Responsible AI webinar, 20th August
Kasia Jakimowicz, Advocacy and Resource Mobilization Lead, PathCheck Foundation
For many of us, it is still unclear what GAEN-based contact tracing is all about. The Responsible AI initiative embarked on an effort to bring some clarity to the general public and states. On August 20th, MIT Media Lab held a panel with renowned experts to discuss Google Apple Exposure Notification (GAEN) based contact tracing and the challenges of “impossible apps.” “Impossible apps,” as Ramesh Raskar, MIT Media Lab professor, called them, refer to exposure notification apps that use Bluetooth to alert users exposed to Covid-19.
Why are these apps considered impossible? For exposure notification apps to be effective, many challenges need to be dealt with: overcoming privacy concerns, encouraging wide-range adoption by local and national health authorities and by the population itself, reducing the high cost to develop and deploy the app (which can be done with open source solutions such as PathCheck), and following up with manual contact tracing. However, once these digital exposure notification apps are adopted, they can effectively enhance the response to Covid-19 and other pandemics.
How National Bluetooth Server works (Microsoft: Nate Yohannes, Jeff Day; PathCheck: Lira Martenson, Sam Zimmerman)
GAEN apps use Bluetooth Low Energy protocol, built for Android and iOS. With this technology, phones exchange Bluetooth beacons – randomly assigned numbers – when they are in close proximity. These numbers are then captured and stored. This basic ability enables GAEN apps to identify possible Covid-19 exposure by using your phone.
Here’s how it works: when a user first downloads a GAEN app, they must opt-in to access exposure notification APIs. After doing so, the exposure notification APIs generate two numbers: 1) a temporary exposure key (TEK) and 2) a rolling proximity identifier (RPI) that changes every 10 – 15 minutes.
A TEK is just a random number, but an RPI is derived from its TEK and the current time of day. RPIs are important for maintaining security, as using a TEK alone is not enough to safeguard the privacy of the user. Theoretically, it would be possible to track a person moving from location to location by the TEK. With an RPI, a unique number is generated every 10 – 15 minutes, dramatically reducing the possibility of tracing a user. A log of prior RPIs are then stored securely on a device for 14 days.
If a person becomes infected, s/he can decide to share their TEK and RPI. To do so, a verification server managed by an individual state provides a code to the user, enabling them to confirm their Covid-positive status. This way, the possibility of self-reporting false positives is limited. Only with this state-provided code can a user self-report in the app. After confirmation of infection, the user’s numbers are broadcast by the server, allowing other users to cross-match against the RPIs captured by their phone. If a match is found, the
app will notify its user of detection of possible Covid-19 exposure.
Currently US states are transitioning from state-managed diagnosis key servers to a central national one, hosted by the Association of Public Health Laboratories (APHL) and free for states’ use. It was launched by North Dakota and Wyoming on August 12th.
Building the GAEN app with Exposure Notification server V1.5 (PathCheck: Sam Zimmerman, Lina Martensson)
Sam Zimmerman and Lina Martensson, PathCheck Foundation, shared their experience building the GAEN app, server, and health-response program on a global scale.
Each jurisdiction has individual needs, creating unique challenges for the interface between user, app and server. Customization for each state becomes even more complex due to subtle differences between iOS and Android specifications. This means that there will be more divergence in the implementation of Exposure Notification over the coming weeks – due to multiple OS versions, backward compatibility and varying privacy philosophies. Customized implementation is tricky, but necessary: disastrous complications can result from an incorrect implementation.
For example, various factors make it difficult to precisely calculate the exact day of exposure. As a result, an incorrect implementation of the GAEN app can lead to excessive (or missed) push notifications, called the ‘notification armageddon.’
The benefits of Exposure Notification V1.5 and the ability to use a national keys server brings innovation, but also significant complexity. PathCheck Foundation continues to be a central resource for many states to correctly implement V1.5 and V1.6 specifications, working closely with Google/Apple on a daily basis and contributing to APHL/Microsoft efforts for the server.
The misunderstood relationship between testing and Exposure Notification (Kevin Esvelt, MIT MediaLab)
MIT MediaLab prepared a branching process model for Covid-19 testing within the Covid-19 HPC Consortium that revealed a misconception between testing and exposure notification. According to Kevin Esvelt, the probability of getting an incorrectly negative test is extremely high – about a 20% chance of a false negative result, even when contact tracing is performed. This might be attributed to the fact that the virus is difficult to detect early in the course of infection, according to Johns Hopkins researchers.
There are three major take-aways for contact tracing and exposure notification solutions from the model:
- Do not invite testing until 6 days post exposure
- Identify and notify infectors that were the source of clusters (multiple exposures within the same 3-day window) and encourage them to get tested and self-report
- Exposure notification is more important than self-reporting
GAEN tested in action (Washington State: Landon Cox)
Everybody wonders if GAEN-based apps work in practice. A team composed of researchers from the University of Washington Medical Center and Microsoft volunteers recently gave one a test. They ran the pilot study of the Covid Shield GAEN-based app, using 194 devices in a ward of the University of Washington Medical Center in late July of this year.
The GAEN app proved to be effective – the devices in close proximity for long periods of time exchanged a relatively large number of random identifiers and acted as a reliable approximation of physical proximity. The GAEN app exhibited a strong recall compared to self-reported contact within 6 feet for more than 10 minutes. However, users might not always know each other, or might not always remember whom they have met. Therefore, a GAEN contact alone does not always lead to self-reporting by users.
The developers of Exposure Notification apps are still learning. The more we learn and collaborate, the more we can utilize GAEN – turning the impossible apps into an effective tool for pandemic response.