Learning from Few Subjects with Large Amounts of Voice Monitoring Data

Area of Science:

• Biomedical Engineering
• Machine Learning
• Clinical Informatics

Background:

• High-complexity machine learning models show promise in clinical tasks but often require extensive supervision.
• Overfitting to subject-specific features is a significant challenge in small cohorts with long monitoring periods.
• Manual feature engineering or expert knowledge is typically used to mitigate overfitting, which can be impractical.

Purpose of the Study:

• To develop a robust machine learning approach for voice monitoring that generalizes well, even with limited labeled data.
• To decouple feature learning from classification, enabling unsupervised feature extraction.
• To address the overfitting problem common in clinical machine learning applications.

Main Methods:

• A two-step learning strategy was implemented, separating feature learning and classification.
• Unsupervised learning was employed for the feature extraction stage, eliminating the need for manual feature engineering.
• The model was applied to a voice monitoring dataset for two distinct tasks.

Main Results:

• The proposed model effectively classified patients with vocal fold nodules against controls using extracted features.
• Features captured pathology-relevant information, improving prediction accuracy for vocal use in patients compared to controls.
• The method demonstrated generalization to unseen subjects and across different learning tasks, achieving state-of-the-art results.

Conclusions:

• The presented two-step unsupervised feature learning approach effectively generalizes in voice monitoring tasks with limited data.
• This method overcomes the limitations of traditional supervised learning in small cohort studies.
• The findings highlight the potential of unsupervised learning for robust clinical machine learning applications.