Laryngeal Pressure Estimation With a Recurrent Neural Network.

Area of Science:

• Biomedical Engineering
• Computational Acoustics
• Artificial Intelligence in Medicine

Background:

• Accurate quantification of voice production parameters is crucial for phonation research and clinical diagnosis.
• Current inverse problem methods rely on numerical models, which are often computationally intensive.
• Non-invasive estimation of laryngeal parameters is highly desirable for clinical applications.

Purpose of the Study:

• To develop a computationally efficient method for estimating subglottal pressure using artificial intelligence.
• To train a long short-term memory network (LSTM) for real-time subglottal pressure estimation.
• To validate the AI model's performance against experimental data.

Main Methods:

• A long short-term memory network was trained using synthetic data generated from a numerical two-mass model of voice production.
• The trained LSTM network was validated using experimental high-speed ex vivo video recordings of porcine vocal folds.
• Model performance was evaluated based on the mean absolute percentage error in subglottal pressure estimation.

Main Results:

• The LSTM network achieved comparable accuracy in estimating subglottal pressure (21.2% MAE) to previous methods (17.7% MAE).
• The computational cost for evaluating one sample using the LSTM approach is negligibly small, representing a significant speedup.
• The training process required substantially fewer model evaluations compared to traditional optimization techniques.

Conclusions:

• The developed AI-driven approach enables accurate subglottal pressure estimation at a significantly reduced computational cost.
• This methodology shows promise for broader applications in estimating other voice production parameters and integrating with advanced numerical models.
• The computational efficiency is a critical advancement for potential future clinical applications in voice disorder diagnosis and treatment.