Harnessing machine learning in diagnosing complex hoarseness cases

Area of Science:

• Computational linguistics
• Medical diagnostics
• Artificial intelligence in healthcare

Background:

• Traditional vocal fold pathology diagnosis relies on laryngologist expertise and direct visualization.
• A need exists for supplementary, accessible diagnostic methods.

Purpose of the Study:

• To develop and evaluate a machine-learning algorithm for voice analysis.
• To distinguish between healthy and hoarse voices.
• To identify specific laryngeal disorders from voice recordings.

Main Methods:

• Utilized transfer learning on the HuBERT model with the Saarbruecken Voice Database.
• Employed a two-stage machine learning approach: binary classification (healthy vs. hoarse) and multi-class classification (specific disorders).
• Analyzed data from 2103 sessions, encompassing diverse pathologies and healthy individuals.

Main Results:

• The binary classifier achieved 82% accuracy in differentiating healthy from pathological voices.
• The multi-class classifier demonstrated over 93% accuracy in identifying specific laryngeal disorders, notably Laryngeal Dystonia.
• Laryngeal Dystonia remains a diagnostic challenge, highlighting the potential of AI-assisted methods.

Conclusions:

• Machine learning effectively categorizes voice samples for distinct pathologies.
• This AI-driven approach can enhance patient triage and streamline diagnostic processes.
• The method shows promise for improving care, especially for complex conditions like Laryngeal Dystonia.