Development & Validation of a Machine Learning Model That Uses Voice to Predict Aspiration Risk

Area of Science:

• Medical technology
• Artificial intelligence in healthcare
• Respiratory medicine

Background:

• Aspiration is a significant risk factor for various respiratory diseases.
• Current bedside aspiration assessments lack reliability.
• Gold standard tests like VFSS and FEES are invasive and resource-intensive.

Purpose of the Study:

• To develop and validate a machine learning algorithm for objective aspiration risk screening.
• To analyze voice features for predicting aspiration risk.
• To establish an accessible, non-invasive screening tool.

Main Methods:

• Retrospective analysis of 163 patients' recorded phonations during nasal endoscopy.
• Extraction of acoustic voice features (pitch, jitter, shimmer, HNR).
• Supervised machine learning model trained to differentiate high-risk vs. low-risk aspirators, validated against VFSS ground truth and tested on an external cohort.

Main Results:

• The ML model demonstrated significant differences in risk scores between high- and low-risk aspirators (p<0.001).
• Achieved an AUC of 0.76 in the development cohort and 0.70 in the external cohort.
• Outperformed trained Speech Language Pathologists (SLPs) in accuracy, sensitivity, specificity, and predictive values.

Conclusions:

• Quantifiable voice characteristics differentiate aspiration risk.
• The ML model accurately predicts aspiration risk using sustained phonations.
• This AI-driven approach surpasses human expert performance for aspiration risk evaluation.