CRcrushthecurve·May 17, 2020
Contact tracing (of the human-powered kind) is obviously hugely important in reducing the scale of outbreaks in the early stages.
However digital contact tracing has a fatal flaw: Bluetooth cannot be used to reliably estimate proximity in dynamic, real-world scenarios as objects (especially human bodies) absorb huge amounts of the signal.
In many scenarios this can make two people sitting next to each other look like they're 10-20 metres away compared to line-of-sight equivalents (just by having a phone in a pocket, handbag, or even next to a head taking a call). You can easily see this using, for instance, Apple's Bluetooth Explorer tool as part of Xcode developer tools [1] (or any of the bluetooth signal strength tools in the Play / App Stores).
You don't have to rely on DIY tests from the internet though. While they're extremely static
Read the rest
tests, the Singapore TraceTogether team did some field studies highlighting the significant variability across hardware [2]. Their tests ended in a plea for factory calibration data from hardware manufacturers.
The Singapore team has talked about false positives in depth as a major issue (one was someone in a different apartment, because bluetooth goes through walls), which is why they set a hard, low RSSI value to reduce false positives - this means a lot of true positives will be missed too.
The key Australian dev revealing significant issues in Australia's COVIDSafe app also acknowledged the major limitations of BLE. [3]
The problem of course is you have a situation where you cannot determine if a contact is epidemiologically interesting, because accuracy in real-world situations is really down to the 20-30 metres of Bluetooth range, even over longer time-frames.
This means you either have a huge caseload for human tracers to sort the signal from the noise (and this relies on the memory of all participants) or you have some kind of automated system (such as amber alerts that the NHS talks about) and the challenge there is that no-one knows if they're really interesting epidemiologically, as no-one can tell where each party was in the context.
A recent series of talks by bluetooth experts is extremely informative.
In one, an expert discusses all the significant sources of error which creates the huge variability you can see in DIY tests at home. [4]
There are other great talks in that video, but Jen Watson - who leads a team at MIT engaged in advanced signal processing - delivers a good brief talk of the issues, hoping to use statistical analysis - using detection theory of fluctuating signals to estimate interesting contacts. [5]
The takeaway from all this though is that it's a hard problem, and in Watson's talk she quickly moves on to thinking about additional future capabilities (such as features in upcoming Bluetooth standards) that might help improve the resolution.
This does leave us with a large current problem though. Tracking apps have been thoroughly oversold with little evidence of usefulness, and in the case of the UK and Australia government authorities have refused to publish the algorithms they are using to determine proximity from an RSSI value and a phone model.
There is nothing sensitive about this apart from the fact it may reveal the system is not useful for the stated purposes.
[1] https://twitter.com/crushthecurve_/status/125911361479693926...
[2] https://github.com/opentrace-community/opentrace-calibration...
[3] https://twitter.com/jim_mussared/status/1255498092239036417
[4] https://youtu.be/KgKbllhgESc?t=2991
[5] https://youtu.be/KgKbllhgESc?t=3175