Machine Learning of Acoustic Voice Outcomes in Idiopathic Subglottic Stenosis

There are no objective reliable non-invasive screening tools to identify disease severity in patients with idiopathic subglottic stenosis (iSGS), a debilitating and life-threatening disease where scar tissue narrows the airway. This study aims to identify objective voice measures that characterize iSGS disease severity.

Voice recordings of the Rainbow Passage (RP) were obtained in patients with severe iSGS (n = 10) immediately prior to endoscopic balloon dilation (ED), 2 weeks post-treatment (n = 6), and in healthy age-matched volunteers (n = 10). CT neck was obtained before and after ED, and the luminal area was measured relative to cricoid as a percent stenosis. Quantitative voice analysis was performed using VoiceLab, an automated voice analysis program. Unsupervised machine learning was performed using principal component analysis (PCA), and a simple linear regression compared composite voice outcome to luminal airway stenosis.

Patients with severe iSGS had increased RP duration (p = 0.013) and decreased speech rate (p = 0.012), articulation rate (p = 0.017), and average syllable duration (p = 0.0096) compared to healthy controls. Isolated recorded breathing samples during RP in severe iSGS patients showed increased cepstral peak prominence (p < 0.0001), mean breath count (p = 0.0039), mean breath duration (p < 0.0001), and time spent breathing (p = 0.0002) compared to controls. PCA analysis showed complete separation between normal and severe iSGS, with post-ED iSGS intervening the two groups. Linear regression of the composite principal component 1 strongly correlated with luminal airway caliber.

Objective voice analysis using machine learning provides a novel, non-invasive biomarker which can be used to quantify iSGS disease severity in advance of surgical intervention.