hold out much longer.
Charging for the guide has allowed me to keep everyone working, replete some of my savings, and devote myself to researching COVID-19 full-time. As a result I now publish this daily COVID-19 Research Updates newsletter you are currently reading, am involved in the design of several clinical trials that are in the process of being submitted for registration now, and am able to continually update the guide for free whenever my research warrants it.
By purchasing the guide, you are enabling me to continue devoting my skills to the most important issue we now face. I am genuinely grateful for your contribution. You can purchase the guide using this link.
Disclaimer
I am not a medical doctor and this is not medical advice. I have a PhD in Nutritional Sciences and my expertise is in conducting and interpreting research related to my field. Please consult your physician before doing anything for prevention or treatment of COVID-19, and please seek the help of a physician immediately if you believe you may have COVID-19.
Subscribe
If you aren't subscribed to the research updates, you can sign up here.
Archive
You can access an archive of these updates here.
Comments and Questions
To leave a comment or question, please use the Facebook post for this newsletter.
*Footnotes
* The term “preprint” is often used in these updates. Preprints are studies destined for peer-reviewed journals that have yet to be peer-reviewed. Because COVID-19 is such a rapidly evolving disease and peer-review takes so long, most of the information circulating about the disease comes from preprints.