A diagnostic system may be built in the coming months or years, depending on data availability. Anyone can build it, and enterprises and institutions across the world are encouraged to do so. However, for such a system to be safe enough for free use by the general public, it must use data from people who have all varieties of respiratory conditions. The reasons are explained in some of our papers, and presentations.
As of now, with the data available to most entities worldwide, the systems that can be built will be good for use as screening tools. It is trivial to set up a voice-based screening system. The signatures of Covid-19 infection are as clear in voice as those of common cold, and machine earning procedures, standard features and tools will work just fine to create screening applications. A screening application is one that is used just as an "additional precaution" before you leave home for work, or school or any other public place.
A screening tool is by nature, imprecise. The only situation in which such a tool will be useful is when a) it is easy to use and b) people use it repeatedly.
A diagnostic system is however different. Such a system must have a zero false negative rate. Covid-19 is a deadly disease, and false negatives can cost lives. Not everyone reads disclaimers, and we cannot brush aside the consequences of a user not doing so by maintianing that what happens next is the user's problem. That would be an irresponsible and unethical act on the part of any scientist in the face of such a dangerous pandemic.
While we continue our research, we are slowing down on building a formal diagnostic system ourselves because Carnegie Mellon University is not a medical institution. The large-scale data acquisition required for building such a system is best tackled by the industry, or other institutions and researchers who are better equipped for carrying out clinically curated human subjects data collection.
Recent papers:
Content from our previous webpage (prior to December 2020)
OUR WORK ON COVID-19 VOICE DETECTOR
Current Status as of 1 Oct 2020I gave a seminar on this topic at the Center for Mathematical Modeling, Universidad de Chile, on 29th Sept. This presentation gives details of where we are currently in the detection of Covid-19 from voice and what needs to be done.
I also talked about it in a workshop hed by Cambridge University 25th Sept. That video is not released yet. We are now in the process of documenting our insights obtained from Covid-19 data collected in clinical settings.
In February 2020, we began repurposing our past work on human profiling from voice to apply to the potential detection of Covid-19 from voice. At the time, we were not in possession of acutal voice data, and our hypotheses were entirely based on reports from the medical community on the manner in which Covid-19 afftected respiratory functions in patients. For lack of data, our system for detecting covid-19 from voice then was designed to be a self-learning system that can improve as data come in. An early version of the system was displayed on our website in late March 2020, and was based on inferences and extrapolations made from what is clinically known about the signatures of respiratory illnesses (in general) in voice.
Since then, we have received more data from patients who have tested positive for Covid-19 and are symptomatic. We have been able to gain valuable insights from them and have been able to validate (and invalidate) many of our hypotheses made in the absence of adequate data early on in this pandemic. The more we study these, however, the deeper we find we must explore. We find that the signatures of Covid-19 in voice are (as expected) not easily separable from those of other illnesses that have similar symptoms. We find that while standard features and standard AI architectures may seem to perform well on data, there is catch: they do so only on data that represent a limited number of Covid-19 positive cases with no other confounding factor(s) included in the validation process. These results do not generalize to performace in the real world. Their statistical significance will remain unclear until more data types are included in the studies.
The problems and challenges we have observed must be addressed and solved before we can confidently propose a specific AI driven methodology to detect covid-19 from voice. We have refrained from publishing results that are not results of investigations that close most (if not all) inferential loopholes. We are however in the process of documenting our insights in a paper, with appropriate warnings and caveats clearly mentioned within.
Until we get to the point where we can go for an FDA approval of this
technology, we will not be releasing an actual system for public
use. Until then, our system may be licensed (for evaluation purposes
only) from Carnegie Mellon University.
(Note to commercial entities: You may write to Merlin Inc. in Chile to license their system instead.)
As of now, we continue to collect crowdsourced data. However, in spite of a good response from the public, we find that our most valuable data (in terms of gaining scientific insights) still come from clinical settings.
Please donate your voice for research on covid