Hierarchical Classification of Cough Sound Using Fine and Coarse-Grain Acoustic Feature Representations: A Doctor's...

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Gender and age are crucial for diagnosing and treating health conditions.
• Previous disease prediction models have not extensively explored age and gender identification.
• Respiratory condition diagnosis can be enhanced by incorporating demographic factors.

Purpose of the Study:

• To introduce a novel Hierarchical Multi-Task Learning (H-MTL) model for respiratory condition diagnosis.
• To re-categorize the COUGHVID dataset with hierarchical labels for gender, age group, and respiratory conditions.
• To enable the model to extract fine and coarse-grain acoustic features for nuanced diagnostic classification.

Main Methods:

• Developed a Hierarchical Multi-Task Learning (H-MTL) model.
• Re-structured the COUGHVID dataset into a multi-level taxonomy with hierarchical labels.
• Employed a shared feature extraction backbone and hierarchical task-specific branches for integrated predictions.
• Compared H-MTL performance against traditional Multi-Task Learning (MTL) and Multi-Class Classification (MCC) models.

Main Results:

• The H-MTL model significantly outperformed MTL and MCC models.
• Achieved superior accuracy, recall, and F1 scores in classifying gender, age group, and respiratory conditions.
• Demonstrated effective extraction of hierarchical acoustic feature representations.

Conclusions:

• The H-MTL model offers a precise tool for respiratory illness diagnosis by integrating gender, age, and condition information.
• The model's hierarchical approach mimics healthcare professionals' decision-making processes.
• This patient-specific diagnostic approach can enhance early detection and personalized treatment strategies for respiratory conditions.