Automatic stridor detection using small training set via patch-wise few-shot learning for diagnosis of multiple...

Area of Science:

• Medical technology
• Artificial intelligence in healthcare
• Sleep medicine

Background:

• Stridor is a rare non-motor symptom in multiple system atrophy (MSA), crucial for diagnosis and prognosis.
• Current stridor detection via video-polysomnography is labor-intensive and time-consuming.
• The rarity of stridor poses challenges for collecting diverse patient data for AI model training.

Purpose of the Study:

• To develop an AI-driven method for automatic stridor detection.
• To address data scarcity by employing few-shot learning for stridor diagnosis.
• To improve the efficiency and accuracy of stridor detection in clinical settings.

Main Methods:

• Proposed an AI method combining audio splitting, reintegration, and few-shot learning.
• Utilized video-polysomnography data from patients with and without stridor (MSA, parkinsonism, sleep disorders).
• Validated the few-shot learning approach for processing medical audio signals.

Main Results:

• Achieved over 96% detection accuracy using data from only eight stridor patients for training.
• Demonstrated substantial performance improvements (4%-13%) compared to a state-of-the-art AI baseline.
• Enabled real-time localization of stridor audio patches, supporting physician interpretation.

Conclusions:

• The developed AI method offers high diagnostic performance for stridor detection despite data limitations.
• Few-shot learning is clinically useful for analyzing rare medical audio signals.
• The method provides efficient and interpretable support for physicians in stridor diagnosis and management.