Automatic Voice Disorder Detection from a Practical Perspective.

Area of Science:

• Medical Informatics
• Speech Science
• Machine Learning

Background:

• Voice disorders like dysphonia are prevalent but often diagnosed late.
• Scarcity of annotated data poses challenges for automatic voice disorder detection (AVDD).

Purpose of the Study:

• To develop an effective AVDD system using limited in-domain data combined with out-of-domain data.
• To propose a cost-based metric (normalized expected cost) for evaluating AVDD performance.
• To enhance the interpretability of AVDD system outputs for clinical decision support.

Main Methods:

• Training deep neural networks with a combination of out-of-domain and in-domain voice data.
• Utilizing Bayes decision theory for optimal decision-making based on the normalized expected cost.
• Implementing a calibration stage to improve the interpretability of system confidence scores.

Main Results:

• Demonstrated the feasibility of training AVDD systems with limited in-domain data.
• Showcased the effectiveness of the normalized expected cost metric for performance evaluation.
• Achieved improved score interpretability through a post-hoc calibration method.

Conclusions:

• Combining out-of-domain and in-domain data is a viable strategy for AVDD system development.
• The normalized expected cost metric provides a practical evaluation framework for AVDD.
• Calibrated confidence scores enhance the utility of AVDD systems in clinical practice.