Study population and setting
This study included 5,320 participants from across the world who submitted forced-cough data between April to May 2020 on a website and agreed to participate in the study to train a machine learning model to detect COVID-19 through cough recordings. Using online submission of cough recordings from various browsers and devices, all COVID-19 positive cases collected (n = 2,660, determined by either testing, or physician or self-assessment) and the same number of randomly selected negative cases (determined negative by the same methods) were included. 80% of the total cohort were used in a training set, and the remaining 20% for validation. Sensitivity, specificity, and diagnostic accuracy were estimated.
Summary of Main Findings
The artificial intelligence-based COVID-19 screening tool achieved a reported sensitivity of 98.5% (eg. the model correctly identified 98.5% of positive cases as positive) and specificity of 94.2% (eg. correctly identified 94.2% of negative cases as negative), achieving an overall accuracy of 98.5%. In asymptomatic subjects, it achieved a reported sensitivity of 100% and specificity of 83.2%.
Cough recordings were collected from various devices and platforms, across a variety of symptomatic groups, to generate the artificial intelligence-based model for identifying COVID-19. Advanced, robust convolutional neural networks featuring modules previously shown to be successful in diagnosing Alzheimer’s disease from audio voice recordings were used to train the model which may yield to stronger diagnostic ability.
Participant recruitment was based on self-selection and online-based volunteering which is prone to selection bias and could potentially limit the value of this artificial intelligence tool in the general population. Additionally, this study did not report the performance of the AI to detect COVID-19 within specific subject (e.g. racial, gender, age) and recording device subgroups, which could lead to variability in the value of this tool in different populations. Lastly, the bulk of COVID-19 “cases” used to train the model and evaluate its accuracy were from self-assessment (59%) or doctor assessment (28%), versus results from (undefined) “official” tests (13%), so it is unknown what proportion of the cohort’s COVID-19 status were correctly assigned. Validation of this artificial intelligence-based COVID-19 detection tool in the general population and on verified COVID-19 cases and non-cases is necessary to further gauge its value.
This was a preliminary study that developed and evaluated the performance of an artificial intelligence-based model using recorded coughs to identify COVID-19 cases. The reported diagnostic accuracy was high (98.5%) with an 100% asymptomatic detection rate. This study proposes an alternative low-cost and rapid screening method for evaluating COVID-19 using artificial intelligence.
This review was posted on: 14 January 2021