Deep Learning-Based Acoustic Screening for Penetration-Aspiration Events Using Short Voice Recordings

Area of Science:

• Medical Technology
• Artificial Intelligence in Medicine
• Speech-Language Pathology

Background:

• Post-swallow airway compromise is a significant concern in swallowing disorders.
• Current diagnostic methods like videofluoroscopic swallowing studies (VFSS) can be resource-intensive.
• There is a need for accessible, non-invasive screening tools.

Purpose of the Study:

• To assess the feasibility of a smartphone-based deep learning AI tool for detecting post-swallow airway compromise.
• To evaluate the accuracy of acoustic analysis of voice recordings for this purpose.

Main Methods:

• A multicenter prospective study involving 208 participants referred for VFSS.
• Recording a 1.5-second sustained phonation ("a~") using a smartphone.
• Classifying swallowing safety using the Penetration-Aspiration Scale (PAS) and training an autoencoder-based anomaly detection model.

Main Results:

• The AI model achieved high performance in the validation set: 90.9% sensitivity, 87.5% specificity, 90.4% accuracy, and 0.98 AUC.
• In an independent test set, the model showed 91.9% sensitivity, 50.0% specificity, 85.2% accuracy, and 0.76 AUC.
• The study highlights promising internal performance for screening.

Conclusions:

• A brief, smartphone-based voice analysis using deep learning AI shows potential for screening post-swallow airway compromise.
• This method could serve as a practical adjunct to identify individuals requiring further instrumental swallowing evaluation.
• Further validation in diverse populations is warranted.