Petri Hollmén traveled to Tyrol on the 5th of March. He had a bottle of hand sanitizer with him, used it a lot and washed his hands like never before.

Sunday, the 8th he returned home to hear a day afterwards that Tyrol was declared a COVID-19 epidemic area. He decided to work from home given the higher risk of having been in an epidemic area. On Thursday the 12th he woke up feeling normal but his Oura ring measured that his readiness was down to 54 from being normally at 80-90 which was mostly due to having a 1°C elevated temperature at his finger at night.

Even though he felt normal, he went to the doctor and given that he was from an epidemic area, they decided to test him. He tested positive and went to self-quarantine for 14 days. He measured his temperature several times during the following day and it always came back with 36.5°C. The Oura ring provided evidence that led to his diagnosis that wouldn't have been available otherwise.

While he didn’t have true fever as defined by the official gold standard he did have a kind of clinical relevant fever. It’s my impression that our medical community is too focused on their gold standards that are based on old and outdated technology like mercurial thermometers.

Even when new measurements like nightly finger temperature don’t match with the gold standard there are still cases where the information allows for better clinical decision making.

Today, we have cheap sensors and machine learning that provide us with a different context of making medical decisions then going to the doctors office.

Testing by doctors is very important in the fight against COVID-19 but people need to know when it’s time to go to the doctor. Hollmén needed his Oura to know that it was time to get tested professionally.

We need to get good at catching cases of COVID-19 as fast as possible when they happen in the wild if we want to avoid that millions die without us choking our economy by long-term quarantines.

Analysis of Fitbit users found that their resting heart rate and total amount of sleep can be used to predict the official state numbers for influenza-like illness.

It’s very likely that lower heart rate variance and a higher minimum of the nightly heartrate happens in at least some of the COVID-19 cases. Unfortunately, the WHO is stuck in the last century and the official symptoms charts tell us nothing about how common either of those metrics are in COVID-19 patients. Lack of access to those metrics in the official statistics means it’s harder for people who have an Oura Ring, an Apple watch or another device that can measure nightly heartrate to make good decisions about when to go to the doctor or self-quarantine.

Given that Apple sold around 50 million Apple watches between 2018 and 2019, a sizable portion of people could make better decisions if we would have more information about how COVID-19 affects heart rate.

Even more people have access to a smart phone with a decent camera. Having a sore throat is a typical symptom for many virus infections like COVID-19 and a good machine learning algorithm could produce valuable data from those images.

A priori it’s unclear about how much we can learn from such pictures. If a throat of a patient is red due to inflammation a doctor who looks at it, can’t distinguish whether it’s due to snoring or a virus infection.

If a machine learning algorithm could have access to a steady stream of daily imagine of a person’s throat the algorithm could understand a person’s baseline and use that insight to factor out the effects of snoring.

When the gold standard of diagnosing the throat is to look at one image at a particular point in time at the doctor’s office there’s potentially a big improvement to be gained by looking at a series over multiple days. We don’t know how useful such a diagnostic tool is before building it.

Ideally, users of a new app would take an image of their throat every morning after getting up and every evening before going to sleep. They would also measure their temperature with a normal thermometer at both points and enter information about subjective symptoms. If a person gets a proper COVID-19 test, they should also be able to enter the data.

At first we would train the machine learning algorithm to use the images to predict temperature. With enough users our algorithm can learn how the throat of a person having flu differs from their baseline whether or not they are snoring.

As we have more users and some of our users get COVID-19 lab tests our machine learning algorithm can learn to predict the test results directly. It’s the nature of advanced technology that we don’t know how powerful a tool is before it’s developed. Most clinical trials for new drugs find that they don’t live up to their promise.

We need more dakka for COVID-19. Creating an app that does the above function doesn’t cost much and the cost of the project should be worth the potential benefits of catching COVID-19 cases faster and thus preventing people from unknowingly infecting their friends.

New to LessWrong?

New Comment
11 comments, sorted by Click to highlight new comments since: Today at 11:04 PM

Absolutely!! Turning "pre-symptomatic transmission" into "pre-obviously-symptomatic but post-subtly-symptomatic transmission" would be extremely valuable. That would (partially) replace the job currently being done (poorly in some places) by contact-tracing, testing, and/or wholesale population-wide physical distancing / lockdowns. I say this effort should be a high priority for anyone who can contribute to it.

I agree that a lot could be done with those sorts of data.

One company that already is making some use of a similar dataset is Kinsa, who sells smart thermometers. They started a few years ago, tracking trends in the flu in the US based on the temperature readings of the people using their thermometers (along with location, age, and gender). Now they have a coronavirus tracking website up. It looks like the biggest useful thing that they've been able to do so far with their data is to quickly identify hotspots - parts of the country where there has been a spike in the number of people with a fever. That used to be a sign of a local flu outbreak, now it's a sign of a local coronavirus outbreak. From the NYTimes:

Just last Saturday, Kinsa’s data indicated an unusual rise in fevers in South Florida, even though it was not known to be a Covid-19 epicenter. Within days, testing showed that South Florida had indeed become an epicenter.

Companies like Fitbit could make a similar pivot, looking to see if they can find atypical trends in their data in the Seattle area Feb 28 - Mar 9, the Miami area Mar 2-19, etc. And they might be able to take the extra step of identifying new indicators that help identify individuals who may have coronavirus (unlike Kinsa, as high body temperature was already a known indicator).

There are potentially a bunch more useful things that could be done with all of these datasets, if more researchers had access to them. For example, it might be possible to get much more accurate estimates of the number of people who have been infected with coronavirus. I may make another post about this soon.

Eran Segal, a Israeli researcher, developed this questionnaire.

he got trends of thousands of Israelis to fill out daily. and he got a map of suspect area in the country....

Holy crap, I just noticed that most of the latest Fitbit models (including mine) have a built-in pulse oximeter!

This would presumably make it even easier to quickly map outbreaks at the population level, in the way discussed in the Lancet article (leaving aside privacy issues).

At the individual level, unfortunately Fitbit won't give me disaggregated data or a percentage reading. It only shows a graph of 'high' and 'low' variations in blood oxygen overnight, which I have no idea how to interpret. I believe this is due to FDA restrictions.

here's an example (not mine):

EDIT: can't figure out how to make image non-obnoxiously enormous, here's a link instead.

Does anyone know:

a) how accurate is a Fitbit pulse oximeter likely to be, compared to a dedicated device?

b) is there a way to access the data directly, and get a percentage reading?

c) is the 'high' and 'low' variations overnight in any way useful?

It might very well be that they can measure the variations with higher accuracy then they can measure the percentage because there are constant factors like skin color that affect the percentage but that can be factored out when comparing daily values. I guess that for doing self diagnosis you want to know how your values derivate from normal.

From a comment on the Quantified Self forum:

Just today @madprime & I launched a small data collection effort to see if there’s any interesting signal that could be used for those predictions: 1
If there’d be enough data collected one might even be able to see if there’s personalized thresholds

On facebook a person left a comment to an interesting story about military funded research:

Using its own globally-collected data sets, Philips was able to develop a unique algorithm for the Defense Department. Using 165 distinct biomarkers across 41,000 cases, the Philips team was able to create the Rapid Analysis of Threat Exposure, or RATE, algorithm, which the company says can “predict infection 48 hours before clinical suspicion” with better than 85% accuracy.

Now that it has been months since the pandemic started, how is this trend going? It seems to me really compelling. What are the pain points of this movement, or things that keep it from really taking off?

Oura run trials that suggest a decent ability to detect presymptomatic COVID-19 infections. As far as I know nobody cared about creating the app that creates photos of the throat.