A framework for biomarkers of COVID-19 based on neuromotor coordination in speech

Abstract: A framework is proposed to detect and track COVID-19 based on changes in neuromotor coordination across speech subsystems involved in respiration, phonation and articulation. The approach is motivated by evidence of widespread inflammation of COVID-19 throughout the body including lower (i.e., bronchial tubes, diaphragm, lower trachea) and upper (i.e., laryngeal, pharyngeal, oral and nasal) tract injury, as well as by the growing evidence of the virus’ neurological impact. An exploratory study is described involving a small set of pre-COVID-19 (pre-exposure) versus post-COVID-19 (after positive diagnosis but presumed asymptomatic) audio interviews and a larger cohort of control versus post-COVID-19 participants in an online protocol designed by Voca.ai in collaboration with Carnegie Melon University.

For each cohort pair, Cohen’s d effect sizes were measured using coordination of respiration (as measured through the acoustic speech envelope) and laryngeal motion (fundamental frequency and cepstral peak prominence), and coordination of laryngeal and articulatory (formant center frequencies) motion. While there is a strong subject-dependence, group-level morphology of effect sizes indicates a reduced complexity of subsystem coordination. For the later (larger) cohort, an encouraging detection/false alarm tradeoff was estimated using a Gaussian mixture-based classifier. Validation is needed with larger more controlled datasets and addressing confounding influences such as different recording conditions, unbalanced data quantities, and changes in underlying vocal status from pre-to-post time recordings including changes in emotional state.

BiologyComputer sciencePhysics

Audience: general audience

SENSE.nano Symposium from MIT.nano