Reviewing the connection between speech and obstructive sleep apnea.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Speech Signal Processing

Background:

• Obstructive Sleep Apnea (OSA) is a prevalent sleep disorder caused by upper airway (UA) blockage during sleep.
• Altered UA structure/function in OSA patients suggests speech analysis as a potential diagnostic tool.
• This review critically examines machine learning (ML) approaches for OSA detection using speech analysis.

Purpose of the Study:

• To critically review and evaluate ML-based speech analysis methods for OSA detection and severity prediction.
• To identify and discuss limitations and potential pitfalls in applying ML for medical diagnostic applications.

Main Methods:

• Utilized a speech database of 426 male Spanish speakers with suspected OSA.
• Evaluated AHI prediction using supervectors/i-vectors and Support Vector Regression (SVR).
• Reviewed OSA classification approaches and analyzed the influence of clinical variables (age, BMI, etc.).

Main Results:

• Poor results were obtained when estimating AHI using supervectors/i-vectors with SVR, contrasting with prior research.
• A careful review revealed methodological limitations and deficiencies in previous studies, potentially leading to overoptimistic findings.
• Identified issues with data sources correlated with confounding patient characteristics and overfitting in feature selection/validation.

Conclusions:

• Methodological deficiencies in ML diagnostic applications, including confounding factors and overfitting, can lead to false discoveries.
• Highlights the importance of rigorous validation and consideration of confounding variables in ML-based medical diagnosis.
• Provides insights for future research on the relationship between speech and OSA detection.