Automated contact tracing is not a coronavirus panacea
I think with all the excitement about using technology for contact tracing, as the product lead for TraceTogether, I feel compelled to call out the following section of the policy brief and white paper that we published to accompany the open-sourcing of the BlueTrace protocol and OpenTrace codebase.
If you ask me whether any Bluetooth contact tracing system deployed or under development, anywhere in the world, is ready to replace manual contact tracing, I will say without qualification that the answer is, No. Not now and, even with the benefit of AI/ML and — God forbid — blockchain 😂 (throw whatever buzzword you want), not for the foreseeable future.
There are critical factors (like ventilation — see below; update: or singing!) that a purely automated system will not have access to. You cannot “big data” your way out of a “no data” situation. Period.
Any attempt to believe otherwise, is an exercise in hubris, and technology triumphalism. There are lives at stake. False positives and false negatives have real-life (and death) consequences. We use TraceTogether to supplement contact tracing — not replace it.
One thing that sets TraceTogether apart from most private efforts to build a Bluetooth contact tracer, is that we have been working closely with the public health authorities from day 1. Theory is theory. Application and execution is something else entirely. The team has shadowed actual real-life contact tracers in order to empathise with their challenges.
Consequently, there has never been any daylight between TraceTogether’s design, and what will best allow our epidemiologists to fight COVID-19 and protect us all.
We look forward to working with both Apple and Google, to provide our input into the draft API specifications. We share a common objective in ensuring that the contact detection service integrates well with, and supports public health authorities around the world, to harness #techforpublicgood in our shared fight against COVID-19.
12 Apr 2020: I received feedback from the chief medical officer of a 15,000-strong organisation, pointing to the Washington choir super-spreader event, and ongoing research into how the coronavirus spreads. In this case, a contact tracing app tuned to detect 2 metre proximity would not have flagged any encounters — but a human adjusting the sensitivity thresholds of an algorithm might. Without a human-in-the-loop, matching histories across cases to contribute new knowledge to the medical community’s understanding of how the coronavirus spreads, would not be possible.
14 Apr 2020: Edits to concluding paragraph.
Human-in-the-loop vs Human-out-of-the-loop
It is possible to implement the BlueTrace protocol and have automated notification of probable close contacts of persons who have been diagnosed with COVID-19. In theory, we appreciate the privacy and scalability benefits of doing so. In practice, our ongoing conversations with public health authority officials performing epidemic surveillance and conducting contact tracing operations compel us to recommend otherwise.
An automated algorithm will necessarily generate both false negatives and false positives. A human contact tracer will similarly make mistakes. However, because a human contact tracer would seek to incorporate information beyond just physical proximity, he/she can correct for systematic biases introduced by automated notification system.
Encounters between individuals can be classified into close, casual and transient contacts for epidemiological purposes, based on proximity and duration of contact. However, these classifications depend on factors such as location/environment. For example, short-duration encounters in enclosed spaces without fresh ventilation often constitute close contact, even if encounter proximity and duration do not meet algorithmic thresholds.
Since Bluetooth-based contact tracing solutions do not, by themselves, record location/environment data, this information needs to be obtained through other means — a human-led contact tracing interview.
A human-in-the-loop system is also necessary to allow judgment to be applied, given the high likelihood of pre-symptomatic transmission of the SARS-CoV-2 virus. Since time is of the essence, contact tracers may preemptively wish to trace selected second-degree close contacts of a COVID-19 patient, in cases where there is a high likelihood of exposure and infection, even if the first-degree close contact has yet to test positive. For example, there may be epidemiological value in tracing close contacts of a close relative of an infected person.
A human-out-of-the-loop system will certainly yield better results than having no system at all, but where a competent human-in-the-loop system with sufficient capacity exists, we caution against an over-reliance on technology.
Finally, the experience of Singapore’s contact tracers suggest that contact tracing should remain a human-fronted process. Contact tracing involves an intensive sequence of difficult and anxiety-laden conversations, and it is the role of a contact tracer to explain how a close contact might have been exposed — while respecting patient privacy — and provide assurance and guidance on next steps.
Singapore’s contact tracers are on the frontline of the fight against COVID-19; they are able to do this because they incorporate multiple sources of information, demonstrate sensitivity in their conversations with Singaporeans who have had probable exposure to SARS-CoV-2, and help to minimise unnecessary anxiety and unproductive panic. These are considerations that an automated algorithm may have difficulty explaining to worried users.
Jason Bay is Senior Director (Government Digital Services) at the Government Technology Agency, Singapore. He is also the product lead for TraceTogether, the world’s first nationwide Bluetooth contact tracing system. The views, thoughts, and opinions expressed in this article belong solely to the author, and do not necessarily reflect those of the Government Technology Agency, or any other group or individual.
