Audio- and sensor-based classification of essential vocal tremor with and without respiratory involvement

Essential vocal tremor (EVT) involves modulation of the fundamental frequency and intensity related to oscillation within the respiratory, laryngeal, or vocal tract subsystems. Current methods for clinical assessment of EVT focus on laryngeal involvement detected using laryngoscopy. In addition, the most common treatment for EVT (i.e., botulinum toxin injections) targets laryngeal involvement. Detection of tremor outside the larynx requires additional procedures that increase clinical training and time demands. Thus, more readily accessible and automated methods are needed for detection of tremor affecting each subsystem. The current study analyzed microphone and neck-surface vibration sensor signals collected from 12 speakers with EVT producing sustained vowels. Based on visual and tactile assessments and respiratory inductive plethysmography, nine of the 12 participants exhibited respiratory tremor. The participants with respiratory tremor had higher extents of intensity modulation in the microphone and vibration sensor signals than participants without respiratory tremor. Machine learning analyses using the modulation patterns in both signals revealed a trial-level classification accuracy of 85% and a person-level classification accuracy of 100%. These findings indicate that patterns of modulation in microphone and vibration sensor signals may differentiate the subsystems affected by tremor and have the potential to advance clinical assessment of EVT. [Work supported by NIDCD R21DC017001.]