Post-swallowing voice-based aspiration screening in dysphagia using a deep learning approach: insights from audio...

Area of Science:

• Medical Science
• Biomedical Engineering
• Speech Pathology

Background:

• Dysphagia poses a significant aspiration risk, necessitating continuous patient monitoring.
• Current monitoring methods can be resource-intensive and may not be suitable for remote assessment.
• Voice analysis offers a non-invasive approach to assess swallowing function and detect aspiration risk.

Purpose of the Study:

• To develop and validate a voice-based method for detecting aspiration risk in individuals with dysphagia.
• To establish standardized 2-second voice segments derived from physiological constraints for aspiration detection.
• To evaluate the performance of machine learning models in identifying aspiration risk from voice recordings.

Main Methods:

• A prospective cohort study involving 198 participants (≥40 years) with and without aspiration risk (Penetration-Aspiration Scale ≥5).
• Voice data were collected and analyzed using standardized 2-second segments, processed into mel-spectrograms.
• Three machine learning models (male, female, integrated) based on MobileNetV3 within the EfficientAT framework were developed and evaluated using 10-fold cross-validation.

Main Results:

• Phonation duration was significantly shorter in participants with aspiration risk compared to healthy controls (2.22-2.48s vs. 6.19s).
• The integrated model achieved the highest performance with an Area Under the Curve (AUC) of 0.8090 and 82.77% sensitivity.
• The male-specific model showed an AUC of 0.7586 and 91.88% sensitivity, while the female model had an AUC of 0.7376 and 61.11% sensitivity.

Conclusions:

• A physiologically grounded, 2-second voice segment analysis shows potential for accurate aspiration detection in dysphagia.
• Machine learning models, particularly the integrated model, can effectively identify aspiration risk from voice features.
• This approach holds promise for developing telemedicine-based aspiration screening tools, improving accessibility and monitoring.