How fast does tracing need to be?

Background was this post in the data science & algorithms channel (Diarmid's post)

Some sketch notes based on that…..

I've been spending quite a long time digging into the NHSX app, and the apparent mess around whether they will or won't use GAEN.

It took me quite a while to get my head around why they didn't want to use GAEN, because there was a lot of stuff around being able to track the spread etc. - which was all of some value, but didn't explain to me to be a good trade-off for the fundamental issues with Background Bluetooth on iOS if you don't use GAEN.

Then I found this series of tweets from an Oxford virologist who is advising the project.
https://twitter.com/OxfordViromics/status/1253383951676329984

The fundamental point is that why contact tracing failed to stop the original epidemics in e.g. Europe & the US wasn't a question of preparedness, or scale, but simply of SPEED.

Pre-symptomatic transmission is such a big factor in the spread that the speed of contact tracing really matters.

And based on this insight, the NHSX approach is therefore to act based on symptom-reporting, rather than on diagnosis - because waiting for a diagnosis is simply too slow to be effective.That is the fundamental reason why they couldn't use GAEN, because GAEN offers no mechanism for weeding out spoofed reports (hence it only works with an authenticated diagnosis, by code or similar).

Now, I don't know if they are correct, or not, that waiting for a diagnosis makes the contact tracing process too slow. But I have to think that they might be right.

If so, then GAEN solutions won't be effective enough to stop the spread. They may help a bit, but not enough.
The implications for Safe Paths are interesting. With a human contact tracer in the loop, we are in a much better position to prevent spoofed data getting into our network than GAEN is. But at the moment, I think all our plans are based on reporting on diagnosis, rather than on symptoms. Could we change this? If we have to do so, to be fast enough to be effective, can we shift to a model where we drive notifications based on symptoms, rather than diagnosis? I don't see any fundamental reasons why the existing flows won't work - with just slightly different messaging needed. But I haven't thought it all through yet. Has anybody been looking at this already?

https://twitter.com/OxfordViromics/status/1253383951676329984

Ramesh: 9/5

We at safepaths believe that covid-tracing is too slow. It shd be fever-tracing. i.e. the moment a user has fever, and has been verified, we shd start tracing contacts. This is how most countries are going to manage the second wave.

Kyle questions:

Does some of the NHSX conclusions have to do with how slow testing is/was though? Do we know how Korea is approaching it?

Do we have specific insight on how to remove spoofed reports from symptom reporting?

Do we know for sure that a mature network effect is too slow? Sure the first diagnosis in the graph of infected carriers is slow, but doesn't it pick up speed if the 2nd node on the graph gets tested/diagnosed much more quickly, because they are being tested as a result of the notification not purely hospitalization

Seems like self-reporting is about overcoming lack of or too slow testing vs the point at which contact tracing starts, yes?

Let’s take these one at a time…

Does some of the NHSX conclusions have to do with how slow testing is/was though?

This is from the paper linked in the tweet: https://twitter.com/OxfordViromics/status/1253383951676329984

The time period here is time from symptoms onset not diagnosis.

We already have a ~12 hour delay for data upload (potentilaly longer if there are few cases, and we need to delay to preserve privacy). I don’t know what a realistic delay is for test + test results, but it seems unlikely that this will be much less than 12 hours. Addin time to arrange & hold teh contact tracing interview, and we are looking at > 24 hours, so we are already into the 2nd graph from the right, possibly the 3rd.

In short, I don’t think this is solely a reflection of how slow testing was at the start. Even with best-case rapid testing, time is still incredibly tight.

Do we know how Korea is approaching it?

No idea. Would need looking into.

Do we have specific insight on how to remove spoofed reports from symptom reporting?

It’s not clear to me what spoofing would look like in our system.

Wiith Bluetooth, you presumably get a Bluetooth phone, leave it all day in some location you want to spike, and then call in reporting symptoms with your unique ID. Then all the people who were near your phone get told to self-isolate.

WIth Safe Paths, you can’t get your data into the system without a contact tracing interview. Yes, someone could make up a load of stuff, but it seems much less likely when this involved a detailed interview with a real person.

The NHSX plan was to monitor how many people who were close to the intiial reporting person later report symptoms. If that number is low/zero, then the whole group can be notified that they can leave quarantine. Being able to detect this depends on matching on a centralized server.

In our model (like GAEN) exposure detection is performned at the phone, and we never learn who was exposed..

This makes it hard to detect a spoofed report.

However, where we are able to follow up every symptom report with a test, we can remove the data from the data set once we get back a negative test result. What’s not clear is how we correlate tests to original symptom reports, if it is all done anonymously.

Also, if Safe Places data counatins both data reflecting a diagnosis & data reflecting self-reports, it seems like it might be good to indicate the two different types of data, and give a different app behaviour based on the type of contact detected….

Do we know for sure that a mature network effect is too slow?

Sure the first diagnosis in the graph of infected carriers is slow, but doesn't it pick up speed if the 2nd node on the graph gets tested/diagnosed much more quickly, because they are being tested as a result of the notification not purely hospitalization

This is interesting - this chart (Tomas Pueyo, but derived from same Oxford report) is probably the relevant one.

Symptoms emerge after 4-8 days (for most)

Most infections happen from 2-8 days.

So when symptoms emerge, others may have been infected 6 to 0 days previously, with 2d probably a typical lag.

2d post-infection, this next group of contacts are about to enter the 6 day period when most onward infections occur. They will be at peak transmission 3d later.

If you can test them within 24-48 hours, then you can indeed cut out the bulk of the transmission from this 2nd wave. Any longer, and you’ll be at or around the peak already, and therefore still allowing > 50% of the transmission.

On the other hand, symtoms won’t emerge in these patients for 4-8 days anyway, which will definitely be too late.

So for the 2nd tier and subsequent contacts in a chain, the only viable solution is testing within < 48 hours, and notification based on that.

If testing within 48 hours is not possible, then we would really need to consider cascading immediately to a 3rd tier of contacts. That would require the 2nd tier of contacts go through a full contact tracing interview, redaction etc. solely on the basis of the fact that they had had exposure to someone who is showing symptoms. Is that plausible? Seems far fetched.

This suggests testing for all notified people within 48 hours is essential

Seems like self-reporting is about overcoming lack of or too slow testing vs the point at which contact tracing starts, yes?

Self-reporting is only relevant for the 1st contact, it’s not relevant for the 2nd tier & beyond, for whom self-reporting is too late, because symptoms peak after transmission peaks.

SInce most people in the system will be 2nd tier & later, not 1s tier, maybe this is not such a big deal…