Prediction of Aspiration Risk by Using Vocal Biomarkers: Machine Learning Development and Validation Study

Area of Science:

• Otolaryngology
• Speech Science
• Artificial Intelligence

Background:

• Aspiration poses risks for respiratory diseases, but current diagnostic methods are invasive or unreliable.
• Subjective bedside evaluations lack consistency, while tests like VFSS and FEES are resource-intensive.

Purpose of the Study:

• To develop and validate a machine learning (ML) algorithm for predicting aspiration risk.
• The algorithm analyzes acoustic features of simple vowel phonations.

Main Methods:

• Retrospective analysis of [i] vowel phonations from 163 patients, recording acoustic features.
• Supervised ML model trained to differentiate high-risk vs. low-risk aspirators, using VFSS for ground truth.
• Model validated on an external cohort and compared to Speech Language Pathologists (SLPs).

Main Results:

• ML model showed significant differences in risk scores between high-risk (0.530) and low-risk (0.243) aspiration groups (p<0.001).
• Achieved an Area Under the Curve (AUC) of 0.76 in the development cohort and 0.70 in the external cohort.
• ML model performance was comparable to trained SLPs in classifying aspiration risk.

Conclusions:

• Quantifiable voice characteristics in Otolaryngology (ENT) patients correlate with aspiration risk.
• An ML model analyzing sustained phonation can effectively detect differences between high- and low-risk aspirators.
• This approach offers a promising, non-invasive method for aspiration risk assessment.