‘Cough against COVID’ is an initiative by The Wadhwani Institute for Artificial Intelligence, with support from the Bill and Melinda Gates Foundation, to build AI-based tool that uses cough sounds, symptoms, and other contextual information to screen for possible COVID-19 infection. Dr Rahul Panicker, Chief Research and Innovation Officer, Wadhwani Institute for Artificial Intelligence gives more details about the endeavour and explains how it can contribute to the fight against the novel coronavirus, in an interview with Lakshmipriya Nair
Tell us about the Cough against COVID initiative? Which are the entities involved in this project and what will be their roles in this endeavour? How did the partnership come about?
Cough against COVID is a global data-crowdsourcing and open-innovation initiative to try and build an artificial intelligence (AI) tool that uses cough sounds, symptoms, and other contextual information to screen for possible COVID-19 infection. Such a tool requires nothing more than a phone and can be used by people at home to be triaged for testing and can help healthcare systems save limited tests for the most likely cases.
Cough sounds carry vital information about the respiratory tract, and anecdotal evidence suggests that a COVID-19 patient’s cough sounds different from other coughs. The goal of the initiative is to collect and analyse cough sounds to try and find the early signs of COVID-19 through AI. An AI tool developed would not be a replacement for an established COVID-19 diagnostic test, but rather an ‘in-between’ screening tool to potentially detect who is likely to have COVID-19, allowing them to get tested early and thereby support public health systems in containing the pandemic.
We are supported by the Bill and Melinda Gates Foundation.
You are developing a tool that can diagnose COVID-19 from cough sounds. So, how will that work exactly? What are the parameters being used to determine the diagnosis?
We propose a home-based triaging tool for the public that will combine an analysis of solicited cough sounds as an objective measurement along with self-reported symptoms and contextual information (coarse location to obtain local prevalence) to identify the most probable potential COVID-19 cases and to enable wider but targeted testing.
The tool will require a user to record a cough sound and report the symptoms they are experiencing. The interface could be WhatsApp, a web app, a Facebook Messenger bot, or an API call from any number of third-party symptom checker apps. This can be a very valuable tool as countries grapple with limited and unevenly distributed testing capacity.
Apart from cough recordings, we will also collect socio-demographic information, travel and contact history, other symptoms like fever, cough, shortness of breath, etc., clinical examination parameters like temperature, vitals, etc. and clinical history including co-morbidities. At in-patient wards, we plan to install one smartphone per bed to collect cough recordings from admitted patients. At testing labs, we will collect data from subjects who present for testing.
Is there any study/evidence that provides proof of concept? How are you validating the data received to establish the accuracy of the tool?
There is some prior work suggesting ability to detect respiratory disease from cough. For COVID, this is evidence is in the early stages, and we’re seeing interesting results. Accuracy is established against ground truth from RT-PCR tests.
What have been the results so far? How many correct identifications through this tool?
‘Cough against COVID’ started off as an experiment. We are seeing early promising results, and hope that this bold experiment bears fruit. We didn’t know if it would work, but we felt was worth trying given the potential impact during this crisis around us. We are looking at clinical evaluation in August.
As mentioned earlier, we are running a large global crowdsourced citizen science campaign called Cough against COVID, to encourage COVID-19 tested people to contribute their cough sounds and complete a short survey – this dataset will help build the tool, and also be made available to researchers across the world free of cost. We will validate and anonymise the data we collect before we make the dataset open.
Anecdotal evidence suggests that the COVID cough may be different from the most common coughs. We suspect that, for screening suspected COVID-19 cases, the acoustic signature of a solicited cough could provide a useful signal for improving specificity beyond typical early-stage symptoms. This will be a very valuable tool in any country or region with a limited testing capacity.
Have you applied for any regulatory approvals to commercialise it? Any plans for tie-ups with public and private healthcare partners for large scale deployment of this tool?
We will seek regulatory approvals after clinical evaluation. We are a non-profit and don’t plan to commercialise what we build. We intend to give free access to public health systems.
The Indian Council of Medical Research is eager to support this effort and has recommended sites and has offered to facilitate expedited IRB clearance. We also have potential collaborators who could collect data in South and West Africa. Other facility-based data collection opportunities are welcome. In India, we started data collection at multiple hospitals that offer testing sites as well as isolation/treatment wards for COVID-19 patients.
What are the key learnings garnered about COVID-19 and infectious diseases from this endeavour? Is it possible to modify the tool and leverage the learnings for the management of other infectious diseases, for instance, TB?
We see how important non-invasive and rapidly scalable screening is for containing infectious diseases and the role digital technologies can play. Screening for TB and other respiratory diseases in a similar fashion is something we are certainly keen on exploring.