This blog is the high-level summary of the talk given by Serina Chang, Stanford University, and Cecilia Mascolo, the University of Cambridge at the PathCheck Global Health Innovators seminar, I co-organize! Check out https://tiny.cc/pcfresearchtalks to know more.
This episode focused on two important directions of analyzing the current pandemic — understanding mobility patterns and using acoustic signals. Serina Chang talked about — Modeling COVID-19 with large-scale dynamic mobility networks — based on her Nature paper from 2020 and the follow-up KDD paper in 2021. The impetus for this work is based on the premise that SARs-CoV-2, like many other viruses, spreads through people coming in contact with each other at homes, offices, and other random locations. Therefore, understanding the mobility patterns of the people is of paramount importance in efficiently modeling the virus.
The solution for this is 2 pronged — firstly, they make use of a fine-grained hourly mobility network as a proxy for the contact network data. The mobility data were aggregated by the Google/Apple teams in a privacy-preserving manner. Secondly, the dynamics of mobility are fused into the disease dynamic models.
The talk primarily focuses on 3 parts, initially understanding the mobility networks and the disease transmission dynamics, secondly analyzing the inequities that are evident from the mobility patterns, and finally, talks about the decision support tool that the team as build to help the state government with insight and foresight into the pandemic! Check out the talk to more about the amazing work done by Serina and the team!
Later in the talk, we had Cecilia Mascolo, present her work on detecting COVID-19 based on acoustic signals. Her team developed a novel method that could predict the presence of COVID-19 based on voice, coughing, etc. She drew on from how clinical practitioners are trained without any curated datasets of sounds (heartbeat/murmur etc) — here, the expert listens to the patient’s heartbeat, then the students do the same and later they discuss to arrive at a conclusion!
Owing to the respiratory symptoms of COVID-19, we can understand the difference in respiratory sounds produced by infected and non-infected individuals. The team deployed an app that enabled crowdsourcing of voice/cough data. It is important to know that there was a lot of noise in the data, but the team made use of some heuristics to clean it to an extent.
There were interesting instances where the deep learning model picked up acoustic signals dependent on language to determine the presence of COVID-19. That is, the model predicted Italian speakers to be more likely positive as opposed to English speakers. Yet another classic example of DL models picking up spurious correlations. Consequently, the team then developed the model to be cognitive of uncertainty and defer the judgment to an (expert) human-in-the-loop!
More importantly, the acoustic method of analysis can help us understand the progression of the disease in a much scalable way!
Further information on these works can be found in the recent papers published by Cecilia’s group!
Q&A Highlights
Knowing what you know now, what would you do differently?
Serina Chang — The follow-up works we do are actually addressing the gaps in our work… one fundamental aspect to our work is we change mobility and then look for impact, but in the real world this might not happen. The policymakers have access to implement interventions and mobility changes as a consequence of that. Therefore implementing a feedback loop, where policy drives patterns of mobility which in turn drives infection patterns would be important to understand the complex relationship among them!
Cecilia Mascolo — Make changes to the way the data was collected with more explicit instructions! For instance, while coughing, many participants kept their phones too close to their mouth so that the researchers can get the “best quality audio”, but quite the opposite happens when one does that…Furthermore, the focus would be on getting more longitudinal data from people as this tool makes the most sense with its personalization aspect!
If these discussions look interesting, check out the full talk from the video pinned above!
I would like to thank both our speakers — Serina and Cecilia — and my amazing co-organizers — Ramesh Raskar, Christin Glorioso, Nina Reščič, Shanice Hudson, and Tavpritesh Sethi!
Also, if you’re interested in joining our efforts at PathCheck Foundation, head over to — tiny.cc/pathcheckslack. Feel free to drop any feedback or let me know if you would like to attend/speak at our event!